Designed for the agents you already use

Auteur is a full studio in its own right, and it does not need an outside agent to run. Drive it on its own, from its dashboard, from the Control Deck on a tablet, or from Telegram on your phone, in plain language, with hands-on control over every part of a build, a campaign, a film, or a novel. Or seat the AI harness you already use, Claude Code, Codex, Cursor, or Grok, on top of it for a familiar, fire-and-forget way to make progress. The harness is optional, never required.

That is the real choice this page is about. Bring a harness and Auteur lends it the power of essentially every major AI platform at once, orchestrated, governed, and receipted. Bring none and the studio still runs, entirely on local models if you want, with no frontier harness in the loop at all. Either way Auteur is an amplification and accountability layer, never a replacement: the agent can change, or there may be no agent at all, and the project, process, authority, and evidence remain.

Anyone can hand an agent an MCP connection; that part is commonplace now. What matters is what travels across it. Auteur bakes premade, engineered toolsets straight into the harness you already use, so the agent is not left rooting around your application, guessing at what to do and burning tokens and time to rediscover it every session. It arrives with a roadmap: the shortcuts to the result you asked for in plain language, with the technical steps already solved behind software-engineered guardrails that keep the harness from overreaching, misfiring, or taking a destructive action. Think of it as three things your agent does not have on its own: an efficiency layer that removes the trial and error, a memory layer that already knows your project, and pre-built, battle-tested toolsets that make reaching the right result simpler, faster, and safer.

Every harness, one project

Chat harnesses and coding harnesses seat here side by side. Each one keeps its own page, its own measured status, and its own session mockup:

• Claude Code, the conductor's chair: long-context planning, cross-family review, and orchestration.
• Codex, the workhorse: the most-used builder in our own ledger, housed on your subscription.
• Cursor, the editor that keeps its editor: campaign structure and cross-model review around the tool you code in.
• Grok, the improviser with a studio: receipts, reach, and order over multi-agent chaos.
• Google Antigravity, Gemini's agentic environment, with vision that parses footage in intake and review.
• OpenClaw, Hermes, and any spec-compliant MCP client at their published status.

How they interact

The point is not eight separate integrations. It is one project where they work each other's blind spots:

• Orchestrator and builders. One harness conducts a campaign while others build inside it: separate worktrees, allowlists, and spend ceilings per seat.
• No seat grades its own homework. Diffs and drafts from one family are reviewed by another. The 16 seat failures caught on this site's own build were caught exactly this way.
• Shared memory, shared receipts. Whatever seat did the work, the project remembers it: canon, decisions, costs, and verification live in one ledger every harness can read.
• Handoffs without loss. Start a scene in Claude Code at the desk, approve renders from Grok on your phone, let Codex run the overnight build: the project, not the chat scroll, is the continuity.

Three ways a harness connects

Every harness below is a designed seat for Auteur over the Model Context Protocol; each mode's current, measured state lives in its chips.

• Client, the harness is your interface and calls Auteur as an MCP control surface. Auteur holds the project, the authority, and the evidence.
• Executor, Auteur drives the harness as a campaign seat under an executor profile. Requires headless operation and a machine-readable output stream.
• Direct model, Auteur calls the vendor's API directly for measured, accounted work, with no harness in the loop.

The harnesses

Each harness row lists its designed connection modes - client, executor, or both - with per-mode state carried by the chips.

• Grok
• Claude Code
• Codex
• Cursor
• Google Antigravity
• OpenClaw
• Hermes
• Generic MCP client

What every harness page tells you

The same eleven things, in the same order, at the same depth, for every harness: what it does well, what Auteur adds, how it connects, which products it reaches, what is measurable, what the integration does not do, the security and authority boundary, the current status, the proof, how to set it up, and the affiliation statement.

What is measurable

Auteur measures the work that flows through it. When a harness runs as an Auteur executor or a direct-model lane, token and cost accounting is derived from the record. When work happens outside Auteur's visibility, Auteur never invents missing usage data; missing = unknown, never zero.

Affiliation

Auteur is designed to connect with the tools named on these pages. No partnership, endorsement, or certification is claimed, and none exists.

Proof

Status source: site_data/integration_page_schema.yaml · Claims: COND-004, COND-003, COND-001, BUILD-002, BUILD-004, PLAT-007, PLAT-005 · Evidence: harness registry, MCP release packet, and campaign history · Verified: 2026-08-28.