AgentValet

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 the actions that matter, and an immutable audit log of everything it did.

Open core. This repo holds the open-source MIT-licensed client surface: the MCP server, the CLI, the Claude Desktop bundle, and the runtime adapters. The managed proxy runs the credential vault, the policy engine, and the audit store. A self-host reference for the proxy is on the roadmap.

Live at agentvalet.ai. Docs at docs.agentvalet.ai.

Quickstart

npx @agentvalet/register

That generates an RS256 keypair for your agent, registers it, and wires up the config. The private key never leaves your machine. From then on your agent signs a short-lived JWT per request and calls platforms through the AgentValet proxy. Approve the agent in the dashboard, grant it scopes, and you are running.

Use as an MCP server

npx @agentvalet/register writes this block into your client's MCP config for you. To wire it up by hand — in Claude Desktop, Claude Code, Cursor, or any MCP-compatible client — add:

{
  "mcpServers": {
    "agentvalet": {
      "command": "npx",
      "args": ["-y", "@agentvalet/mcp-server"],
      "env": {
        "AGENT_ID": "agt_your_agent_id",
        "OWNER_ID": "your_owner_id",
        "PROXY_URL": "https://api.agentvalet.ai",
        "AGENT_PRIVATE_KEY_PATH": "~/.agentvalet/agent.key"
      }
    }
  }
}

The server exposes seven tools: list_platforms, use_platform, authzen_evaluate, agent_register, agent_status, list_my_pending_actions, and report_self_diagnostic. The private key is read from AGENT_PRIVATE_KEY_PATH and never leaves your machine; every platform call goes through the proxy with scope checks, audit logging, and human approval on the actions that matter.

The problem: credential inheritance

Credential inheritance is what happens when an AI agent runs on your credentials instead of its own. The moment a token lands in .mcp.json or an environment variable, every agent in that project inherits the full reach of that token. It can do anything you can do, on every platform the token touches, and nothing records which agent did what.

How it works

Agent (holds its RS256 private key)
    |
    |  signs a 60-second JWT per request
    v
+-------------------------------------------+
|              AgentValet proxy             |
|  1. verify JWT signature                  |
|  2. check scope grant (deny by default)   |
|  3. scan request for leaked secrets       |
|  4. AuthZEN policy evaluation             |
|  5. human approval gate, if required      |
|  6. inject real credential in memory      |---> SaaS platform
|  7. append-only audit log entry           |
+-------------------------------------------+
    |
    v
Dashboard: approve registrations, manage scopes, review the audit log, monitor usage
MCP server: exposes AgentValet as tools for Claude and any MCP-compatible agent

Credentials use envelope encryption: a unique AES-256 data key per credential, wrapped by a master key held in a Key Vault HSM, decrypted in memory only at call time and never written to a log.

Features

• Per-agent RS256 cryptographic identity, SPIFFE-format URIs, 60-second signed JWTs
• Deny-by-default scopes, granular per-agent-per-platform-per-action grants
• Human-in-the-loop approval for destructive or financial scopes, one-click magic-link
• Immutable, append-only audit log, filterable and exportable
• Three-strike circuit breaker that auto-suspends a misbehaving agent
• Native MCP server plus a one-command CLI
• Self-hostable: PostgreSQL-backed, runs in your own infrastructure
• Standards-aligned: SPIFFE, RFC 7591 Dynamic Client Registration, AuthZEN 1.0, IETF AIMS

How AgentValet compares

Honest framing. These are strong tools that reached agent governance from an adjacent starting point.

* The client surface (MCP server, CLI, bundle, adapters) is MIT in this repo. The proxy is a managed service today, with a self-host reference on the roadmap.

If you already run Aembit or Akeyless at enterprise scale, AgentValet is not trying to replace your identity provider. If you are a developer or small team shipping agents this week, AgentValet is built for you.

Roadmap and known limitations

Building in public, so the rough edges are listed here rather than discovered.

Known limitations today

• The SSE stream for approval status can close prematurely on long waits. Reconnect logic is planned.
• CLI rate limiting is rudimentary.
• There is no clear or delete UI yet for expired pending registrations.

On the roadmap

• Self-host reference for the proxy (the open client surface already runs anywhere)
• SIEM export for the audit log (Enterprise)
• Multi-region self-hosting guides
• Custom integrations UI (today these are requested through the roadmap system)
• PII handling Phase 2: detection at the broker layer

What is open and what is managed

Open source in this repo, MIT licensed: the MCP server, the @agentvalet/register CLI, the Claude Desktop bundle, the runtime adapters, and @agentvalet/mcp-broker.

The two MCP packages are mirror images of each other: @agentvalet/mcp-server lets an agent call platforms through the AgentValet proxy, whereas @agentvalet/mcp-broker lets you embed AgentValet policy and credential enforcement inside your own MCP server, with a genuinely useful local mode and zero required account.

Managed service today: the proxy that holds the credential vault, runs the policy engine, and writes the audit log. A self-host reference for the proxy is on the roadmap. See CONTRIBUTING.md for local development of the open packages.

Security

Found a vulnerability? Please report it privately, see SECURITY.md. Do not open a public issue for security reports.

License

MIT. See LICENSE.