Saphan StudioDocs
Concepts

How you talk with the fleet

Five surfaces over one record - CLI, web console, client API, MCP, editor.

A control plane is only as good as its doors. Saphan Studio ships five, all projections of the same record — pick per moment, not per religion:

SurfaceWhat it is forStatus
saphan CLIThe operator's surface: every verb, every act, paste-ready commands; the only writer of actsShipped
Web consoleThe fleet in a browser: overview, streams, runners, the cost ledger, the owner's inbox — see the consoleShipped; listens on loopback by default and serves nothing without a signed-in session
Client API and gatewayThe versioned HTTP client contract, served by saphan-gateway — see the gatewayShipped
MCP serverConversation with the fleet — see belowShipped
VS Code extensionThe fleet inside the editor: the fleet tree, stream detail, and gate acts whose command lines are composed by the engine — see the VS Code extensionShipped (early)

The runner-side binary, saphan-agent, is not a surface you talk to: it is what dispatched runs execute through on admitted machines.

The MCP server deserves its own paragraph, because it changes who can operate a fleet. MCP (Model Context Protocol) is the standard by which AI assistants call external tools — and Saphan Studio ships an MCP server that exposes the fleet's projections to any MCP-capable client, behind an OAuth 2.1 front door. In practice this means: you open a chat and talk to your fleet. "What waits on my gate?" "Show me this stream's history." "What did today cost, and where?" — answered from the record, with the same rows the CLI reads, in plain language. The surface is deliberately read-only: conversation informs, acts stay acts — the assistant you talk to can hand you the exact command the record composes for your hand, but only your hand executes it. This is not a chatbot bolted onto a product; it is the record made conversational. (The MCP server shows what it serves, how a client proves who it is, and worked examples of real calls.)

The doors compose. A typical day uses all five: the fleet reports into the record; you ask about it in chat; you decide; you paste one command in a terminal — or, for a gate decision, record the same act in the console; the editor shows the runs land.