Agent MCP Host

MCP hosting with agent-mcp-host

agent-mcp-host is a CLI binary installed on $PATH. It is the operator tool of the agent-* family: it runs every mounted tool's MCP server behind ONE https origin, with ONE OAuth 2.1 authorization server and a separate login per tool. Each tool stays a full MCP server in its own binary — the host mounts it behind a path (/<name>/mcp) and reverse-proxies to it.

Run agent-mcp-host usage for the full LLM-optimized reference card.

Mental model

• One origin (e.g. https://hub.tailnet.example) fronts N tools.
• Each tool is a mount: --mount slack=agent-slack serves https://<host>/slack/mcp. A person adds one Claude connector per tool.
• The host owns pairing (who may connect), OAuth (token minting), and the browser enrollment pages. Tools validate the host's Ed25519 tokens and mint nothing themselves (delegate mode). A slack token is useless at /lin/mcp by construction.

Serve (long-running — run in the background)

agent-mcp-host serve --tailscale funnel \
    --mount slack=agent-slack --mount lin=lin

• --tailscale funnel|serve fronts the host with a Tailscale tunnel and derives --public-url from MagicDNS — no URL to figure out; the tunnel is torn down on exit (--tailscale-port 443|8443|10000).
• Without --tailscale, pass --public-url https://… (the OAuth issuer) and front the --http listener (default 127.0.0.1:8000) yourself.
• serve does not exit. Run it in the background and read stdout for NDJSON lifecycle events — {"event":"ready"} then one {"event":"mount_ready","tool":…,"url":…} per mount means it's up. Connector URLs and the pairing banner go to stderr.
• There is no separate step to start mounted tools: serve spawns each one as <binary> mcp --http 127.0.0.1:<port> --oauth <public-url> with its audience and verify key injected via env. The binaries just need to be installed (brew install shhac/tap/agent-slack); stopping the host stops them.

Attach mounts (run a tool yourself)

To run a tool under your own control (debugger, launchd) instead of having serve spawn it:

agent-mcp-host mount-env lin=lin --http 127.0.0.1:9410 --tailscale funnel
# → prints the exact launch command (env + flags) for that tool; run it, then:
agent-mcp-host serve --tailscale funnel --mount lin=lin@127.0.0.1:9410 ...

The binary name is still required (the host execs it for mcp schema and mcp enroll). mount-env's --public-url/--tailscale must match what serve runs with — --tailscale here only derives the URL; no tunnel starts.

People (pairing + per-tool credentials)

agent-mcp-host pair add alice --bind slack:workspace=acme --bind lin:workspace=acme
agent-mcp-host pair list              # principals + bindings (codes never shown)
agent-mcp-host pair show alice        # prints alice's code (a secret)
agent-mcp-host pair rotate alice      # fresh code, bindings preserved
agent-mcp-host pair remove alice      # revokes code + refresh tokens + sessions

Bindings are namespaced per tool (slack:workspace=acme); each tool's token carries only its own slice, prefix stripped. Without --bind for a tool that supports enrollment, the person enters their own credentials in the browser the first time they connect that tool (fields come from the tool's mcp schema; secrets go tool-ward via stdin, never argv).

How a person connects: (1) add the connector URL — the approval page opens; (2) enter the pairing code once, then the tool's enrollment form if no binding was provisioned; (3) the next tool skips the code (a 30-day browser session covers identity) and prompts only for that tool's enrollment.

Output contract

• stdout — NDJSON event stream, one line per lifecycle moment: ready, mount_ready, client_registered, paired, session_started, enrolled, authorized (with tool, principal, client, via, time fields as applicable). Events never contain secrets.
• stderr — the human boot banner (connector URLs + pairing code) and each tool's own stderr prefixed with its mount name.
• pair commands print human-readable text, not NDJSON.

Handling secrets

Pairing codes are passwords. pair add, pair rotate, and pair show print one, and the serve stderr banner shows the shared code — never quote a code into summaries, commit messages, or shared documents; share it only with the person it belongs to. If a code may have leaked, pair rotate <name>; to revoke a person entirely, pair remove <name>. Stored secrets (signing key, codes, sessions, refresh tokens) live in the keychain service app.paulie.agent-mcp-host.mcp and are never printed by list.

Global flags

• --format json|yaml|jsonl — output format
• --color auto|always|never, --timeout <ms>, --debug

References

• references/commands.md: full command map + all flags
• references/output.md: event stream + output shapes