Is mnki-mcp an MCP server with its own tools?
No. It is a proxy. It adds no tools of its own; it verifies tool calls to the server it wraps and forwards everything else.

@MNKIAgentOS
Basic information
Agent Trust — open profile + reference implementation for proving what an autonomous AI agent is authorized to do: identity, delegation with attenuation, proof-of-possession, attestations, tamper-evident provenance. Buil






Add this server to your MCP-compatible client using the configuration below.
{
"mcpServers": {
"agent-trust": {
"command": "npx",
"args": [
"mnki-cli",
"demo",
"#",
"the",
"15-second",
"story,",
"no",
"account:",
"a",
"€47,000",
"refund",
"denied,",
"€420",
"allowed"
]
}
}
}Tool list not available yet
This listing does not currently have a tool list to display. This does not mean the server has no tools; check its documentation for supported capabilities.
Fetch the live tool list by running this server in a temporary sandbox using the button above.
Agent Trust (mnki-mcp) is a local MCP proxy that sits between your MCP client and any MCP server and checks every tools/call against an agent's identity, delegated authority and policy before it reaches the server, returning signed evidence for each decision.
It is part of Agent Trust by MNKI, an open profile and reference implementation for proving what an autonomous AI agent is authorized to do: identity, delegation with attenuation, proof-of-possession, attestations and tamper-evident provenance, built on JWS, OAuth/OIDC, SPIFFE/WIMSE, AuthZEN, OpenID Federation, MCP and A2A. The proxy, CLI, SDKs and verifier are open source under Apache-2.0 and work in local mode with no account. A hosted console at mnki.com adds approvals, a ledger and a fleet view.
You need Node.js. Start with the demo, which needs no account:
npx mnki-cli demo
An invoice agent may refund up to €5,000: €47,000 is denied with the evidence list, €420 is allowed and €3,200 needs a human.
To protect the MCP servers you already run, scan the machine, then let mnki protect rewrite your client config. It is a dry run by default and starts in observe mode:
npx mnki-cli scan
npx mnki-cli protect --client cursor --server github --agent my-agent
npx mnki-cli protect --client cursor --server github --agent my-agent --apply
npx mnki-cli protect report --server github
mnki protect rollback undoes the change. Add --mode enforce when you are ready to block calls.
By hand, a stdio server entry wraps the original command after --upstream --:
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "mnki-mcp", "--agent", "agt_your_agent", "--server", "github", "--mode", "observe", "--upstream", "--", "npx", "-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "YOUR_GITHUB_TOKEN", "MNKI_URL": "https://mnki.com", "MNKI_API_KEY": "YOUR_MNKI_API_KEY" }
}
}
}
MNKI_URL and MNKI_API_KEY (or mnki init) tell the proxy which console verifies; MNKI_IDENTITY_FILE points at the agent key for signed requests.
Add the same entry as JSON:
claude mcp add-json github '{"command":"npx","args":["-y","mnki-mcp","--agent","my-agent","--server","github","--mode","observe","--upstream","--","npx","-y","@modelcontextprotocol/server-github"]}'
For a remote HTTP server, point the proxy at the upstream URL:
{
"mcpServers": {
"billing": {
"command": "npx",
"args": ["-y", "mnki-mcp", "--agent", "agt_your_agent", "--server", "billing", "--mode", "enforce", "--amount-map", "create_payment=amount:currency", "--upstream-url", "https://mcp.example.com/mcp", "--upstream-header", "Authorization=Bearer YOUR_UPSTREAM_TOKEN"]
}
}
}
{
"servers": {
"playwright": {
"type": "stdio",
"command": "npx",
"args": ["-y", "mnki-mcp", "--agent", "agt_your_agent", "--server", "playwright", "--mode", "warn", "--upstream", "--", "npx", "@playwright/mcp@latest"]
}
}
}
The repository also has a Windsurf snippet and a Docker sidecar image, ghcr.io/mnkiagentos/mnki-mcp.
tools/call and forwards everything else untouched.observe (log only, the default), warn, require_approval and enforce.-32003 denied with evidence, -32001 approval required, rejected or expired, -32005 verification quota exceeded and -32006 Agent Trust unreachable.pip install mnki) and Go, checked by open conformance vectors for levels 1 to 3.mnki protect report to see what would have been denied.create_payment amounts with --amount-map and enforce a refund limit, so €47,000 is denied and €420 runs.require_approval or enforce mode, risky calls wait for a person to approve from the console.npx mnki-cli scan lists every MCP server and API key agents can use on a machine.observe, which only logs. Switch to --mode enforce once the report looks right.-32006 errors. Agent Trust is unreachable. In observe and warn modes the proxy forwards instead; check MNKI_URL and your network.-32005 errors. You hit the verification quota. The free cloud plan includes 10,000 hosted verifications a month.mnki protect shows a diff first; use mnki protect rollback to undo an applied change.No. It is a proxy. It adds no tools of its own; it verifies tool calls to the server it wraps and forwards everything else.
The proxy, CLI, SDKs and verifier are Apache-2.0 open source and work locally with no account. The hosted console has a free tier with 10,000 verifications a month, plus paid plans.
No for local mode and the demo. Connecting the console at mnki.com needs an API key from Settings, API keys.
The docs list Claude Desktop, Claude Code, Cursor, Windsurf and VS Code, with ready-made snippets for Claude Desktop, Cursor, Windsurf and VS Code.
Yes. Use `--upstream-url` and `--upstream-header` to wrap a Streamable HTTP endpoint.
Yes. `mnki protect rollback` restores the client config.
Set up and manage the Eddyter AI rich text editor by talking to your AI agent — for Claude, Cursor, Codex and other MCP tools. A Model Context Protocol server that lets AI tools add Eddyter to a React, Next.js, Vite or R

Automated codebase documentation for GitHub: searchable docs, MCP context for AI agents, doc drift detection, and Friday Cleanup PRs.

DrUrls audits websites for SEO, and its MCP server lets Claude, ChatGPT, Claude Code, Cursor or VS Code work with those audits directly: list your sites with their latest score and grade, read each issue across SEO, perf
Query your team's drift, vulnerability, and migration data from any AI assistant. Vibgrate MCP connects Cursor, Claude, ChatGPT, Windsurf, or VS Code to Vibgrate Cloud: 51 tools for DriftScores, CVEs and EOL runtimes, up
Modelglass turns "which model should I use?" into a lookup instead of a guess. Seven tools give your agent live pricing across image, language, video, and audio models. Sourced and versioned, not scraped once and forgott
Comments