Overview
Some of the most popular agent surfaces are MCP clients with no pre-tool hook. Claude Desktop and Kimi run tools exclusively through Model Context Protocol servers and expose no place to run a policy check before a tool fires. Codex fires hooks for Bash but not for MCP calls. For this entire class of client, the enforcement point is the MCP transport itself.@sigilcore/mcp-proxy sits between the client and its MCP servers. Every
tools/call request is evaluated against your operator’s warranty.md policy via
Sigil Sign /v1/authorize before it reaches the real server, and is wrapped in a
Sigil Intent Attestation. You change one line in your MCP client config to point a
server at the proxy. No client code changes are required.
What this governs. The proxy governs MCP tool calls only. It does not see
a client’s built-in capabilities (Claude Desktop connectors and web search, Codex’s
native Bash and file edits). Pair the proxy with a native hook where one exists:
Codex hooks for Bash, Claude Code
and Hermes for full tool coverage. The proxy is the right
tool precisely where no native hook exists.
Quick Start
MCP Client Config
Wrap an existing server by changing one line in your MCP client config. The proxy launches and supervises the upstream server and authorizes every call to it. This form works in any client that reads a standardmcpServers block, including
Claude Desktop, Kimi (kimi-cli / Kimi Code), and Codex MCP config.
Configuration
Generate a starter config withnpx @sigilcore/mcp-proxy --init. Precedence is
CLI flags > environment variables > config file > defaults.
| CLI Flag | Env Var | Config Key | Default |
|---|---|---|---|
--key | SIGIL_API_KEY | (env only) | (required) |
--sign-url | SIGIL_SIGN_URL | signUrl | https://sign.sigilcore.com |
--log-level | SIGIL_LOG_LEVEL | logLevel | info |
--server-id | — | serverId | derived from command/URL |
--server-name | — | serverName | same as serverId |
--pending-timeout | SIGIL_PENDING_TIMEOUT | pendingTimeout | 30000 |
--unsafe-bypass | (CLI only) | (CLI only) | false |
--remote | — | — | — |
--port | — | — | auto |
Server Identity
serverIdis the binding identity and is security-critical. It is used in thetxCommitpreimage and in policy evaluation. It is auto-derived from the package name (stdio) or the full URL (HTTP) when not set explicitly.serverNameis a display label for logs only and defaults toserverId.
Fail-Closed by Default
The proxy is fail-closed: when Sigil Sign is unreachable, tool calls are blocked. To allow ungoverned calls during a Sign outage, pass--unsafe-bypass. This option
is intentionally CLI-only (no env var, no config key) so it is always visible in
your MCP client config.
ungoverned_tool_call error-level log. Authentication
failures (401) are never bypassed.
HTTP/SSE Transport
Proxy a remote MCP server with--remote:
$IDENTIFIER env var; raw secrets are rejected at load time.
Extractors
Map tool arguments to Sigil policy fields insigil.config.json so the right
warranty.md rules apply. For example, route a fetch tool’s url argument to the
web_fetch policy field and a write tool’s path argument to file_write:
Error Codes
-32001— Sigil policy denial (DENIED, fail-closed block, or hold timeout)-32002— Sigil authentication failure (invalid API key)
Conformance
The proxy is a client of the SOF enforcement specification, not a signer. It submits intents and acts on decisions; the authorization itself is issued by Sigil Sign or any conforming signer. The intent wire format is the same/v1/authorize contract used by every agent-hooks adapter, documented in
Getting Started and sigil-attestations.
Source
- npm: @sigilcore/mcp-proxy — MIT License