Saphan Studio explained

For a reader deciding whether the approach is sound, before or instead of installing anything. This is the only section written to be read straight through.

After reading it you can use the product's vocabulary correctly and explain to someone else why a fleet needs a record rather than a log.

Before you start: nothing. If you would rather act first, install from Getting started and come back.

In one paragraph

Saphan Studio is a control plane for fleets of AI coding agents — Claude Code, Codex, local models behind an OpenAI-compatible endpoint, and plain deterministic tools — built on one idea carried to its end: every consequential step is a signed, recorded act under human gates. Agents plan, implement, test, review and report at machine speed; nothing irreversible happens on model output alone. A human accepts plans, gates merges, ratifies the fleet's laws. Everything else — dispatch, isolation, metering, verification, bookkeeping — the engine does, and writes down. The result is an agent fleet that a serious engineering organization can actually run: attributable, bounded, priced to the cent, and auditable after the fact by someone who was not in the room.

Why it exists

An AI coding agent is a new kind of actor inside your perimeter: capable, cheap, fast — and forgetful, persuadable, and unaccountable by default. One agent on a laptop is a productivity story. A fleet of them doing real work for weeks raises the questions every engineering organization already knows how to ask: who decided this change ships? what did it cost? what was refused, and why? which of three parallel changes actually merged — and can two agents quietly share one login? Sessions forget; organizations cannot afford to. Saphan Studio's answer is deliberately boring: a record of acts, and a fleet that cannot step around it.

The foundation: Saphan Protocol

Saphan Studio works on Saphan Protocol — public at saphan.ai: "the written discipline the engine enforces — orders, gates, evidence standards, decision records — versioned, human-readable, and hardened daily against our own fleets."

Two properties of the protocol matter to everything below. First, the method is not locked inside the tool: the protocol publishes openly, so machines execute it while auditors and engineers can simply read it — the discipline outlasts any single vendor implementation, including this one. Second, the engine speaks the protocol in both directions: the orders agents receive, the evidence standards their returns must meet, the gate procedures humans decide by, and the decision records (ADRs, findings) the work leaves behind are one language across every surface — editor, CLI, dashboard, or a chat with the fleet. So where these pages describe the working method, they are describing a published discipline being enforced rather than one product's house habit.

The sentence to remember

The fleet does the work, the record does the remembering, and you do the meaning.

What this section does not cover

⚠ Nothing here is a procedure. No command, no install path, no configuration — those are Getting started and Deployment. ⚠ It does not state which capabilities are shipped; where a page names a mechanism, check its own section for the state that mechanism is in.