@AgentValet
基本情報
IGA for AI agents. IETF AIMS-compliant identity, credential governance & MCP proxy. SPIFFE · RFC 7591 · AuthZEN · CIBA. npx @agentvalet/register
以下の設定を使って、このサーバーを MCP 対応クライアントに追加してください。
{
"mcpServers": {
"agentvalet": {
"command": "npx",
"args": [
"-y",
"@agentvalet/mcp-server"
],
"env": {
"AGENT_ID": "agt_...",
"OWNER_ID": "...",
"PROXY_URL": "https://api.agentvalet.ai",
"AGENT_PRIVATE_KEY_PATH": "/path/to/agent.key"
}
}
}
}list_platforms: List the platforms and permission scopes this agent has access to. Input: None. Returns: { platforms: [{ platformId, platformName, scopes, requireApproval, connections? }], version: "<hex>" }. A platform includes `connections` only when it has 2+ usable accounts; pass the chosen connection_id to use_platform (omit for the default). Version: a deterministic hash over the effective spec — it changes whenever the platform set, the scopes, or the requireApproval flags change. Compare it against the `Current permissions version` stamped on this description: if they differ, your cached result is stale and you must call list_platforms again before trusting it. Auth: Bearer JWT.
use_platform: Call an external platform API (Airtable, GitHub, Slack, Metabase, etc.) through the AgentValet proxy. Input: platform (string), endpoint (string), method (GET|POST|PUT|PATCH|DELETE), scope (string), body (object, optional — JSON request body for POST/PUT/PATCH/DELETE), reason (string, optional — why you are making this call; shown to the human approver when the action is gated). Returns: upstream API response body. May take up to 50 seconds when the action requires owner approval — the call will block while we wait, then return the approved result transparently. If approval doesn't land in time, returns a `pending_approval` envelope and the action runs asynchronously; the user is notified when it completes. Auth: Bearer JWT. Note: legacy clients passing `data` instead of `body` are still accepted for backwards compatibility, but `body` is the canonical name.
agent_register: Self-register this agent with an owner. No auth required. Input: owner_id (string), agent_name (string), requested_scopes (array of {platformId, scopes}). Returns: registration_token, poll_url, client_id, scope, expires_in. Auth: None.
agent_status: Poll registration status using the token from agent_register. Input: token (string, required). Returns: status ("pending_approval"|"approved"|"rejected"), agent_id (if approved), mcp_config (if approved). Auth: None.
authzen_evaluate: Evaluate whether this agent has access to a specific platform scope. Call this BEFORE use_platform when you want to pre-check without making the upstream call. Input: platform_id (string), scope (string). Returns: decision (boolean), reason ("approved"|"denied"|"revoked"|"scope_not_granted"). Auth: Bearer agent JWT (sent automatically by this MCP server).
report_self_diagnostic: Lodge a self-report (error/warning/info) with the AgentValet owner. Use after a use_platform error returns a report_hint, OR proactively when you encounter a problem the user should know about. Input: severity (debug|info|warn|error|critical), message (string, required, max 4096 bytes), code (string, optional, max 128 chars), platform (string, optional), endpoint (string, optional), correlation_id (uuid string, optional — copy from the failing call's report_hint to stitch this report to the broker-side audit row), context (object, optional, JSON-serialised must be < 16 KiB). Returns: { id, received_at } on success. Auth: Bearer agent JWT (sent automatically).
list_my_pending_actions: Returns this agent's currently-pending approval requests AND any that completed in the last 24 hours. Use this at session start when the user mentions an earlier action, or when use_platform's long-poll timed out and the user comes back asking what happened. Input: None. Returns: { pending: [{approval_id, platform_id, scope, created_at, expires_at}], recently_completed: [{approval_id, platform_id, scope, status, executed_at, result_summary, execution_error}] }. Auth: Bearer agent JWT (sent automatically).
request_platform_access: Ask an org admin to grant this agent access to a platform it is currently blocked from. Call this when use_platform returns an access-denied error. Input: platform (string, required), scope (string, optional — the specific scope you need), reason (string, optional — why you need it). The call waits up to ~50s for an admin decision. Returns: { status: "approved"|"pending"|"denied", message }. On "approved", retry your original use_platform call. On "pending", the request is queued; retry later. Auth: Bearer agent JWT (sent automatically).
Agentvalet is an identity and credential governance broker for AI agents and MCP servers. It gives every agent its own cryptographic identity, scoped and short-lived credentials per platform, human approval gates on critical actions, and an immutable audit log. The open-source client surface (MCP server, CLI, Claude Desktop bundle, adapters) is MIT-licensed; the managed proxy runs the credential vault, policy engine, and audit store.
Run npx @agentvalet/register to generate an RS256 keypair, register the agent, and wire up the configuration. Alternatively, manually add the MCP server to your client’s config with environment variables AGENT_ID, OWNER_ID, PROXY_URL, and AGENT_PRIVATE_KEY_PATH. The server exposes seven tools: list_platforms, use_platform, authzen_evaluate, agent_register, agent_status, list_my_pending_actions, and report_self_diagnostic.
@agentvalet/register)Credential inheritance occurs when an AI agent runs on a developer’s token instead of its own. This gives every agent full scope on every platform the token touches and leaves no record of which agent did what. Agentvalet solves this by giving each agent its own scoped, short-lived credential.
Agentvalet focuses on agent‑first identity and governance, with self‑serve onboarding under five minutes and open‑source client components (MIT). Aembit and Akeyless are enterprise‑sales‑led; Infisical is open‑source core but network‑level. Agentvalet is built for developers and small teams shipping agents this week.
The client surface (MCP server, CLI, Claude Desktop bundle, adapters, and `@agentvalet/mcp‑broker`) is MIT‑licensed in this repo. The proxy (credential vault, policy engine, audit store) is a managed service today; a self‑host reference is on the roadmap.
The SSE stream for approval status can close prematurely on long waits (reconnect logic planned). CLI rate limiting is rudimentary. There is no clear or delete UI for expired pending registrations.
The MCP server communicates over SSE (streaming) and uses RS256 JWTs (60‑second lifetime) signed by the agent’s private key. Private keys never leave the machine. Every platform call goes through the proxy with scope checks, audit logging, and human approval where required.
A one-of-a-kind resume builder that keeps your privacy in mind. Completely secure, customizable, portable, open-source and free forever. Try it out today!
🚀 The fast, Pythonic way to build MCP servers and clients.
Yuppsy MCP: build and publish real estate websites from your AI assistant. Yuppsy lets your AI assistant create live websites for real estate agents and individual property listings. You don't need a designer or any code
A free remote MCP server for The Living Bread. Bible passages read word for word from a stored public domain text with an evidence label, keyword search and cross references, real churches, gatherings and prayer groups w
A search tool helps dev to solve the naming things problem.
コメント