Preparing components...
Skip to content

Docs

Bring your own agent.

Connect Switchboard to Claude, Codex or Cursor. Your agent gets only the tools each task calls.

Connect an agent

Quickstart

Package publication is pending. This website uses its browser-local planner. A separate HydraDB-backed API has passed hosted planning and guard smoke tests; use only synthetic or public data there. The current benchmark measures seven fixture tasks, not a general latency guarantee.

terminal
1# After switchhboard-mcp is published on npm
2npx -y switchhboard-mcp stats
3
4# Get a local plan
5npx -y switchhboard-mcp plan --task "Open a PR on acme/api fixing the flaky login test"
6
7# Optional HTTP API on port 8787
8npx -y switchhboard-mcp serve --port 8787

Connect your agent

One stdio MCP server, once the switchhboard-mcp package is published. It exposes plan, report_run and catalog_stats. The product supports a real HydraDB server. Configure HYDRA_HTTP_URL and HYDRA_TOKEN for your own instance. The separate public API is an unauthenticated synthetic-data demo, not private account storage. Schemas are included where available; missing schemas must be fetched from the actual MCP server.

terminal
1# Available after npm publication
2claude mcp add switchboard -- npx -y switchhboard-mcp
3claude mcp list

Run it in your project. Check it with claude mcp list.

Call the API directly

Send a task to get tool identities, typed inputs and call order. The current product measured 700 mean plan tokens against 110,238 preload-all tokens over seven fixture tasks. This is not a general performance guarantee.

HTTP API
1POST http://127.0.0.1:8787/api/plan
2{ "task": "Find open login bugs in Linear and post a summary to #eng",
3 "held": [{ "type": "slack.channel_id", "value": "C0123ENG" }] }

The plan contract

POST /api/plan accepts task, held, exclude and allow_destructive. Steps come in call order. Full input schemas are available for 51 of 420 starter tools; fetch missing schemas from the actual MCP server. The official ingested catalog has schemas for all 37 tools.

response · /api/plan
{
  "note": "Illustrative response shape. Actual fields and values depend on the task.",
  "plan_id": "pl_example",
  "steps": [
    {
      "order": 1,
      "tool_id": "server.tool",
      "consumes": [],
      "yields": [],
      "schema": "included only when present in the catalog"
    }
  ],
  "guard": {
    "status": "ok",
    "note": "Plan-time heuristics only, not runtime protection"
  }
}

Three more endpoints:

/api/plan_baseline
BM25 top-k over the same catalog, for comparison.
/api/report_run
Log what ran. Each report strengthens the NEXT edges between the tools you called.
/api/guard/check
Check the plan-time risk pattern. Pass held values to plan or report_run to retain known values.

Add your own servers

List your MCP servers, run the harvest, re-ingest. Parameters map onto a shared type vocabulary, so a GitHub repo and a GitLab repo stay different types. That's what makes the edges typed.

servers.json
1# Import a live server into a local catalog
2npx -y switchhboard-mcp ingest \
3 --server "npx -y @modelcontextprotocol/server-filesystem /absolute/allowed/path" \
4 --name filesystem --catalog ./catalog.json
5
6# Plan using that catalog
7npx -y switchhboard-mcp --catalog ./catalog.json plan --task "Read a file"

Risk flags and the guard

Every tool gets flags at ingest. If a plan holds all three trifecta legs, the edge into the sink is cut before your agent sees it. Watch it happen.

  • reads_private
    reads private data
  • ingests_untrusted
    ingests untrusted content
  • exfil_sink
    can exfiltrate
  • destructive
    destructive
Flags come from heuristics, so they can miss a risky tool. The guard checks plans. It does not sandbox the calls your agent makes.

Where it runs

This website plans in your browser using the bundled catalog and recorded examples. Reports stay in browser-local storage. It is not connected to the separate hosted HydraDB API. That API has no user authentication and is for synthetic or public data only; free hosting can sleep or reset its data.

Credits and licences

  • Planning engine: local embedded typed graph. Original site design and reference catalog: owner-provided Switchboard site.
  • Photos and visual assets: preserved from the owner-provided original site; original third-party asset notices remain applicable.
  • Agent brand marks: Claude, OpenAI/Codex and Cursor, shown only to identify compatible clients. Reference assets: docs.trybounty.ai.
  • Type: Geist and Geist Mono (SIL Open Font License).
  • Icons: Lucide (ISC). Motion: GSAP.

Switchboard questions

What is Switchboard?

Switchboard plans MCP tool calls using typed dependencies, returning the tools a task needs in call order.

Does this website call the hosted database?

No. The website is a browser-local interactive demo with recorded examples. The separate hosted API uses real HydraDB, but this website does not send tasks or run reports to it.

Is My memory private cloud storage?

No. Accounts are not connected. Website reports stay in this browser. The separate API is unauthenticated and must not receive private tasks, credentials or personal memory.

What has been verified?

The real-server integration and learning checks passed. Hosted planning and guard smoke tests passed. These checks do not guarantee security or performance for every task. The API container build remains unresolved; the hosted API uses a native Node deployment.

Can I install the npm package now?

The intended package is switchhboard-mcp, with a double h. Publication is still pending. Do not install similarly named packages.

Search Switchboard

Search the docs, plan a task, or jump to a page