nyrn
/
openclaw
mirror of https://github.com/openclaw/openclaw.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
openclaw/docs/plugins/architecture.md

55 KiB
Raw Blame History

summary read_when title
Plugin architecture internals: capability model, ownership, contracts, load pipeline, runtime helpers
Building or debugging native OpenClaw plugins
Understanding the plugin capability model or ownership boundaries
Working on the plugin load pipeline or registry
Implementing provider runtime hooks or channel plugins
Plugin Architecture

Plugin Architecture

This page covers the internal architecture of the OpenClaw plugin system. For user-facing setup, discovery, and configuration, see Plugins.

Public capability model

Capabilities are the public native plugin model inside OpenClaw. Every native OpenClaw plugin registers against one or more capability types:

Capability Registration method Example plugins
Text inference api.registerProvider(...) openai, anthropic
Speech api.registerSpeechProvider(...) elevenlabs, microsoft
Media understanding api.registerMediaUnderstandingProvider(...) openai, google
Image generation api.registerImageGenerationProvider(...) openai, google
Web search api.registerWebSearchProvider(...) google
Channel / messaging api.registerChannel(...) msteams, matrix

A plugin that registers zero capabilities but provides hooks, tools, or services is a legacy hook-only plugin. That pattern is still fully supported.

External compatibility stance

The capability model is landed in core and used by bundled/native plugins today, but external plugin compatibility still needs a tighter bar than "it is exported, therefore it is frozen."

Current guidance:

  • existing external plugins: keep hook-based integrations working; treat this as the compatibility baseline
  • new bundled/native plugins: prefer explicit capability registration over vendor-specific reach-ins or new hook-only designs
  • external plugins adopting capability registration: allowed, but treat the capability-specific helper surfaces as evolving unless docs explicitly mark a contract as stable

Practical rule:

  • capability registration APIs are the intended direction
  • legacy hooks remain the safest no-breakage path for external plugins during the transition
  • exported helper subpaths are not all equal; prefer the narrow documented contract, not incidental helper exports

Plugin shapes

OpenClaw classifies every loaded plugin into a shape based on its actual registration behavior (not just static metadata):

  • plain-capability -- registers exactly one capability type (for example a provider-only plugin like mistral)
  • hybrid-capability -- registers multiple capability types (for example openai owns text inference, speech, media understanding, and image generation)
  • hook-only -- registers only hooks (typed or custom), no capabilities, tools, commands, or services
  • non-capability -- registers tools, commands, services, or routes but no capabilities

Use openclaw plugins inspect <id> to see a plugin's shape and capability breakdown. See CLI reference for details.

Legacy hooks

The before_agent_start hook remains supported as a compatibility path for hook-only plugins. Legacy real-world plugins still depend on it.

Direction:

  • keep it working
  • document it as legacy
  • prefer before_model_resolve for model/provider override work
  • prefer before_prompt_build for prompt mutation work
  • remove only after real usage drops and fixture coverage proves migration safety

Compatibility signals

When you run openclaw doctor or openclaw plugins inspect <id>, you may see one of these labels:

Signal Meaning
config valid Config parses fine and plugins resolve
compatibility advisory Plugin uses a supported-but-older pattern (e.g. hook-only)
legacy warning Plugin uses before_agent_start, which is deprecated
hard error Config is invalid or plugin failed to load

Neither hook-only nor before_agent_start will break your plugin today -- hook-only is advisory, and before_agent_start only triggers a warning. These signals also appear in openclaw status --all and openclaw plugins doctor.

Architecture overview

OpenClaw's plugin system has four layers:

  1. Manifest + discovery OpenClaw finds candidate plugins from configured paths, workspace roots, global extension roots, and bundled extensions. Discovery reads native openclaw.plugin.json manifests plus supported bundle manifests first.
  2. Enablement + validation Core decides whether a discovered plugin is enabled, disabled, blocked, or selected for an exclusive slot such as memory.
  3. Runtime loading Native OpenClaw plugins are loaded in-process via jiti and register capabilities into a central registry. Compatible bundles are normalized into registry records without importing runtime code.
  4. Surface consumption The rest of OpenClaw reads the registry to expose tools, channels, provider setup, hooks, HTTP routes, CLI commands, and services.

The important design boundary:

  • discovery + config validation should work from manifest/schema metadata without executing plugin code
  • native runtime behavior comes from the plugin module's register(api) path

That split lets OpenClaw validate config, explain missing/disabled plugins, and build UI/schema hints before the full runtime is active.

Channel plugins and the shared message tool

Channel plugins do not need to register a separate send/edit/react tool for normal chat actions. OpenClaw keeps one shared message tool in core, and channel plugins own the channel-specific discovery and execution behind it.

The current boundary is:

  • core owns the shared message tool host, prompt wiring, session/thread bookkeeping, and execution dispatch
  • channel plugins own scoped action discovery, capability discovery, and any channel-specific schema fragments
  • channel plugins execute the final action through their action adapter

For channel plugins, the SDK surface is ChannelMessageActionAdapter.describeMessageTool(...). That unified discovery call lets a plugin return its visible actions, capabilities, and schema contributions together so those pieces do not drift apart.

Core passes runtime scope into that discovery step. Important fields include:

  • accountId
  • currentChannelId
  • currentThreadTs
  • currentMessageId
  • sessionKey
  • sessionId
  • agentId
  • trusted inbound requesterSenderId

That matters for context-sensitive plugins. A channel can hide or expose message actions based on the active account, current room/thread/message, or trusted requester identity without hardcoding channel-specific branches in the core message tool.

This is why embedded-runner routing changes are still plugin work: the runner is responsible for forwarding the current chat/session identity into the plugin discovery boundary so the shared message tool exposes the right channel-owned surface for the current turn.

For channel-owned execution helpers, bundled plugins should keep the execution runtime inside their own extension modules. Core no longer owns the Discord, Slack, Telegram, or WhatsApp message-action runtimes under src/agents/tools. We do not publish separate plugin-sdk/*-action-runtime subpaths, and bundled plugins should import their own local runtime code directly from their extension-owned modules.

For polls specifically, there are two execution paths:

  • outbound.sendPoll is the shared baseline for channels that fit the common poll model
  • actions.handleAction("poll") is the preferred path for channel-specific poll semantics or extra poll parameters

Core now defers shared poll parsing until after plugin poll dispatch declines the action, so plugin-owned poll handlers can accept channel-specific poll fields without being blocked by the generic poll parser first.

See Load pipeline for the full startup sequence.

Capability ownership model

OpenClaw treats a native plugin as the ownership boundary for a company or a feature, not as a grab bag of unrelated integrations.

That means:

  • a company plugin should usually own all of that company's OpenClaw-facing surfaces
  • a feature plugin should usually own the full feature surface it introduces
  • channels should consume shared core capabilities instead of re-implementing provider behavior ad hoc

Examples:

  • the bundled openai plugin owns OpenAI model-provider behavior and OpenAI speech + media-understanding + image-generation behavior
  • the bundled elevenlabs plugin owns ElevenLabs speech behavior
  • the bundled microsoft plugin owns Microsoft speech behavior
  • the bundled google plugin owns Google model-provider behavior plus Google media-understanding + image-generation + web-search behavior
  • the bundled minimax, mistral, moonshot, and zai plugins own their media-understanding backends
  • the voice-call plugin is a feature plugin: it owns call transport, tools, CLI, routes, and runtime, but it consumes core TTS/STT capability instead of inventing a second speech stack

The intended end state is:

  • OpenAI lives in one plugin even if it spans text models, speech, images, and future video
  • another vendor can do the same for its own surface area
  • channels do not care which vendor plugin owns the provider; they consume the shared capability contract exposed by core

This is the key distinction:

  • plugin = ownership boundary
  • capability = core contract that multiple plugins can implement or consume

So if OpenClaw adds a new domain such as video, the first question is not "which provider should hardcode video handling?" The first question is "what is the core video capability contract?" Once that contract exists, vendor plugins can register against it and channel/feature plugins can consume it.

If the capability does not exist yet, the right move is usually:

  1. define the missing capability in core
  2. expose it through the plugin API/runtime in a typed way
  3. wire channels/features against that capability
  4. let vendor plugins register implementations

This keeps ownership explicit while avoiding core behavior that depends on a single vendor or a one-off plugin-specific code path.

Capability layering

Use this mental model when deciding where code belongs:

  • core capability layer: shared orchestration, policy, fallback, config merge rules, delivery semantics, and typed contracts
  • vendor plugin layer: vendor-specific APIs, auth, model catalogs, speech synthesis, image generation, future video backends, usage endpoints
  • channel/feature plugin layer: Slack/Discord/voice-call/etc. integration that consumes core capabilities and presents them on a surface

For example, TTS follows this shape:

  • core owns reply-time TTS policy, fallback order, prefs, and channel delivery
  • openai, elevenlabs, and microsoft own synthesis implementations
  • voice-call consumes the telephony TTS runtime helper

That same pattern should be preferred for future capabilities.

Multi-capability company plugin example

A company plugin should feel cohesive from the outside. If OpenClaw has shared contracts for models, speech, media understanding, and web search, a vendor can own all of its surfaces in one place:

import type { OpenClawPluginDefinition } from "openclaw/plugin-sdk";
import {
  buildOpenAISpeechProvider,
  createPluginBackedWebSearchProvider,
  describeImageWithModel,
  transcribeOpenAiCompatibleAudio,
} from "openclaw/plugin-sdk";

const plugin: OpenClawPluginDefinition = {
  id: "exampleai",
  name: "ExampleAI",
  register(api) {
    api.registerProvider({
      id: "exampleai",
      // auth/model catalog/runtime hooks
    });

    api.registerSpeechProvider(
      buildOpenAISpeechProvider({
        id: "exampleai",
        // vendor speech config
      }),
    );

    api.registerMediaUnderstandingProvider({
      id: "exampleai",
      capabilities: ["image", "audio", "video"],
      async describeImage(req) {
        return describeImageWithModel({
          provider: "exampleai",
          model: req.model,
          input: req.input,
        });
      },
      async transcribeAudio(req) {
        return transcribeOpenAiCompatibleAudio({
          provider: "exampleai",
          model: req.model,
          input: req.input,
        });
      },
    });

    api.registerWebSearchProvider(
      createPluginBackedWebSearchProvider({
        id: "exampleai-search",
        // credential + fetch logic
      }),
    );
  },
};

export default plugin;

What matters is not the exact helper names. The shape matters:

  • one plugin owns the vendor surface
  • core still owns the capability contracts
  • channels and feature plugins consume api.runtime.* helpers, not vendor code
  • contract tests can assert that the plugin registered the capabilities it claims to own

Capability example: video understanding

OpenClaw already treats image/audio/video understanding as one shared capability. The same ownership model applies there:

  1. core defines the media-understanding contract
  2. vendor plugins register describeImage, transcribeAudio, and describeVideo as applicable
  3. channels and feature plugins consume the shared core behavior instead of wiring directly to vendor code

That avoids baking one provider's video assumptions into core. The plugin owns the vendor surface; core owns the capability contract and fallback behavior.

If OpenClaw adds a new domain later, such as video generation, use the same sequence again: define the core capability first, then let vendor plugins register implementations against it.

Need a concrete rollout checklist? See Capability Cookbook.

Contracts and enforcement

The plugin API surface is intentionally typed and centralized in OpenClawPluginApi. That contract defines the supported registration points and the runtime helpers a plugin may rely on.

Why this matters:

  • plugin authors get one stable internal standard
  • core can reject duplicate ownership such as two plugins registering the same provider id
  • startup can surface actionable diagnostics for malformed registration
  • contract tests can enforce bundled-plugin ownership and prevent silent drift

There are two layers of enforcement:

  1. runtime registration enforcement The plugin registry validates registrations as plugins load. Examples: duplicate provider ids, duplicate speech provider ids, and malformed registrations produce plugin diagnostics instead of undefined behavior.
  2. contract tests Bundled plugins are captured in contract registries during test runs so OpenClaw can assert ownership explicitly. Today this is used for model providers, speech providers, web search providers, and bundled registration ownership.

The practical effect is that OpenClaw knows, up front, which plugin owns which surface. That lets core and channels compose seamlessly because ownership is declared, typed, and testable rather than implicit.

What belongs in a contract

Good plugin contracts are:

  • typed
  • small
  • capability-specific
  • owned by core
  • reusable by multiple plugins
  • consumable by channels/features without vendor knowledge

Bad plugin contracts are:

  • vendor-specific policy hidden in core
  • one-off plugin escape hatches that bypass the registry
  • channel code reaching straight into a vendor implementation
  • ad hoc runtime objects that are not part of OpenClawPluginApi or api.runtime

When in doubt, raise the abstraction level: define the capability first, then let plugins plug into it.

Execution model

Native OpenClaw plugins run in-process with the Gateway. They are not sandboxed. A loaded native plugin has the same process-level trust boundary as core code.

Implications:

  • a native plugin can register tools, network handlers, hooks, and services
  • a native plugin bug can crash or destabilize the gateway
  • a malicious native plugin is equivalent to arbitrary code execution inside the OpenClaw process

Compatible bundles are safer by default because OpenClaw currently treats them as metadata/content packs. In current releases, that mostly means bundled skills.

Use allowlists and explicit install/load paths for non-bundled plugins. Treat workspace plugins as development-time code, not production defaults.

Important trust note:

  • plugins.allow trusts plugin ids, not source provenance.
  • A workspace plugin with the same id as a bundled plugin intentionally shadows the bundled copy when that workspace plugin is enabled/allowlisted.
  • This is normal and useful for local development, patch testing, and hotfixes.

Export boundary

OpenClaw exports capabilities, not implementation convenience.

Keep capability registration public. Trim non-contract helper exports:

  • bundled-plugin-specific helper subpaths
  • runtime plumbing subpaths not intended as public API
  • vendor-specific convenience helpers
  • setup/onboarding helpers that are implementation details

Load pipeline

At startup, OpenClaw does roughly this:

  1. discover candidate plugin roots
  2. read native or compatible bundle manifests and package metadata
  3. reject unsafe candidates
  4. normalize plugin config (plugins.enabled, allow, deny, entries, slots, load.paths)
  5. decide enablement for each candidate
  6. load enabled native modules via jiti
  7. call native register(api) hooks and collect registrations into the plugin registry
  8. expose the registry to commands/runtime surfaces

The safety gates happen before runtime execution. Candidates are blocked when the entry escapes the plugin root, the path is world-writable, or path ownership looks suspicious for non-bundled plugins.

Manifest-first behavior

The manifest is the control-plane source of truth. OpenClaw uses it to:

  • identify the plugin
  • discover declared channels/skills/config schema or bundle capabilities
  • validate plugins.entries.<id>.config
  • augment Control UI labels/placeholders
  • show install/catalog metadata

For native plugins, the runtime module is the data-plane part. It registers actual behavior such as hooks, tools, commands, or provider flows.

What the loader caches

OpenClaw keeps short in-process caches for:

  • discovery results
  • manifest registry data
  • loaded plugin registries

These caches reduce bursty startup and repeated command overhead. They are safe to think of as short-lived performance caches, not persistence.

Performance note:

  • Set OPENCLAW_DISABLE_PLUGIN_DISCOVERY_CACHE=1 or OPENCLAW_DISABLE_PLUGIN_MANIFEST_CACHE=1 to disable these caches.
  • Tune cache windows with OPENCLAW_PLUGIN_DISCOVERY_CACHE_MS and OPENCLAW_PLUGIN_MANIFEST_CACHE_MS.

Registry model

Loaded plugins do not directly mutate random core globals. They register into a central plugin registry.

The registry tracks:

  • plugin records (identity, source, origin, status, diagnostics)
  • tools
  • legacy hooks and typed hooks
  • channels
  • providers
  • gateway RPC handlers
  • HTTP routes
  • CLI registrars
  • background services
  • plugin-owned commands

Core features then read from that registry instead of talking to plugin modules directly. This keeps loading one-way:

  • plugin module -> registry registration
  • core runtime -> registry consumption

That separation matters for maintainability. It means most core surfaces only need one integration point: "read the registry", not "special-case every plugin module".

Conversation binding callbacks

Plugins that bind a conversation can react when an approval is resolved.

Use api.onConversationBindingResolved(...) to receive a callback after a bind request is approved or denied:

export default {
  id: "my-plugin",
  register(api) {
    api.onConversationBindingResolved(async (event) => {
      if (event.status === "approved") {
        // A binding now exists for this plugin + conversation.
        console.log(event.binding?.conversationId);
        return;
      }

      // The request was denied; clear any local pending state.
      console.log(event.request.conversation.conversationId);
    });
  },
};

Callback payload fields:

  • status: "approved" or "denied"
  • decision: "allow-once", "allow-always", or "deny"
  • binding: the resolved binding for approved requests
  • request: the original request summary, detach hint, sender id, and conversation metadata

This callback is notification-only. It does not change who is allowed to bind a conversation, and it runs after core approval handling finishes.

Provider runtime hooks

Provider plugins now have two layers:

  • manifest metadata: providerAuthEnvVars for cheap env-auth lookup before runtime load, plus providerAuthChoices for cheap onboarding/auth-choice labels and CLI flag metadata before runtime load
  • config-time hooks: catalog / legacy discovery
  • runtime hooks: resolveDynamicModel, prepareDynamicModel, normalizeResolvedModel, capabilities, prepareExtraParams, wrapStreamFn, formatApiKey, refreshOAuth, buildAuthDoctorHint, isCacheTtlEligible, buildMissingAuthMessage, suppressBuiltInModel, augmentModelCatalog, isBinaryThinking, supportsXHighThinking, resolveDefaultThinkingLevel, isModernModelRef, prepareRuntimeAuth, resolveUsageAuth, fetchUsageSnapshot

OpenClaw still owns the generic agent loop, failover, transcript handling, and tool policy. These hooks are the extension surface for provider-specific behavior without needing a whole custom inference transport.

Use manifest providerAuthEnvVars when the provider has env-based credentials that generic auth/status/model-picker paths should see without loading plugin runtime. Use manifest providerAuthChoices when onboarding/auth-choice CLI surfaces should know the provider's choice id, group labels, and simple one-flag auth wiring without loading provider runtime. Keep provider runtime envVars for operator-facing hints such as onboarding labels or OAuth client-id/client-secret setup vars.

Hook order and usage

For model/provider plugins, OpenClaw calls hooks in this rough order. The "When to use" column is the quick decision guide.

# Hook What it does When to use
1 catalog Publish provider config into models.providers during models.json generation Provider owns a catalog or base URL defaults
-- (built-in model lookup) OpenClaw tries the normal registry/catalog path first (not a plugin hook)
2 resolveDynamicModel Sync fallback for provider-owned model ids not in the local registry yet Provider accepts arbitrary upstream model ids
3 prepareDynamicModel Async warm-up, then resolveDynamicModel runs again Provider needs network metadata before resolving unknown ids
4 normalizeResolvedModel Final rewrite before the embedded runner uses the resolved model Provider needs transport rewrites but still uses a core transport
5 capabilities Provider-owned transcript/tooling metadata used by shared core logic Provider needs transcript/provider-family quirks
6 prepareExtraParams Request-param normalization before generic stream option wrappers Provider needs default request params or per-provider param cleanup
7 wrapStreamFn Stream wrapper after generic wrappers are applied Provider needs request headers/body/model compat wrappers without a custom transport
8 formatApiKey Auth-profile formatter: stored profile becomes the runtime apiKey string Provider stores extra auth metadata and needs a custom runtime token shape
9 refreshOAuth OAuth refresh override for custom refresh endpoints or refresh-failure policy Provider does not fit the shared pi-ai refreshers
10 buildAuthDoctorHint Repair hint appended when OAuth refresh fails Provider needs provider-owned auth repair guidance after refresh failure
11 isCacheTtlEligible Prompt-cache policy for proxy/backhaul providers Provider needs proxy-specific cache TTL gating
12 buildMissingAuthMessage Replacement for the generic missing-auth recovery message Provider needs a provider-specific missing-auth recovery hint
13 suppressBuiltInModel Stale upstream model suppression plus optional user-facing error hint Provider needs to hide stale upstream rows or replace them with a vendor hint
14 augmentModelCatalog Synthetic/final catalog rows appended after discovery Provider needs synthetic forward-compat rows in models list and pickers
15 isBinaryThinking On/off reasoning toggle for binary-thinking providers Provider exposes only binary thinking on/off
16 supportsXHighThinking xhigh reasoning support for selected models Provider wants xhigh on only a subset of models
17 resolveDefaultThinkingLevel Default /think level for a specific model family Provider owns default /think policy for a model family
18 isModernModelRef Modern-model matcher for live profile filters and smoke selection Provider owns live/smoke preferred-model matching
19 prepareRuntimeAuth Exchange a configured credential into the actual runtime token/key just before inference Provider needs a token exchange or short-lived request credential
20 resolveUsageAuth Resolve usage/billing credentials for /usage and related status surfaces Provider needs custom usage/quota token parsing or a different usage credential
21 fetchUsageSnapshot Fetch and normalize provider-specific usage/quota snapshots after auth is resolved Provider needs a provider-specific usage endpoint or payload parser

If the provider needs a fully custom wire protocol or custom request executor, that is a different class of extension. These hooks are for provider behavior that still runs on OpenClaw's normal inference loop.

Provider example

api.registerProvider({
  id: "example-proxy",
  label: "Example Proxy",
  auth: [],
  catalog: {
    order: "simple",
    run: async (ctx) => {
      const apiKey = ctx.resolveProviderApiKey("example-proxy").apiKey;
      if (!apiKey) {
        return null;
      }
      return {
        provider: {
          baseUrl: "https://proxy.example.com/v1",
          apiKey,
          api: "openai-completions",
          models: [{ id: "auto", name: "Auto" }],
        },
      };
    },
  },
  resolveDynamicModel: (ctx) => ({
    id: ctx.modelId,
    name: ctx.modelId,
    provider: "example-proxy",
    api: "openai-completions",
    baseUrl: "https://proxy.example.com/v1",
    reasoning: false,
    input: ["text"],
    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
    contextWindow: 128000,
    maxTokens: 8192,
  }),
  prepareRuntimeAuth: async (ctx) => {
    const exchanged = await exchangeToken(ctx.apiKey);
    return {
      apiKey: exchanged.token,
      baseUrl: exchanged.baseUrl,
      expiresAt: exchanged.expiresAt,
    };
  },
  resolveUsageAuth: async (ctx) => {
    const auth = await ctx.resolveOAuthToken();
    return auth ? { token: auth.token } : null;
  },
  fetchUsageSnapshot: async (ctx) => {
    return await fetchExampleProxyUsage(ctx.token, ctx.timeoutMs, ctx.fetchFn);
  },
});

Built-in examples

  • Anthropic uses resolveDynamicModel, capabilities, buildAuthDoctorHint, resolveUsageAuth, fetchUsageSnapshot, isCacheTtlEligible, resolveDefaultThinkingLevel, and isModernModelRef because it owns Claude 4.6 forward-compat, provider-family hints, auth repair guidance, usage endpoint integration, prompt-cache eligibility, and Claude default/adaptive thinking policy.
  • OpenAI uses resolveDynamicModel, normalizeResolvedModel, and capabilities plus buildMissingAuthMessage, suppressBuiltInModel, augmentModelCatalog, supportsXHighThinking, and isModernModelRef because it owns GPT-5.4 forward-compat, the direct OpenAI openai-completions -> openai-responses normalization, Codex-aware auth hints, Spark suppression, synthetic OpenAI list rows, and GPT-5 thinking / live-model policy.
  • OpenRouter uses catalog plus resolveDynamicModel and prepareDynamicModel because the provider is pass-through and may expose new model ids before OpenClaw's static catalog updates.
  • GitHub Copilot uses catalog, auth, resolveDynamicModel, and capabilities plus prepareRuntimeAuth and fetchUsageSnapshot because it needs provider-owned device login, model fallback behavior, Claude transcript quirks, a GitHub token -> Copilot token exchange, and a provider-owned usage endpoint.
  • OpenAI Codex uses catalog, resolveDynamicModel, normalizeResolvedModel, refreshOAuth, and augmentModelCatalog plus prepareExtraParams, resolveUsageAuth, and fetchUsageSnapshot because it still runs on core OpenAI transports but owns its transport/base URL normalization, OAuth refresh fallback policy, default transport choice, synthetic Codex catalog rows, and ChatGPT usage endpoint integration.
  • Google AI Studio and Gemini CLI OAuth use resolveDynamicModel and isModernModelRef because they own Gemini 3.1 forward-compat fallback and modern-model matching; Gemini CLI OAuth also uses formatApiKey, resolveUsageAuth, and fetchUsageSnapshot for token formatting, token parsing, and quota endpoint wiring.
  • OpenRouter uses capabilities, wrapStreamFn, and isCacheTtlEligible to keep provider-specific request headers, routing metadata, reasoning patches, and prompt-cache policy out of core.
  • Moonshot uses catalog plus wrapStreamFn because it still uses the shared OpenAI transport but needs provider-owned thinking payload normalization.
  • Kilocode uses catalog, capabilities, wrapStreamFn, and isCacheTtlEligible because it needs provider-owned request headers, reasoning payload normalization, Gemini transcript hints, and Anthropic cache-TTL gating.
  • Z.AI uses resolveDynamicModel, prepareExtraParams, wrapStreamFn, isCacheTtlEligible, isBinaryThinking, isModernModelRef, resolveUsageAuth, and fetchUsageSnapshot because it owns GLM-5 fallback, tool_stream defaults, binary thinking UX, modern-model matching, and both usage auth + quota fetching.
  • Mistral, OpenCode Zen, and OpenCode Go use capabilities only to keep transcript/tooling quirks out of core.
  • Catalog-only bundled providers such as byteplus, cloudflare-ai-gateway, huggingface, kimi-coding, modelstudio, nvidia, qianfan, synthetic, together, venice, vercel-ai-gateway, and volcengine use catalog only.
  • Qwen portal uses catalog, auth, and refreshOAuth.
  • MiniMax and Xiaomi use catalog plus usage hooks because their /usage behavior is plugin-owned even though inference still runs through the shared transports.

Runtime helpers

Plugins can access selected core helpers via api.runtime. For TTS:

const clip = await api.runtime.tts.textToSpeech({
  text: "Hello from OpenClaw",
  cfg: api.config,
});

const result = await api.runtime.tts.textToSpeechTelephony({
  text: "Hello from OpenClaw",
  cfg: api.config,
});

const voices = await api.runtime.tts.listVoices({
  provider: "elevenlabs",
  cfg: api.config,
});

Notes:

  • textToSpeech returns the normal core TTS output payload for file/voice-note surfaces.
  • Uses core messages.tts configuration and provider selection.
  • Returns PCM audio buffer + sample rate. Plugins must resample/encode for providers.
  • listVoices is optional per provider. Use it for vendor-owned voice pickers or setup flows.
  • Voice listings can include richer metadata such as locale, gender, and personality tags for provider-aware pickers.
  • OpenAI and ElevenLabs support telephony today. Microsoft does not.

Plugins can also register speech providers via api.registerSpeechProvider(...).

api.registerSpeechProvider({
  id: "acme-speech",
  label: "Acme Speech",
  isConfigured: ({ config }) => Boolean(config.messages?.tts),
  synthesize: async (req) => {
    return {
      audioBuffer: Buffer.from([]),
      outputFormat: "mp3",
      fileExtension: ".mp3",
      voiceCompatible: false,
    };
  },
});

Notes:

  • Keep TTS policy, fallback, and reply delivery in core.
  • Use speech providers for vendor-owned synthesis behavior.
  • Legacy Microsoft edge input is normalized to the microsoft provider id.
  • The preferred ownership model is company-oriented: one vendor plugin can own text, speech, image, and future media providers as OpenClaw adds those capability contracts.

For image/audio/video understanding, plugins register one typed media-understanding provider instead of a generic key/value bag:

api.registerMediaUnderstandingProvider({
  id: "google",
  capabilities: ["image", "audio", "video"],
  describeImage: async (req) => ({ text: "..." }),
  transcribeAudio: async (req) => ({ text: "..." }),
  describeVideo: async (req) => ({ text: "..." }),
});

Notes:

  • Keep orchestration, fallback, config, and channel wiring in core.
  • Keep vendor behavior in the provider plugin.
  • Additive expansion should stay typed: new optional methods, new optional result fields, new optional capabilities.
  • If OpenClaw adds a new capability such as video generation later, define the core capability contract first, then let vendor plugins register against it.

For media-understanding runtime helpers, plugins can call:

const image = await api.runtime.mediaUnderstanding.describeImageFile({
  filePath: "/tmp/inbound-photo.jpg",
  cfg: api.config,
  agentDir: "/tmp/agent",
});

const video = await api.runtime.mediaUnderstanding.describeVideoFile({
  filePath: "/tmp/inbound-video.mp4",
  cfg: api.config,
});

For audio transcription, plugins can use either the media-understanding runtime or the older STT alias:

const { text } = await api.runtime.mediaUnderstanding.transcribeAudioFile({
  filePath: "/tmp/inbound-audio.ogg",
  cfg: api.config,
  // Optional when MIME cannot be inferred reliably:
  mime: "audio/ogg",
});

Notes:

  • api.runtime.mediaUnderstanding.* is the preferred shared surface for image/audio/video understanding.
  • Uses core media-understanding audio configuration (tools.media.audio) and provider fallback order.
  • Returns { text: undefined } when no transcription output is produced (for example skipped/unsupported input).
  • api.runtime.stt.transcribeAudioFile(...) remains as a compatibility alias.

Plugins can also launch background subagent runs through api.runtime.subagent:

const result = await api.runtime.subagent.run({
  sessionKey: "agent:main:subagent:search-helper",
  message: "Expand this query into focused follow-up searches.",
  provider: "openai",
  model: "gpt-4.1-mini",
  deliver: false,
});

Notes:

  • provider and model are optional per-run overrides, not persistent session changes.
  • OpenClaw only honors those override fields for trusted callers.
  • For plugin-owned fallback runs, operators must opt in with plugins.entries.<id>.subagent.allowModelOverride: true.
  • Use plugins.entries.<id>.subagent.allowedModels to restrict trusted plugins to specific canonical provider/model targets, or "*" to allow any target explicitly.
  • Untrusted plugin subagent runs still work, but override requests are rejected instead of silently falling back.

For web search, plugins can consume the shared runtime helper instead of reaching into the agent tool wiring:

const providers = api.runtime.webSearch.listProviders({
  config: api.config,
});

const result = await api.runtime.webSearch.search({
  config: api.config,
  args: {
    query: "OpenClaw plugin runtime helpers",
    count: 5,
  },
});

Plugins can also register web-search providers via api.registerWebSearchProvider(...).

Notes:

  • Keep provider selection, credential resolution, and shared request semantics in core.
  • Use web-search providers for vendor-specific search transports.
  • api.runtime.webSearch.* is the preferred shared surface for feature/channel plugins that need search behavior without depending on the agent tool wrapper.

Gateway HTTP routes

Plugins can expose HTTP endpoints with api.registerHttpRoute(...).

api.registerHttpRoute({
  path: "/acme/webhook",
  auth: "plugin",
  match: "exact",
  handler: async (_req, res) => {
    res.statusCode = 200;
    res.end("ok");
    return true;
  },
});

Route fields:

  • path: route path under the gateway HTTP server.
  • auth: required. Use "gateway" to require normal gateway auth, or "plugin" for plugin-managed auth/webhook verification.
  • match: optional. "exact" (default) or "prefix".
  • replaceExisting: optional. Allows the same plugin to replace its own existing route registration.
  • handler: return true when the route handled the request.

Notes:

  • api.registerHttpHandler(...) is obsolete. Use api.registerHttpRoute(...).
  • Plugin routes must declare auth explicitly.
  • Exact path + match conflicts are rejected unless replaceExisting: true, and one plugin cannot replace another plugin's route.
  • Overlapping routes with different auth levels are rejected. Keep exact/prefix fallthrough chains on the same auth level only.

Plugin SDK import paths

Use SDK subpaths instead of the monolithic openclaw/plugin-sdk import when authoring plugins:

  • openclaw/plugin-sdk/core for the smallest generic plugin-facing contract. It also carries small assembly helpers such as definePluginEntry, defineChannelPluginEntry, defineSetupPluginEntry, and createChannelPluginBase for bundled or third-party plugin entry wiring.
  • Domain subpaths such as openclaw/plugin-sdk/channel-config-helpers, openclaw/plugin-sdk/channel-config-schema, openclaw/plugin-sdk/channel-policy, openclaw/plugin-sdk/channel-runtime, openclaw/plugin-sdk/config-runtime, openclaw/plugin-sdk/agent-runtime, openclaw/plugin-sdk/lazy-runtime, openclaw/plugin-sdk/reply-history, openclaw/plugin-sdk/routing, openclaw/plugin-sdk/runtime-store, and openclaw/plugin-sdk/directory-runtime for shared runtime/config helpers.
  • Narrow channel-core subpaths such as openclaw/plugin-sdk/discord-core, openclaw/plugin-sdk/telegram-core, openclaw/plugin-sdk/whatsapp-core, and openclaw/plugin-sdk/line-core for channel-specific primitives that should stay smaller than the full channel helper barrels.
  • openclaw/plugin-sdk/compat remains as a legacy migration surface for older external plugins. Bundled plugins should not use it, and non-test imports emit a one-time deprecation warning outside test environments.
  • Bundled extension internals remain private. External plugins should use only openclaw/plugin-sdk/* subpaths. OpenClaw core/test code may use the repo public entry points under extensions/<id>/index.js, api.js, runtime-api.js, setup-entry.js, and narrowly scoped files such as login-qr-api.js. Never import extensions/<id>/src/* from core or from another extension.
  • Repo entry point split: extensions/<id>/api.js is the helper/types barrel, extensions/<id>/runtime-api.js is the runtime-only barrel, extensions/<id>/index.js is the bundled plugin entry, and extensions/<id>/setup-entry.js is the setup plugin entry.
  • openclaw/plugin-sdk/telegram for Telegram channel plugin types and shared channel-facing helpers. Built-in Telegram implementation internals stay private to the bundled extension.
  • openclaw/plugin-sdk/discord for Discord channel plugin types and shared channel-facing helpers. Built-in Discord implementation internals stay private to the bundled extension.
  • openclaw/plugin-sdk/slack for Slack channel plugin types and shared channel-facing helpers. Built-in Slack implementation internals stay private to the bundled extension.
  • openclaw/plugin-sdk/signal for Signal channel plugin types and shared channel-facing helpers. Built-in Signal implementation internals stay private to the bundled extension.
  • openclaw/plugin-sdk/imessage for iMessage channel plugin types and shared channel-facing helpers. Built-in iMessage implementation internals stay private to the bundled extension.
  • openclaw/plugin-sdk/whatsapp for WhatsApp channel plugin types and shared channel-facing helpers. Built-in WhatsApp implementation internals stay private to the bundled extension.
  • openclaw/plugin-sdk/line for LINE channel plugins.
  • openclaw/plugin-sdk/msteams for the bundled Microsoft Teams plugin surface.
  • Additional bundled extension-specific subpaths remain available where OpenClaw intentionally exposes extension-facing helpers: openclaw/plugin-sdk/acpx, openclaw/plugin-sdk/bluebubbles, openclaw/plugin-sdk/feishu, openclaw/plugin-sdk/googlechat, openclaw/plugin-sdk/irc, openclaw/plugin-sdk/lobster, openclaw/plugin-sdk/matrix, openclaw/plugin-sdk/mattermost, openclaw/plugin-sdk/memory-core, openclaw/plugin-sdk/minimax-portal-auth, openclaw/plugin-sdk/nextcloud-talk, openclaw/plugin-sdk/nostr, openclaw/plugin-sdk/synology-chat, openclaw/plugin-sdk/test-utils, openclaw/plugin-sdk/tlon, openclaw/plugin-sdk/twitch, openclaw/plugin-sdk/voice-call, openclaw/plugin-sdk/zalo, and openclaw/plugin-sdk/zalouser.

Compatibility note:

  • openclaw/plugin-sdk remains supported for existing external plugins.
  • New and migrated bundled plugins should use channel or extension-specific subpaths; use core plus explicit domain subpaths for generic surfaces, and treat compat as migration-only.
  • Capability-specific subpaths such as image-generation, media-understanding, and speech exist because bundled/native plugins use them today. Their presence does not by itself mean every exported helper is a long-term frozen external contract.

Channel target resolution

Channel plugins should own channel-specific target semantics. Keep the shared outbound host generic and use the messaging adapter surface for provider rules:

  • messaging.inferTargetChatType({ to }) decides whether a normalized target should be treated as direct, group, or channel before directory lookup.
  • messaging.targetResolver.looksLikeId(raw, normalized) tells core whether an input should skip straight to id-like resolution instead of directory search.
  • messaging.targetResolver.resolveTarget(...) is the plugin fallback when core needs a final provider-owned resolution after normalization or after a directory miss.
  • messaging.resolveOutboundSessionRoute(...) owns provider-specific session route construction once a target is resolved.

Recommended split:

  • Use inferTargetChatType for category decisions that should happen before searching peers/groups.
  • Use looksLikeId for "treat this as an explicit/native target id" checks.
  • Use resolveTarget for provider-specific normalization fallback, not for broad directory search.
  • Keep provider-native ids like chat ids, thread ids, JIDs, handles, and room ids inside target values or provider-specific params, not in generic SDK fields.

Config-backed directories

Plugins that derive directory entries from config should keep that logic in the plugin and reuse the shared helpers from openclaw/plugin-sdk/directory-runtime.

Use this when a channel needs config-backed peers/groups such as:

  • allowlist-driven DM peers
  • configured channel/group maps
  • account-scoped static directory fallbacks

The shared helpers in directory-runtime only handle generic operations:

  • query filtering
  • limit application
  • deduping/normalization helpers
  • building ChannelDirectoryEntry[]

Channel-specific account inspection and id normalization should stay in the plugin implementation.

Provider catalogs

Provider plugins can define model catalogs for inference with registerProvider({ catalog: { run(...) { ... } } }).

catalog.run(...) returns the same shape OpenClaw writes into models.providers:

  • { provider } for one provider entry
  • { providers } for multiple provider entries

Use catalog when the plugin owns provider-specific model ids, base URL defaults, or auth-gated model metadata.

catalog.order controls when a plugin's catalog merges relative to OpenClaw's built-in implicit providers:

  • simple: plain API-key or env-driven providers
  • profile: providers that appear when auth profiles exist
  • paired: providers that synthesize multiple related provider entries
  • late: last pass, after other implicit providers

Later providers win on key collision, so plugins can intentionally override a built-in provider entry with the same provider id.

Compatibility:

  • discovery still works as a legacy alias
  • if both catalog and discovery are registered, OpenClaw uses catalog

Read-only channel inspection

If your plugin registers a channel, prefer implementing plugin.config.inspectAccount(cfg, accountId) alongside resolveAccount(...).

Why:

  • resolveAccount(...) is the runtime path. It is allowed to assume credentials are fully materialized and can fail fast when required secrets are missing.
  • Read-only command paths such as openclaw status, openclaw status --all, openclaw channels status, openclaw channels resolve, and doctor/config repair flows should not need to materialize runtime credentials just to describe configuration.

Recommended inspectAccount(...) behavior:

  • Return descriptive account state only.
  • Preserve enabled and configured.
  • Include credential source/status fields when relevant, such as:
    • tokenSource, tokenStatus
    • botTokenSource, botTokenStatus
    • appTokenSource, appTokenStatus
    • signingSecretSource, signingSecretStatus
  • You do not need to return raw token values just to report read-only availability. Returning tokenStatus: "available" (and the matching source field) is enough for status-style commands.
  • Use configured_unavailable when a credential is configured via SecretRef but unavailable in the current command path.

This lets read-only commands report "configured but unavailable in this command path" instead of crashing or misreporting the account as not configured.

Package packs

A plugin directory may include a package.json with openclaw.extensions:

{
  "name": "my-pack",
  "openclaw": {
    "extensions": ["./src/safety.ts", "./src/tools.ts"],
    "setupEntry": "./src/setup-entry.ts"
  }
}

Each entry becomes a plugin. If the pack lists multiple extensions, the plugin id becomes name/<fileBase>.

If your plugin imports npm deps, install them in that directory so node_modules is available (npm install / pnpm install).

Security guardrail: every openclaw.extensions entry must stay inside the plugin directory after symlink resolution. Entries that escape the package directory are rejected.

Security note: openclaw plugins install installs plugin dependencies with npm install --ignore-scripts (no lifecycle scripts). Keep plugin dependency trees "pure JS/TS" and avoid packages that require postinstall builds.

Optional: openclaw.setupEntry can point at a lightweight setup-only module. When OpenClaw needs setup surfaces for a disabled channel plugin, or when a channel plugin is enabled but still unconfigured, it loads setupEntry instead of the full plugin entry. This keeps startup and setup lighter when your main plugin entry also wires tools, hooks, or other runtime-only code.

Optional: openclaw.startup.deferConfiguredChannelFullLoadUntilAfterListen can opt a channel plugin into the same setupEntry path during the gateway's pre-listen startup phase, even when the channel is already configured.

Use this only when setupEntry fully covers the startup surface that must exist before the gateway starts listening. In practice, that means the setup entry must register every channel-owned capability that startup depends on, such as:

  • channel registration itself
  • any HTTP routes that must be available before the gateway starts listening
  • any gateway methods, tools, or services that must exist during that same window

If your full entry still owns any required startup capability, do not enable this flag. Keep the plugin on the default behavior and let OpenClaw load the full entry during startup.

Example:

{
  "name": "@scope/my-channel",
  "openclaw": {
    "extensions": ["./index.ts"],
    "setupEntry": "./setup-entry.ts",
    "startup": {
      "deferConfiguredChannelFullLoadUntilAfterListen": true
    }
  }
}

Channel catalog metadata

Channel plugins can advertise setup/discovery metadata via openclaw.channel and install hints via openclaw.install. This keeps the core catalog data-free.

Example:

{
  "name": "@openclaw/nextcloud-talk",
  "openclaw": {
    "extensions": ["./index.ts"],
    "channel": {
      "id": "nextcloud-talk",
      "label": "Nextcloud Talk",
      "selectionLabel": "Nextcloud Talk (self-hosted)",
      "docsPath": "/channels/nextcloud-talk",
      "docsLabel": "nextcloud-talk",
      "blurb": "Self-hosted chat via Nextcloud Talk webhook bots.",
      "order": 65,
      "aliases": ["nc-talk", "nc"]
    },
    "install": {
      "npmSpec": "@openclaw/nextcloud-talk",
      "localPath": "extensions/nextcloud-talk",
      "defaultChoice": "npm"
    }
  }
}

OpenClaw can also merge external channel catalogs (for example, an MPM registry export). Drop a JSON file at one of:

  • ~/.openclaw/mpm/plugins.json
  • ~/.openclaw/mpm/catalog.json
  • ~/.openclaw/plugins/catalog.json

Or point OPENCLAW_PLUGIN_CATALOG_PATHS (or OPENCLAW_MPM_CATALOG_PATHS) at one or more JSON files (comma/semicolon/PATH-delimited). Each file should contain { "entries": [ { "name": "@scope/pkg", "openclaw": { "channel": {...}, "install": {...} } } ] }.

Context engine plugins

Context engine plugins own session context orchestration for ingest, assembly, and compaction. Register them from your plugin with api.registerContextEngine(id, factory), then select the active engine with plugins.slots.contextEngine.

Use this when your plugin needs to replace or extend the default context pipeline rather than just add memory search or hooks.

export default function (api) {
  api.registerContextEngine("lossless-claw", () => ({
    info: { id: "lossless-claw", name: "Lossless Claw", ownsCompaction: true },
    async ingest() {
      return { ingested: true };
    },
    async assemble({ messages }) {
      return { messages, estimatedTokens: 0 };
    },
    async compact() {
      return { ok: true, compacted: false };
    },
  }));
}

If your engine does not own the compaction algorithm, keep compact() implemented and delegate it explicitly:

import { delegateCompactionToRuntime } from "openclaw/plugin-sdk/core";

export default function (api) {
  api.registerContextEngine("my-memory-engine", () => ({
    info: {
      id: "my-memory-engine",
      name: "My Memory Engine",
      ownsCompaction: false,
    },
    async ingest() {
      return { ingested: true };
    },
    async assemble({ messages }) {
      return { messages, estimatedTokens: 0 };
    },
    async compact(params) {
      return await delegateCompactionToRuntime(params);
    },
  }));
}

Adding a new capability

When a plugin needs behavior that does not fit the current API, do not bypass the plugin system with a private reach-in. Add the missing capability.

Recommended sequence:

  1. define the core contract Decide what shared behavior core should own: policy, fallback, config merge, lifecycle, channel-facing semantics, and runtime helper shape.
  2. add typed plugin registration/runtime surfaces Extend OpenClawPluginApi and/or api.runtime with the smallest useful typed capability surface.
  3. wire core + channel/feature consumers Channels and feature plugins should consume the new capability through core, not by importing a vendor implementation directly.
  4. register vendor implementations Vendor plugins then register their backends against the capability.
  5. add contract coverage Add tests so ownership and registration shape stay explicit over time.

This is how OpenClaw stays opinionated without becoming hardcoded to one provider's worldview. See the Capability Cookbook for a concrete file checklist and worked example.

Capability checklist

When you add a new capability, the implementation should usually touch these surfaces together:

  • core contract types in src/<capability>/types.ts
  • core runner/runtime helper in src/<capability>/runtime.ts
  • plugin API registration surface in src/plugins/types.ts
  • plugin registry wiring in src/plugins/registry.ts
  • plugin runtime exposure in src/plugins/runtime/* when feature/channel plugins need to consume it
  • capture/test helpers in src/test-utils/plugin-registration.ts
  • ownership/contract assertions in src/plugins/contracts/registry.ts
  • operator/plugin docs in docs/

If one of those surfaces is missing, that is usually a sign the capability is not fully integrated yet.

Capability template

Minimal pattern:

// core contract
export type VideoGenerationProviderPlugin = {
  id: string;
  label: string;
  generateVideo: (req: VideoGenerationRequest) => Promise<VideoGenerationResult>;
};

// plugin API
api.registerVideoGenerationProvider({
  id: "openai",
  label: "OpenAI",
  async generateVideo(req) {
    return await generateOpenAiVideo(req);
  },
});

// shared runtime helper for feature/channel plugins
const clip = await api.runtime.videoGeneration.generateFile({
  prompt: "Show the robot walking through the lab.",
  cfg,
});

Contract test pattern:

expect(findVideoGenerationProviderIdsForPlugin("openai")).toEqual(["openai"]);

That keeps the rule simple:

  • core owns the capability contract + orchestration
  • vendor plugins own vendor implementations
  • feature/channel plugins consume runtime helpers
  • contract tests keep ownership explicit