Add mcp-bridge: use any MCP server's tools by voice - #391

What does this Ability do?

MCP Bridge connects the speaker to one MCP server the user chooses and lets them use every tool on it by voice. The user adds the server's address, and a token if it needs one, under Settings, API Keys. MCP Bridge then lists the server's tools, picks the right one for what the user says, fills in the details, and speaks back a short answer. Nothing has to be set up per tool. One Zapier MCP server alone reaches thousands of apps.

No existing ability or template uses MCP.

• n8n Commander needs a webhook wired for each workflow.
• The API template is one hardcoded endpoint per ability.
• OpenClaw, Hermes, the Local Link templates and Voice Dev Terminal need software running on the user's computer.

MCP Bridge finds a hosted server's tools by itself, through the open standard. The README covers the differences.

Suggested Trigger Words

• MCP bridge
• use my MCP
• ask my server
• use my tools
• Zapier

Type

• New community Ability
• Improvement to existing Ability
• Bug fix
• Documentation update

External APIs

• No external APIs
• Uses external API(s): the MCP server the user configures, over HTTPS. API keys: mcp_server_url (required) and mcp_server_token (optional, sent as Authorization: Bearer), both read with get_api_keys(). These public servers need no key and work for testing: https://docs.mcp.cloudflare.com/mcp, https://mcp.deepwiki.com/mcp.

Testing

• Tested in OpenHome Live Editor
• All exit paths tested (said "stop", "exit", etc.)
• Error scenarios tested (API down, bad input, etc.)

Live Editor spike. A throwaway spike ran in the Live Editor and confirmed the sandbox supports everything this needs:

• socket.getaddrinfo
• requests streaming with allow_redirects=False
• a 10-second HTTP wait inside a session
• untagged optional API keys

The full ability has since been tested in the Live Editor and works.

Offline tests (103 scripted cases). These run against a fake MCP server in three modes: modern JSON, modern SSE, and the legacy session handshake. They cover:

• 401s, redirects, oversized replies, session expiry, unsupported versions and invalid x-mcp-header tools
• private and metadata addresses being refused
• exit at every prompt, plus the confirm, change and decline paths
• read-only versus write gating
• a broken timezone, a corrupt Social Memory file and a failed storage read

Live checks against real servers. I also ran the transport against DeepWiki, Cloudflare Docs and Hugging Face. All three connect and their tool calls work (4 to 9 seconds). Zapier without a token gives the "needs sign-in" message.

Checklist

• Files are in community/mcp-bridge/
• main.py follows SDK pattern (extends MatchingCapability, has register_capability + call)
• README.md included with description, suggested triggers, and setup
• resume_normal_flow() called on every exit path
• No print() — using editor_logging_handler
• No hardcoded API keys — using placeholders
• No blocked imports (redis, user_config)
• No asyncio.sleep() or asyncio.create_task() — using session_tasks
• Error handling on all external calls
• Tested in OpenHome Live Editor

Anything else?

Protocol

• It speaks the MCP 2026-07-28 stateless Streamable HTTP revision: one POST per call with MCP-Protocol-Version, Mcp-Method and Mcp-Name, base64 header values for non-ASCII names, and Mcp-Param-* mirroring for x-mcp-header.
• It falls back automatically to the 2025-11-25 initialize / Mcp-Session-Id handshake, following the spec's era-detection rules, and caches which one the server uses.
• It handles JSON and SSE replies, and follows tools/list pagination.
• Limits, stated in the README:
  • OAuth-only servers aren't supported.
  • Servers must be public and use HTTPS.
  • Only one server at a time.

Safety

• Addresses: https only, no redirects followed. Private, loopback, link-local and metadata addresses are refused, both as typed and after the name is resolved.
• Size limits: replies are capped at 512 KB, and only the first 6,000 characters are summarized.
• Untrusted text: tool descriptions and tool output go to the LLM as fenced, untrusted data. There's exactly one tool call per user request, and output never triggers another call.
• Confirmation:
  • A tool runs straight away only when the server marks it read-only, nothing about it looks like a write, and none of its values came from the user's stored details.
  • Everything else asks "Go ahead, or change something?". The user can confirm, decline, or change the details.
• Secrets: tokens and query strings are never logged, and no token or tool output is stored.

Following earlier review feedback

• Triggers: no does_match(); triggers live in the dashboard.
• Routing: an LLM router with natural examples, and an LLM confirmation classifier with a whole-utterance fast path.
• Exits: checked at every prompt, using whole-phrase matches.
• State: saved in Context Storage with create_key, then update_key, checking success. Reads unwrap ["value"], and a failed read blocks saves.
• HTTP: runs through asyncio.to_thread. No gather, no PEP 604 annotations.
• Prompts: missing details are collected in at most two prompts, as asked in Add flight-booking community ability #209.
• Speech:
  • Plain spoken English under 30 words, with natural dates and 12-hour times. Emails and ids are never read out.
  • No tool counts in the greeting.
  • Context-aware fillers that don't repeat, and plain error lines without setup jargon.
  • The session stays open after each answer.
• Context reuse (read-only):
  • Data from other abilities: user_profile.md, user_summary.md, memory_mirror_corrections.md (Memory Mirror) and the names in social_memory.json (Social Memory; its private snippets are skipped).
  • Platform calls: get_region_data(), and get_timezone() with a UTC fallback.
  • These fill details like "my city" or "tomorrow" instead of asking. Any value taken from them always goes through the spoken confirmation, because it's about to leave for a third-party server.
• Marketplace: I couldn't check the Marketplace for an overlapping ability from outside a logged-in account. I'll confirm before merge.