Chirindo

A fail-closed cryptographic gate for the MCP tool-call boundary — the watchtower for your AI agents. By Headless Oracle.

Chirindo — Shona for "watchtower."

A stdio MCP proxy that intercepts tools/call requests from a real MCP client (Claude Desktop, Cursor), evaluates a policy, and either forwards the call to the real downstream server (ALLOW) or returns a tool-failure response WITHOUT forwarding (DENY) — emitting a signed receipt in either case.

Chirindo emits calibrated evidence: a verifying party can prove that the gate fired for a given call and that the chain is recomputable from the signed records. The receipts do not prove an action was "safe" — only that the gate's decision is captured, signed, and tamper-evident (not tamper-proof: any actor with the chain file can rewrite history, but a mutation breaks the hash chain and is caught by recorder verify). The receipt format and signing reuse the existing recorder engine — no reimplementation of JCS, hashing, or Ed25519.

Posture: fail-closed (the opposite of the recorder)

The recorder is observe-only: it never blocks the agent, even when its own signer crashes. The gate is the inverse: when it cannot evaluate policy, it denies. When it cannot write a receipt for an action that already ran, it withholds the result from the client (the action's side effect already happened; we cannot un-do it, but we can prevent the agent from acting on an un-receipted result).

Failure mode	Recorder	Gate
Signer throws	log + permit	DENY
Policy missing / invalid	n/a	DENY
Receipt write fails	log + permit	DENY result back to client

Architecture

MCP client (Claude Desktop)
        │
        ▼ stdio (JSON-RPC 2.0, newline-delimited)
   chirindo proxy   ◀── policy.json (deny rules)
        │
        ▼ stdio
  downstream MCP server (the real one)

The client launches chirindo proxy as its MCP server. The proxy spawns the real downstream MCP server as a child process. Every JSON-RPC frame in either direction passes through the proxy. tools/call requests are evaluated against the loaded policy:

• ALLOW → forward to downstream; on the response, write an ALLOW receipt and pass the response back to the client.
• DENY → synthesize a tool-execution-error response ({result: {content:[...], isError: true}}) per MCP spec § Error Handling, send it to the client, write a DENY receipt. The downstream NEVER sees the call.
• FAIL-CLOSED → if the policy can't be loaded or the evaluator throws, DENY the call.

All other frames (initialize, tools/list, responses, notifications) pass through unmodified.

The signed receipt

Each intercepted tools/call produces a single line appended to a per-session JSONL chain file. The receipt is a regular evidence.action/0 record from the recorder schema, with the gate block populated:

{
  "v": "evidence.action/0",
  "seq": 0,
  "session_id": "<uuid>",
  "ts": "...",
  "agent": {"vendor": "chirindo", "version": "0.0.1"},
  "event": {
    "type": "mcp_call",
    "outcome": "denied",                     // or "executed" on ALLOW
    "server": "<server-label>",
    "tool_name": "delete",
    "args_hash": "sha256:...",               // SHA-256 of JSON-stringified arguments
    "result_hash": "sha256:...",             // present only on ALLOW
    "decision": "deny",                      // or "allow"
    "decision_source": "config"
  },
  "request_commitment": "sha256:<a>",        // SHA-256 over JCS(request_descriptor)
  "gate": {
    "request_commitment": "sha256:<a>",      // MUST equal record.request_commitment
    "gate_receipt": "sha256:<self entry_hash>",  // self-anchored for the spike
    "gate_family": "permit",
    "result": "halt"                         // "act" on ALLOW
  },
  "prev_hash": "sha256:...",
  "kid": "ed25519/<fingerprint>",
  "sig": "<base64url Ed25519>"
}

The chain verifies via recorder verify (no code change to the recorder) — same hash chain, same signature scheme, same canonical bytes. Cross-tool interop is the point: the gate's output is evidence the recorder's verifier already understands.

Commands

chirindo init   [--dir <path>]
chirindo proxy  --policy <file> --server-label <name>
                [--dir <path>] [--chain <file>] [--session-id <id>]
                -- <downstream-command> [<args>...]
chirindo verify <chain-file> [--key <identity.json> | --jwks <url>]
                [--max-skew-ms <ms>]

Defaults: --dir = ./.gate/, identity at <dir>/identity.json, chain at <dir>/sessions/<session-id>.jsonl, --key = <dir>/identity.json. --jwks without a value resolves to $RECORDER_JWKS_URL, falling back to the recorder's published default.

Policy file format

{
  "deny": [
    { "tool": "delete", "reason": "destructive: blocked by policy" },
    { "tool": "shell_exec", "server": "everything" }
  ]
}

A rule matches if tool matches the call's name, AND server (if present) matches the proxy's --server-label. Anything not matched by a deny rule is allowed. The shipped policy.json is {"deny": []} — records everything, blocks nothing, observe-only by default. Enforcement is opt-in (see step 6 below). Fail-closed: an unreadable or malformed policy file still denies all calls.

Getting started

The goal of this section: take you from "I use Cursor or Claude Code with my own MCP server" to "Chirindo is observing it, and I can independently verify a receipt." Six steps, all done locally except the verify hop which contacts a public JWKS endpoint.

1. Install

<!-- INSTALL: TBD at publish — clone-and-run vs npm install. Until the install mechanism lands, assume you have a local checkout of Chirindo built (npm install && npm run build) and a working absolute path to its dist/cli.js. The steps below write that path as <ABSOLUTE-PATH-TO-CHIRINDO>. -->

For now: git clone Chirindo, npm install && npm run build. Note the absolute path to the repo — the next step uses it.

Generate the gate's signing identity:

node <ABSOLUTE-PATH-TO-CHIRINDO>/dist/cli.js init --dir <ABSOLUTE-PATH-TO-CHIRINDO>/.gate
# -> initialized chirindo at <abs path>
#    kid: ed25519/...

2. Configure your client

Copy the template that matches your MCP client into the right place:

• Cursor: config-examples/cursor-mcp.json → <your-project>/.cursor/mcp.json (or ~/.cursor/mcp.json)
• Claude Desktop: config-examples/claude_desktop_config.json → ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows)

Then edit two things — see config-examples/README.md:

1. Replace every <ABSOLUTE-PATH-TO-CHIRINDO> with the absolute path to your Chirindo checkout.
2. Replace the line after "--" with your real downstream MCP server's command (the template ships with npx -y @your-org/your-mcp-server as a deliberately-invalid placeholder so a forgotten edit fails loudly). The documented default is npx-form on every platform; a node + absolute-path fallback is documented for clients whose PATH does not include npx.

Restart your client. You should see my-server-gated in its MCP indicator. Chirindo is now wrapping your server.

3. Run

Use your agent as you normally would — anything that calls a tool on your downstream server is being observed.

4. Observe

Each MCP session writes a chain file:

ls <ABSOLUTE-PATH-TO-CHIRINDO>/.gate/sessions/

One JSONL line per tools/call. Each line is a signed receipt covering the request and its outcome. Inspect one:

head -n 1 <ABSOLUTE-PATH-TO-CHIRINDO>/.gate/sessions/<session-id>.jsonl

You'll see event.type:"mcp_call", event.decision:"allow", gate.result:"act", and an Ed25519 signature in sig.

5. Verify (the payoff)

node <ABSOLUTE-PATH-TO-CHIRINDO>/dist/cli.js verify \
  <ABSOLUTE-PATH-TO-CHIRINDO>/.gate/sessions/<session-id>.jsonl \
  --jwks

The bare --jwks form resolves the gate's public key from the recorder's published JWKS document over HTTPS, then verifies every record's signature and the hash-chain linkage. Expected output:

VALID — N entries, chain intact, all signatures verified, session <id>

You just verified, against a public key over the internet, what your gate recorded — no trust in this repo, no trust in the binary you ran, no trust in the client you used. The receipts prove the gate fired for each call and that the chain is recomputable from the signed records. They do not prove the action was "safe," only that the decision is captured, signed, and tamper-evident (not tamper-proof: any actor with the chain file can rewrite history, but a mutation breaks the hash chain and is caught by chirindo verify).

Offline alternative (no network): --key <ABSOLUTE-PATH-TO-CHIRINDO>/.gate/identity.json.

6. Enforce (opt-in)

Enforcement is one line in policy.json. Add a deny rule for a tool on your downstream server that you'd rather never have happen:

{
  "deny": [
    { "tool": "shell_exec", "reason": "blocked by policy" }
  ]
}

Restart your client. Ask the agent to call shell_exec. The downstream never receives the call; the agent sees isError: true; a DENY receipt with event.decision:"deny" and gate.result:"halt" is appended to the chain. Run chirindo verify again — still VALID.

That's the observe→enforce transition: same gate, same receipts, one extra line in policy.json.

Honesty about what isError: true does and doesn't do

MCP's spec describes tool execution errors this way:

Tool Execution Errors contain actionable feedback that language models can use to self-correct and retry with adjusted parameters.

So isError: true blocks the action (we never forward to the downstream) but does NOT block the agent — the LLM may retry. Each retry is independently evaluated by the gate and will be denied again if it matches policy. The destructive side-effect is prevented; the agent's attempt count is not capped. Productization may want a stronger "this conversation cannot perform this action" mechanism than per-call denial (e.g. a session-scoped lockout, or a protocol-error escalation after N denials).

What the harness proves and what it does NOT

Proves (from the test suite)

• The proxy correctly parses and mediates newline-delimited JSON-RPC per MCP stdio transport spec.
• On DENY, the downstream server never receives the call — proven by the in-memory test asserting a flag set on the fake downstream's data handler stays false, AND by the spawn test asserting the fake server's "DESTRUCTIVE delete tool ran" stderr line never appears.
• ALLOW + DENY + FAIL-CLOSED each emit a receipt with gate populated per the schema, including the continuity invariant gate.request_commitment == record.request_commitment.
• The recorder's verify accepts every produced chain unchanged, and catches a tampered receipt with the legible request_commitment mismatch reason.
• The OS-pipe path works (spawn integration test).
• chirindo verify (CLI e2e test) reports VALID on a fresh chain, TAMPERED on a mutated chain, and exit 2 on conflicting --key + --jwks — same vocabulary the recorder uses, because it is the recorder's verify engine wired into the chirindo binary.

Does NOT prove

• The real client honors isError: true as a block in the agent loop — for clients other than the ones already tested. Cursor's agent halts cleanly on a deny-shaped result (proven live, see SPIKE_RESULT.md); whether Claude Desktop and other MCP clients surface it the same way to the LLM must be confirmed per client.
• COSE output mode. Receipts are currently JSON/Ed25519/base64url. A COSE_Sign1 variant is the productization step for ecosystem interop.
• Argument-level policy. The current policy matches on tool name (+ optional server label) only. Production needs argument matchers (e.g. "deny shell_exec whose command starts with rm").
• Per-action UI — the proxy emits stderr logs and writes JSONL; Claude Desktop / Cursor will not surface "this action was gated" in any operator-visible way without further integration.

License

Apache-2.0. See LICENSE.