Switchboard

A firewall and password manager for your AI coding agents.

Switchboard controls what Claude Code and Codex can reach in a repo. It finds the MCP servers and tokens your agents can already touch, gets secrets out of plaintext config and behind named keychain refs, and puts each agent on a scoped pass that expires on its own.

Everything runs locally. No account, no hosted service, no telemetry.

What it does

• switchboard scan shows which MCP servers and tokens agents can reach in this repo, and which routes bypass Switchboard entirely.

• switchboard import consolidates scattered Claude/Codex MCP config into one Switchboard route, with timestamped backups and exact rollback commands.

• Secrets live in your OS keychain as named refs. Config files carry secretRef: "github/ci/token", never the token itself.

• switchboard grant puts the agent on a pass: named tools reachable, everything else denied, gone in a few hours.

• switchboard status answers whether a pass is live right now. switchboard revoke ends one early.

• Every tool call routed through Switchboard lands in a local audit log. switchboard logs reads it.

Related MCP server: agent-vault

Install

npm install -g @switchboard-mcp/cli

Or without installing:

npx -y @switchboard-mcp/cli@latest scan

Requires Node 22 or newer. Daily use is on macOS, CI runs on Linux, and a Windows keychain backend exists but gets less exercise.

Quickstart

Start in a repo where your agents already work:

$ switchboard scan

This looks like acme-app.

Detected:
- Codex Switchboard route missing
- Claude Switchboard route missing
- Claude direct MCP server "github" detected

Authority bypasses:
  high claude:github (github; direct-mcp-server, secret-env-name)

Warnings:
- 1 direct MCP bypass finding(s), including 1 high-risk finding(s), were detected.

That high-risk finding is a GitHub token sitting in plaintext in .mcp.json. Move it behind Switchboard:

switchboard import --dry-run          # see the plan first
switchboard import --write --cleanup-client
switchboard secrets set <ref>         # store the token; import prints the exact command
switchboard install claude --write    # route the agent through Switchboard

Import rewrites client config to a single Switchboard route, replaces the plaintext token with a named keychain ref, and leaves a backup plus the exact rollback command. Import never reads secret values itself; the secrets set step is where the token actually enters your keychain. Then scope the agent:

switchboard grant --for 4h

You get the pass above. When you want it gone early:

$ switchboard revoke
Revoked pass grant-main (main). The agent's scoped access is off now.

switchboard doctor tells you the next thing to fix at any point. If a repo has no MCP config yet, switchboard setup github-ci starts from a safe provider template instead.

What Switchboard does not do

• It is not a sandbox. Switchboard governs the paths routed through it: Switchboard MCP endpoints and switchboard run. An agent with raw shell access, a provider CLI, a browser session, or a direct MCP route can bypass it. switchboard scan reports those bypass routes so you can clean them up or accept them deliberately.

• Backups keep your old config exactly as it was. If a token was in plaintext before import, the backup still contains it. Rotate old tokens after migrating, and keep backups private.

• A pass only binds routed agents. switchboard grant says so itself when no client is wired up yet, and switchboard install <claude|codex> closes the gap.

• It is alpha software. Local-first workflows, conservative claims, rough edges. The pass and audit contracts are versioned JSON, but expect change.

Commands

Commands that report state take --json for scripts and harnesses, with versioned schemas. switchboard <command> --help has the rest.

How it works

Switchboard reads layered YAML config (global, then .switchboard.yaml, then .switchboard.local.yaml). Each profile names an upstream MCP server and the secret refs it needs. At runtime, Switchboard mounts permitted profiles as one MCP endpoint, resolves secret refs from the OS keychain only at launch, and namespaces every tool so a pass can allow github_ci_* and deny everything else. Passes, approvals, and audit entries are plain local files that reference secrets by name, never by value.

Secret storage uses OS keychain backends by default. Plaintext fallbacks exist for dev machines and CI, and require an explicit SWITCHBOARD_ALLOW_UNSAFE_SECRET_BACKENDS=1 opt-in.

For harnesses and subagent systems

switchboard pass create --from github-ci --json returns a workspaceLease.mcpLaunch payload: the exact command to launch a pass-scoped stdio MCP endpoint, plus the pass policy and lease. Switchboard grants and audits authority; your harness owns scheduling, retries, and agent processes. JSON contracts are documented in docs/use-cases/harness-json-contracts.md.

Alternatives

• Project MCP config alone works, but it is static wiring with tokens in files. Switchboard imports that wiring, moves the tokens, and adds scoped expiring access on top.

• Docker MCP Gateway and MCP runtimes run and package MCP servers well. Switchboard is the local authority layer above them, deciding which profile, tools, and lease an agent gets for a task. A gateway can be an upstream behind a Switchboard profile.

• Hosted tool platforms (Composio, Arcade) offer managed OAuth and broad SaaS coverage. Switchboard is local-first for coding-agent repos: local secrets, local audit, no hosted dependency.

• Just giving the agent a token is fast until the token is broad, live, or copied into the wrong file. That is the failure mode Switchboard exists for.

Development

git clone https://github.com/wkoverfield/switchboard.git
cd switchboard
pnpm install
pnpm build
pnpm switchboard --help
pnpm test

Contributions are welcome; see CONTRIBUTING.md, and report vulnerabilities privately per SECURITY.md.

Smoke tests and fresh-agent evals live in scripts/; CI runs the full set. The top demo is a VHS tape: brew install vhs && pnpm build && vhs examples/switchboard.tape. Deeper docs: quickstart, threat model, provider safety templates, roadmap.

Agents can read these docs over MCP: npx -y @switchboard-mcp/docs-mcp serves list_docs, read_doc, and search_docs, and llms.txt is kept current. The landing and docs site lives in site/ (static build, pnpm --filter @switchboard-mcp/site build).

License

MIT. See LICENSE.