Build your MCP server. Your agent keeps one URL.

$ sb dev my-server --port 3000
my-server · relay target local → http://127.0.0.1:3000/mcp
relay attached; switching my-server to local…
✓ my-server → local (relay) · 9 tools
  agents using my-server__* now reach your machine. Ctrl-C to stop and switch back to staging.
14:02:07 tools/call get_forecast 41 ms 200 1.8 kB
✓ tools changed: +get_tides · catalog refreshed (10 tools)
14:02:31 tools/call get_tides 37 ms 200 912 B
^C
✓ restored my-server → staging · 9 tools

$ sb use my-server staging
Refreshing staging…
✓ my-server now answers from staging · 9 tools (+1, −0)
  + get_alerts
  Agents see the new tools from their next call (direct-mode clients may keep the old list up to ~60 s or until they reconnect).

$ sb use my-server prod
Refreshing prod…
✓ my-server now answers from prod · 8 tools (+0, −1)
  - get_alerts
  Agents see the new tools from their next call (direct-mode clients may keep the old list up to ~60 s or until they reconnect).

• One server name, every build

  Keep local, staging and prod under one server. Your agent always calls my-server__get_forecast, whichever build answers.

• Switch in one command

  sb use my-server staging reads the build's tools, then swaps. Your agent gets them from its next call. If the build does not answer, nothing changes.

• Your localhost, no tunnel

  sb dev my-server --port 3000 puts the server on your laptop behind your hosted MCP URL, with no public address. Tools you add show up within seconds. Ctrl-C switches back.

$ sb compare call_7Hk2pQx9LmT4vR8sWd3e local staging
my-server/get_forecast  local vs staging
        LOCAL   STAGING
result  ok      ok
time    41 ms   58 ms
size    1.8 KB  1.9 KB
tokens  912     951

Diff (json)
~ $.alerts: 0 → 1
+ $.advisory: "Small craft advisory"

✗ Results differ (2 changes).

$ sb calls --tail
14:02:07  my-server@local  get_forecast  ok  41 ms  1840→912  claude-code  call_7Hk2pQx9LmT4vR8sWd3e
14:02:19  my-server@local  get_tides  error  9 ms  156  claude-code  call_Q2nV8cJr5ZyB1tKa6Ufh
14:02:31  my-server@local  get_tides  ok  37 ms  612→402  claude-code  call_mX4Lw0dG7sPe3NqY9Tbc

{
  "issues": [
    {
      "id": "ISS-142",
      "title": "Retry webhook on 502",
      "state": "open",
      "priority": 1,
      "assignee": "ana",
      "labels": "",
      "estimate": null,
      "url": "https://tracker.example/ISS-142"
    },
    {
      "id": "ISS-143",
      "title": "Paginate audit export",
      "state": "in_progress",
      "priority": 2,
      "assignee": "chen",
      "labels": "",
      "estimate": null,
      "url": "https://tracker.example/ISS-143"
    }
  ],
  "pageInfo": {
    "hasNextPage": true,
    "endCursor": "Y3Vyc29yOjUw"
  }
}

issues[2]{id,title,state,priority,assignee,url}:
  ISS-142,Retry webhook on 502,open,1,ana,"https://tracker.example/ISS-142"
  ISS-143,Paginate audit export,in_progress,2,chen,"https://tracker.example/ISS-143"
pageInfo:
  hasNextPage: true
  endCursor: Y3Vyc29yOjUw

• The smaller format wins

  Lists of similar records go out as TOON: one header, then one line per record. Everything else goes out as minified JSON. TOON is used whenever it is smaller by 5% or more; at scale, every percent counts.

• Leaner tool lists

  Tool descriptions and schemas lose the noise: $schema, titles, examples and run-on text. Property names and validation rules stay exactly as the server wrote them.

• Big results capped, out loud

  An oversized result is cut from the end of its longest list, with a note saying how many items were left out and how to page for the rest. Errors, images and prose pass through untouched.

On by default. Turn it off for the whole workspace in Settings, or for one server that needs the raw output. The dashboard shows tokens saved per server, tool and client.

find_tools  "why did last night's deploy fail?"

  tool   github__list_workflow_runs
  tool   sentry__search_issues
  tool   linear__search_issues
  skill  deploy-triage

• Add by URL, sign in once

  Paste a remote MCP server URL. If it uses OAuth, a popup signs you in and Switchboard keeps the tokens, encrypted, and refreshes them. Header and bearer auth work too.

• Tools picked from the conversation

  Semantic search over every tool and skill description. Your agent sees the few that fit, not hundreds of schemas.

• Tool lists kept fresh

  Switchboard re-reads each server's tools in the background and backs off when a server is down. Switch a server or skill off without removing it.

• Any server, latest spec

  Old and new MCP servers alike reach your agent as current MCP. Switchboard negotiates each one's protocol version and falls back when needed.

• Structured output, checked

  Output schemas pass through and results are validated. JSON-only text arrives as structured data too. A broken result falls back to text, not an error.

• Titles and safety hints kept

  Tool titles, icons and annotations like read-only and destructive reach your agent exactly as each server declares them.

• Skills, uploaded or imported

  Upload a SKILL.md bundle, or connect a server that publishes skills and Switchboard imports them. Agents load them through the same URL.

• Metrics and cost

  Calls, success rate, p50/p95 latency and tokens saved, by server, tool and client. Estimated cost saved for Claude clients, labelled as an estimate.

• Call inspector

  Opt-in capture of the request, what the server sent and what the agent got. Secrets redacted, kept 7 days.

• Evals for your server

  Cases in a YAML suite, run on any build with sb eval run or from the dashboard. Regressions are compared with the last run, and the exit code can gate CI.

• Skill lint

  sb skills lint checks skills, plugins and marketplaces against the Agent Skills spec, offline, with GitHub annotations for CI.

• Passkeys, GitHub and OAuth

  Sign in with a passkey or GitHub. Agents connect with OAuth or a personal access token. Upstream tokens are encrypted at rest.

• Runs on Cloudflare

  The gateway, dashboard and dev relay run on Cloudflare Workers, D1, R2 and Durable Objects. Nothing to host.