tmcproxy
Bridges OpenAI's Secure MCP Tunnel to legacy stdio MCP servers by answering server/discover requests that older servers don't support.
README
tmcproxy
Minimal MCP stdio compatibility proxy that bridges OpenAI Secure MCP Tunnel / ChatGPT connectors to legacy stdio MCP servers.
The problem this solves
ChatGPT connectors (via OpenAI Secure MCP Tunnel) probe stdio MCP servers with server/discover — a method introduced in MCP 2026-07-28. Legacy servers like codex mcp-server and local-mcp mcp only speak 2025-06-18 and return method not found for server/discover, causing connector creation to fail.
tmcproxy sits in between, answers server/discover on the downstream's behalf, and forwards every other request verbatim.
ChatGPT → tunnel-client → tmcproxy → codex mcp-server (or local-mcp mcp)
↑
answers server/discover here,
forwards everything else straight through
Requirements
- Node.js >= 23.6.0 (for native TypeScript strip-mode execution — no build step needed)
Usage
Standalone
node src/tmcproxy.ts -- codex mcp-server
node src/tmcproxy.ts -- local-mcp mcp
Or via environment variable:
TMCPROXY_DOWNSTREAM="codex mcp-server" node src/tmcproxy.ts
With tunnel-client
Edit your tunnel-client profile YAML (e.g. ~/.config/tunnel-client/<profile>.yaml):
mcp:
commands:
- channel: main
command: "node /path/to/tmcproxy/src/tmcproxy.ts -- codex mcp-server"
Then:
tunnel-client run --profile <profile>
Switching downstreams is just changing the command after --:
command: "node /path/to/tmcproxy/src/tmcproxy.ts -- local-mcp mcp"
How it works
| Request | Behavior |
|---|---|
server/discover |
Proxy answers with a spec-shaped DiscoverResult. Capabilities are probed from the downstream via tools/list (not guessed). |
initialize |
Forwarded verbatim to downstream. The proxy never calls initialize itself — codex mcp-server allows it exactly once, so the client's handshake must be the one that reaches it. |
tools/list, tools/call, ping, etc. |
Forwarded verbatim with JSON-RPC id preserved. |
| Notifications (no id) | Forwarded; no response expected. |
| Malformed JSON | -32700 Parse error returned to upstream, not forwarded. |
| Unknown method | Forwarded; downstream error flows back. |
Why supportedVersions includes 2026-07-28
ChatGPT probes with 2026-07-28. If the response only lists 2025-06-18, ChatGPT treats it as "requested version not supported" and retries indefinitely — never reaching initialize or tools/list. By listing 2026-07-28, ChatGPT enters modern mode and sends stateless requests (tools/list, tools/call) directly. The legacy downstream answers these without a prior initialize (confirmed for both codex mcp-server and local-mcp mcp), so transparent forwarding works.
Debugging
Debug log
Set TMCPROXY_DEBUG=1 to log server/discover requests and responses to stderr:
TMCPROXY_DEBUG=1 tunnel-client run --profile local-codex
Output on stderr (stdout stays clean — it's the MCP channel):
[tmcproxy] downstream: ["codex","mcp-server"]
[tmcproxy] discover request: {"jsonrpc":"2.0","id":"openai-mcp-discover",...}
[tmcproxy] discover response: {"resultType":"complete","supportedVersions":["2026-07-28","2025-06-18"],...}
Local end-to-end test (no API key needed)
tunnel-client dev proxy runs a local in-memory control plane that reproduces the same MCP probe flow ChatGPT uses:
tunnel-client dev proxy \
--mcp-command "command=node /path/to/tmcproxy/src/tmcproxy.ts -- codex mcp-server,channel=main" \
--url-file /tmp/url.json \
--duration 30s
MCP_URL=$(node -e "console.log(JSON.parse(require('fs').readFileSync('/tmp/url.json','utf8')).mcp_url)")
curl -sS -X POST "$MCP_URL" \
-H 'Content-Type: application/json' \
-H 'Accept: application/json, text/event-stream' \
-H 'MCP-Protocol-Version: 2026-07-28' \
-H 'Mcp-Method: server/discover' \
-d '{"jsonrpc":"2.0","id":1,"method":"server/discover","params":{"_meta":{"io.modelcontextprotocol/protocolVersion":"2026-07-28"}}}'
Baseline comparison
To confirm the proxy is needed, run the downstream directly through dev proxy and observe the server/discover failure:
tunnel-client dev proxy \
--mcp-command "command=codex mcp-server,channel=main" \
--url-file /tmp/url.json --duration 30s
# server/discover → -32601 method not found
Tests
pnpm install
pnpm test # unit + integration
pnpm test:unit # mock downstream only
pnpm test:integration # real codex / local-mcp (auto-skipped if not installed)
pnpm typecheck
Unit tests use a mock downstream (test/fixtures/mock-downstream.js) and cover: discover response shape, capability probing, initialize forwarding, tools/list, tools/call id preservation, ping, unknown methods, malformed JSON, notifications, stderr mirroring, downstream exit handling, and stdout leak detection.
Integration tests spawn the real codex mcp-server and local-mcp mcp binaries if available on PATH.
Security
- Diagnostics go to stderr only; stdout is reserved for MCP traffic.
- No credentials, API keys, or secrets are logged.
- Downstream is spawned via argv (no shell).
- Only
server/discoveris handled by the proxy; all other requests pass through unchanged. - The proxy does not expand the downstream's permissions.
License
Apache-2.0
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.