Agent Identity Plugin

UserPool OIDC login, TIP (Trusted Identity Provider) token via Identity GetWorkloadAccessTokenForJWT, credential 3LO (GetResourceOauth2Token/Oauth2Callback), and session management for OpenClaw.

中文文档请参阅 README-cn.md

Integrates with Volcengine Agent Identity and Permission Management.

Features

• OIDC Login: /identity login returns IdP auth URL (no HTTP start endpoint). User opens URL, IdP redirects to /identity/oauth/callback.
• Pre-LLM Auth Gate (before_dispatch): Blocks unauthenticated messages before the LLM is invoked, saving tokens. Returns an OIDC login URL directly in the static response. Falls back to before_agent_start on older OpenClaw versions.
• TIP Token: before_dispatch / before_agent_start hooks fetch TIP token when session has a logged-in user.
• Credential 3LO: /identity fetch <provider> returns auth URL. IdP redirects to Identity-provided callback (control-plane config).
• Credential Binding: /identity set <provider> <envVar> binds stored credential to env var. Credentials are securely injected per-tool-call, isolated between concurrent multi-user sessions.
• Local Identity Server (UDS): Optional HTTP-over-Unix-Domain-Socket server exposing TIP tokens, OIDC session tokens, and all plugin tools to other local processes. Supports peer credential checking (Linux SO_PEERCRED + /proc) and configurable process allowlist.
• Encrypted Session Storage: sessions.json is encrypted at rest (AES-256-GCM). Plaintext sessions from older versions are auto-migrated on first load.
• In-memory TIP Cache: TIP tokens are stored only in memory (no disk persistence). They are short-lived and re-obtained from the user's session token on demand.
• Dynamic UserPool: Resolve OIDC config by userPoolName + clientName (no manual clientId).
• Credentials: Load AK/SK from env, file, or STS AssumeRole (veadk-style).

HTTP Endpoints

Only the OIDC login callback is exposed. Credential OAuth uses Identity callback. All other logic runs in slash commands.

Slash Commands

Single command /identity (alias /id) with subcommands. Default with no args: status.

OIDC Login Flow

1. User sends /identity login in chat (e.g. Telegram, Discord)
2. Command derives sessionKey from channel/sender, builds IdP authorize URL, stores state
3. Command returns the IdP URL; user opens it in browser
4. User completes login at UserPool IdP
5. IdP redirects to /identity/oauth/callback with code and state
6. Plugin exchanges code, creates session, shows success page and sends message to chat

Credential Fetch Flow

OAuth2 (user federation or M2M):

1. User sends /identity fetch google or /identity fetch google --flow=oauth2-m2m (after /identity login)
2. Command uses TIP to call Identity API; returns auth URL or direct token
3. If auth URL: user opens it; IdP redirects to Identity callback (control-plane provider config)
4. Identity handles callback; token obtained via Identity; user may re-run fetch to pull credential

API Key:

1. User sends /identity fetch openai (provider type api_key in control plane) or /identity fetch openai --flow=apikey
2. Command uses TIP to call GetResourceApiKey; API key stored directly

Flow is auto-inferred from ListCredentialProviders (Type + Flow). Override with --flow=oauth2-user|oauth2-m2m|apikey when needed.

Enable Volcengine Agent Identity Service

Before using this plugin, you must enable Agent Identity and Permission Management in Volcengine and complete authorization. See:

• Enable Agent Identity Service (Chinese)

Get Volcengine AK/SK

The plugin requires Volcengine Access Key and Secret Key to call the Identity API. To create credentials:

• Create Access Key - Volcengine Docs

After creating AK/SK in the console, pass them via config or use environment variables VOLCENGINE_ACCESS_KEY and VOLCENGINE_SECRET_KEY.

Installation

openclaw plugins install @m1a0rz/agent-identity

Or with link for development:

openclaw plugins install --link .

Configuration

Step 2: Configure plugin (minimal setup)

The plugin typically needs three types of config:

A. Platform access (Identity): For TIP Token, credential fetch/hosting, and optional permission checks.

• endpoint (optional): Full Identity API base URL (e.g. https://id.cn-beijing.volcengineapi.com). Highest priority for the API host.
• regionMetadataUrl (optional): HTTP(S) URL that returns a plain-text region id (e.g. cn-beijing). Used only when endpoint is unset: the client builds https://id.{region}.volcengineapi.com. Request timeout ~10s; on failure or invalid body (e.g. unknown), falls back to https://id.cn-beijing.volcengineapi.com. Example metadata URL: http://100.96.0.96/latest/region_id (must be reachable from the gateway).
• accessKeyId / secretAccessKey: For Identity API access. Optional when using env vars or credential file (see below).
• workloadPoolName / workloadName: For issuing TIP Token. Defaults: default, openclaw-agent.
• audience / durationSeconds: Optional, token audience and validity.
• credentialsFile: Path to credential JSON. Default: VOLCENGINE_CREDENTIALS_FILE env or /var/run/secrets/iam/credential.
• credentialsMetadataUrl: Full URL for remote credential fetch. When set with roleTrn, fetches from URL (response: AccessKeyId, SecretAccessKey, SessionToken), then AssumeRole with roleTrn to get final credentials. Same flow as AK/SK + roleTrn. 404 falls through to credential file. Cached by ExpiredTime. Must be explicitly configured.
• roleTrn: Role TRN for STS AssumeRole. When set (and workloadName not set), workload name is omitted; backend uses roleName. Priority: workloadName > roleTrn > params. Used with credentialsMetadataUrl (AssumeRole after fetch) or explicit AK/SK.
• sessionToken: STS session token (or use VOLCENGINE_SESSION_TOKEN env).

Identity API host resolution: endpoint if set → else region from regionMetadataUrl → else https://id.cn-beijing.volcengineapi.com. SigV4 signing region is inferred from the resolved host when it matches id.{region}.volcengineapi.com; otherwise cn-beijing.

Workload pool (credentials / STS): workloadPoolName (default default) scopes ListCredentialProviders, ListRoleCredentialProviders, and control-plane calls such as GetResourceOauth2Token, GetResourceApiKey, GetUserCredential, and GetRoleCredentials via PoolName. Role provider listing also filters by userpool.userPoolName when configured.

Credential resolution order (AK/SK): 1) Explicit config → 2) Env vars (VOLCENGINE_ACCESS_KEY, VOLCENGINE_SECRET_KEY, VOLCENGINE_SESSION_TOKEN) → 3) Remote metadata (credentialsMetadataUrl + roleTrn, fetches from full URL then AssumeRole; 404 falls through) → 4) Credential file (credentialsFile config, or VOLCENGINE_CREDENTIALS_FILE env, or /var/run/secrets/iam/credential). Credential file format (VeFaaS): access_key_id, secret_access_key, session_token (optional), role_trn (optional for AssumeRole). RUNTIME_IAM_ROLE_TRN env can supply role TRN when loading from file.

B. User login (UserPool / OIDC): For /identity login and session setup.

• discoveryUrl (or userPoolName + clientName for dynamic resolution)
• clientId / clientSecret (auto-resolved in dynamic mode)
• callbackUrl: Public callback URL for OpenClaw gateway, e.g. http://127.0.0.1:18789/identity/oauth/callback
• scope: Typically openid profile email
• identityProvider (optional): IdP name for the identity_provider authorize param. When omitted, the first entry from ListIdentityProviders is used.
• useRelayCallback (optional): UserPool relay / redirect_relay_uri flow for OIDC callback. Default false.

C. AuthZ and risk approval (optional): For TIP + CheckPermission + risk evaluation. Each flag is independent; no single "enable" switch.

• agentCheck: Run CheckPermission for agents (resource type agent) in before_agent_start. Verifies the user can invoke the current agent. Uses the outermost actor from TIP delegation chain as resource id. Default false.
• toolCheck: Run CheckPermission for tools (resource type tool). Default false.
• skillReadCheck: Run CheckPermission for read of SKILL.md (resource type skill). Parses available_skills from system prompt. Default false.
• requireRiskApproval: Require user approval for high-risk tool calls. Default false.
• namespaceName: CheckPermission Cedar policy namespace. Default default.
• lowRiskBypass: Skip TIP+CheckPermission for built-in low-risk tools. Default true.
• lowRiskTools: Extra tool names treated as low-risk.
• enableLlmRiskCheck: Use LLM to re-evaluate when rules return medium. Default false.
• llmRiskCheck: LLM config (endpoint, api, model, apiKey, timeoutMs, cacheTtlMs). Required when enableLlmRiskCheck is true.
• approvalTtlSeconds: Approval link/command TTL (seconds). Default 300.

Expected outcome: After config, the plugin can initiate login and obtain TIP Token. With AuthZ flags enabled, agent/tool/skill permission checks and high-risk approvals apply. High-risk tool calls use OpenClaw's native requireApproval mechanism (UI overlay, Telegram buttons, Discord interactions, /approve CLI).

Add to openclaw.json under plugins.entries.agent-identity.config:

{
  "plugins": {
    "entries": {
      "agent-identity": {
        "config": {
          "identity": {
            "endpoint": "https://id.cn-beijing.volcengineapi.com",
            "workloadPoolName": "default",
            "workloadName": "openclaw-agent"
          },
          "userpool": {
            "discoveryUrl": "https://userpool-xxx.userpool.auth.id.cn-beijing.volces.com",
            "clientId": "<client-id>",
            "clientSecret": "<client-secret>",
            "callbackUrl": "https://gateway.example.com/identity/oauth/callback",
            "scope": "openid profile email"
          },
          "authz": {
            "agentCheck": false,
            "toolCheck": false,
            "skillReadCheck": false,
            "requireRiskApproval": false,
            "namespaceName": "default",
            "lowRiskBypass": true,
            "enableLlmRiskCheck": false,
            "approvalTtlSeconds": 300
          },
          "localServer": false,
          "localServerAllowlist": [],
          "localServerFailOpen": true
        }
      }
    }
  }
}

Identity credentials: Omit accessKeyId/secretAccessKey to use env vars (VOLCENGINE_ACCESS_KEY, VOLCENGINE_SECRET_KEY) or credential file (VOLCENGINE_CREDENTIALS_FILE or /var/run/secrets/iam/credential).

identity config (required vs optional)

* AK/SK must be provided via accessKeyId+secretAccessKey, environment variables, credentialsMetadataUrl+roleTrn, or credentialsFile.

Environment variables: VOLCENGINE_ACCESS_KEY, VOLCENGINE_SECRET_KEY, VOLCENGINE_SESSION_TOKEN, VOLCENGINE_CREDENTIALS_FILE, RUNTIME_IAM_ROLE_TRN (for AssumeRole when loading from file). Set IDENTITY_STS_DEBUG=1 to log full STS AssumeRole request/response for debugging.

userpool config (OIDC login)

Explicit mode (required): discoveryUrl, clientId, clientSecret, callbackUrl, scope

Dynamic mode (required): userPoolName, clientName, callbackUrl; autoCreate defaults to true

OAuth2 credential fetch uses control-plane redirect URL and scopes. Override via /identity fetch <provider> --redirectUrl and --scopes.

authz config (optional, each flag independent)

Workload and TIP

TIP token is obtained via GetWorkloadAccessTokenForJWT. Workload behavior:

• workloadName (optional): Workload name sent to the API. Priority: config.workloadName > config.roleTrn > params (session agentId or "openclaw-agent"). When config.workloadName is set, it takes precedence over roleTrn.
• roleTrn (optional): When set (and workloadName not set), the plugin does not pass workload name; the backend uses the role name. Use this for delegated execution (e.g. VeFaaS, K8s IRSA).
• Auto-create workload: When GetWorkloadAccessTokenForJWT returns 404 (workload not found), the plugin calls CreateWorkloadIdentity to create the workload (Category: Agent), then retries. Only applies when a workload name is used (workloadName set or neither workloadName nor roleTrn set). Duplicated (409) from concurrent create is ignored.

Feishu notifications

Login success and credential fetch follow-up messages (e.g. "✓ Credential for google added.") are sent via Feishu when the user runs /identity from a Feishu chat. Credentials are read from channels.feishu in openclaw.json (same as feishu extension: appId, appSecret, optional accounts). No extra config in agent-identity is required.

Approval messages (when a high-risk tool is blocked): The plugin returns requireApproval from before_tool_call, and OpenClaw handles the approval UI natively across all channels (exec overlay, Telegram buttons, Discord interactions, /approve CLI). No plugin-side delivery logic is needed.

WebChat Session Exchange (Gateway WS Methods)

When identity.webchatSessionExchange is true, the plugin registers two gateway WebSocket methods for webchat clients to inject and retrieve session tokens without going through the OIDC redirect flow:

• senderId defaults to "openclaw-control-ui". The effective storage key is agent:main:main:user:<senderId> for main sessions.
• channel is optional; when the session originates from a sendable channel (feishu, telegram, etc.), pass it to enable per-channel-peer key promotion.

Both methods are restricted to webchat WS connections only (isWebchatConnect check). Non-webchat clients receive a FORBIDDEN error.

Config:

{
  "identity": {
    "webchatSessionExchange": true
  }
}

Typical flow (BFF → webchat → plugin):

1. BFF completes 3LO login and obtains an OIDC id_token for the user
2. Webchat client calls identity.session.put with the session key and id_token (optionally refreshToken from the token response if silent renewal is desired)
3. Plugin verifies the token, stores the session, and acquires TIP
4. Subsequent agent runs in that session have a valid identity — no manual login needed

WebChat / TUI

Follow-up messages (login success, credential fetch done) are not delivered when the user runs /identity from WebChat or TUI; the plugin has no API to push to those channels. Use /identity status to confirm results.

Tools

Agent-facing behavior is summarized in skills/SKILL.md. Registered tools:

• identity_whoami — Session identity (sub, TIP)
• identity_status — Login, TIP, credentials, bindings
• identity_login / identity_logout — OIDC login or refresh TIP; clear session
• identitylistcredentials — OAuth/API key providers and stored credentials (page, name, flow, type filters)
• identitylistroles — STS role credential providers (name prefix filter)
• identity_fetch — Add credential (provider, flow, redirectUrl, scopes, returnValue)
• identitygetrole_credentials — STS credentials for a role provider (providerName, useTip)
• identitygettip_token / identitygetsession_token — Raw TIP JWT or session user token (advanced)
• identity_config — Effective plugin config (redacted)
• identityconfigsuggest — Config merge snippets (intent, lang)
• identitysetbinding / identityunsetbinding — Env var bindings for tool injection
• identityriskcheck / identitylistrisk_patterns — Risk evaluation (optional plugin)
• identitylisttips — All valid TIP tokens and bindings (ops / multi-session debug)

Hooks

• before_dispatch - Stateless pre-LLM authentication gate. Checks for a valid TIP token before the LLM is invoked. If no TIP is found, blocks the message with a static response containing an OIDC login URL. Zero LLM token cost for unauthenticated users. When authz.agentCheck is enabled, also runs agent-level CheckPermission. Sets a feature flag so before_agent_start skips redundant auth logic.
• beforeagentstart - Fallback auth gate for older OpenClaw versions that do not support before_dispatch. When before_dispatch is active, this hook is a no-op for authentication. Still fetches TIP token and runs agent CheckPermission when applicable.
• subagent_spawned - Propagate TIP to child session on subagent spawn.
• beforetoolcall - Group context injection, optional AuthZ (TIP check, CheckPermission, risk via requireApproval), and per-tool-call credential injection. High-risk tool calls return requireApproval to the OpenClaw framework, which handles approval UI natively.
• aftertoolcall - Clean up per-tool-call credential injection state.

Local Identity Server (UDS)

When identity.localServer is true, the plugin starts an HTTP server over a Unix Domain Socket, allowing other local processes (scripts, plugins, CLIs) to access identity tokens and execute plugin tools without going through OpenClaw's messaging layer.

Configuration

Socket path: ~/.openclaw/plugins/identity/identity.sock (permissions 0600, owner-only access).

Endpoints

Peer Credential Checking (Linux)

On Linux, the server validates connecting processes via:

1. SO_PEERCRED (kernel-level): Zero-overhead retrieval of peer PID/UID/GID via getsockopt. Requires a registered native provider.
2. /proc filesystem: Derives process name and path from /proc/<pid>/exe, /proc/<pid>/comm, /proc/<pid>/status. No process spawning.

Resolved peer info is checked against the built-in allowlist (curl) plus any custom entries in localServerAllowlist. When localServerFailOpen is true (default), unresolvable peers are allowed (socket file permissions provide baseline security).

Usage Examples

# Get TIP token
curl --unix-socket ~/.openclaw/plugins/identity/identity.sock http://localhost/token

# Get OIDC session token
curl --unix-socket ~/.openclaw/plugins/identity/identity.sock http://localhost/session

# List available tools
curl --unix-socket ~/.openclaw/plugins/identity/identity.sock http://localhost/tools

# Execute a tool
curl --unix-socket ~/.openclaw/plugins/identity/identity.sock \
  -X POST http://localhost/tool/identity_whoami \
  -H 'Content-Type: application/json' \
  -d '{"params": {}, "session": ""}'

See demo/local-server/ for examples in Python, Go, TypeScript, Java, and Rust.

Data Storage

Plugin data at ~/.openclaw/plugins/identity/:

In-memory only (not persisted to disk):