Switchboard is an agent gateway: you say what you want over Slack, the CLI, HTTP or MCP, and the agent it picks answers the question, reviews the pull request, or ships the code. Read this tree here or at switchboard.space.
New here? Use an existing deployment if your team already runs Switchboard, or Get started to set up your own.
| You are… | Start with | Then |
|---|---|---|
| joining an existing deployment | Use an existing deployment | Your first request in Slack |
| trying more requests in Slack | Your first request in Slack | Configure your defaults |
| running it in production | Deploy | Operate production |
| watching runs or spend | Watch a run | Check spend |
| locking it down | Restrict who can do what | Security model |
| changing the code | Run it locally | How a request flows |
A plain sentence is the whole interface. Commands and directives are there for whoever wants to type them: Slack commands.
The tree follows Diataxis: four jobs, never mixed on one page.
| Kind | Answers | Index |
|---|---|---|
| Tutorials | "Walk me through it" | tutorials/ |
| How-to guides | "How do I do X?" | how-to/ |
| Reference | "What is the exact syntax, value, default?" | reference/ |
| Explanation | "Why does it work this way?" | explanation/ |
reference/vocabulary.md is the product's words: the twelve nouns every surface prints, one meaning each — read it before the code map.
For the system picture, start with Architecture, Data model and What an agent is.
For long-running work, read How agents keep context and the Context lifecycle. Their diagrams cover research, verified phase notes, focused child context, recall and source access.
reference/specs/ is the contract: one file per feature, every criterion bound to a test or a procedure. When a page here disagrees with a spec, the spec is right.
README.md is the front door; AGENTS.md is for whoever changes Switchboard.
The shared request failure context design defines direct failure lookup across the gateway and saved runs.