>
Terms.txt
Privacy.txt
Connect your repo. Correct the picture. Every agent builds from it and gets checked against it. Three steps, and none of them change how your team works today.
Point Artifact at your GitHub repo. It reads what exists and drafts the picture of your product. It never writes to your code and never keeps a copy of it.
"No, that flow branches here." You point at what's wrong, in your words. Artifact captures the decision. Where it doesn't know, it asks, and the question stays visible until you answer it.
Claude Code, Cursor, Codex, or your developer gets a brief before every task. When code ships, it's checked against what you meant. Mismatches are shown. You decide.
Five steps between "connect your repo" and "ship what you meant".
Point Artifact at your GitHub repo. Read-only: Artifact never writes to your code, and never keeps a copy of it. Your agent reads in your own session; the server judges what is reported to it and discards the rest. No survey, no template, no forty-question interview.
Have a codebase? Artifact starts from reality. Starting from nothing? It starts from your first descriptions.
It reads what exists and assembles a living picture: every feature it can find, the rules they appear to follow, the screens that are really there. Not a pile of documents. A connected picture, where a feature knows its rules, its screens, and the exact code that implements it. Every claim carries evidence and a confidence. Gaps are reported as gaps, never invented.
This is the job only you can do. "No, that flow should branch here." "That's supposed to show both prices." You point at what's wrong; Artifact captures the decision, in your words. No forms. No PRD homework.
Where the picture is incomplete, Artifact asks. An unanswered question stays a visible open question, not a silent gap. Artifact shows you the problem; you decide.
Artifact speaks MCP, the open protocol your AI tools already use. They connect to your Artifact and pull exactly the context each task needs: the feature, its rules, its acceptance criteria, the decisions behind it.
Your human developer gets the same thing as a brief: a scoped, self-contained work packet with intent, constraints, and done-criteria in one place. Nobody starts from a guess. Nobody needs you to re-explain.
When code ships, Artifact compares what got built to what you meant and shows you every mismatch in plain language, on the parts of your product you recognise, not in a diff you can't read.
A mismatch might mean the build is wrong. It might mean your thinking changed and the picture should update. Either way you see it and you decide on it. That's the difference between drifting for a month and correcting in a morning.
What your agent pulls before it writes a line. Every id resolves to something you wrote or confirmed.
Artifact is not a new place to work. It's two steps added to the loop your team already runs.
Your product, on the record, in a surface you already know how to read.
Your product, on the record, in a surface you already know how to read.
Your product, on the record, in a surface you already know how to read.
GitHub is the repo for code. Artifact is the repo for intent. Here's what that means in practice.
Artifact never writes to your codebase and never keeps a working copy of it. Your agent diffs in your session and posts hunks; the server applies its rules, records a finding with a pointer, and discards the rest. What is kept is an index and a citation, never a codebase.
Nodes for features, requirements, acceptance criteria, rules, decisions, open questions. Deterministic ids, integrity checks, provenance and confidence on every node, full version history. Agents traverse the links to assemble exactly the context a task needs.
The whole graph exports, anytime, as plain markdown and YAML. Diff it, grep it, archive it. Your product's intent is never locked in.
Harness packs hold the how: standards, conventions, and taste, compiled into each agent session per tool (CLAUDE.md, .cursor/rules, AGENTS.md) from one rule set, with project overrides and zero footprint on your repo.
Intent, constraints, acceptance criteria, and the boundary of what must hold, in one packet. An agent or a dev builds from it in a fresh session; the result is checked against the graph, not against vibes.
Gaps, contradictions, and drift surface as warnings with explicit overrides, never as hard gates. Artifact has opinions, not veto power.
This is the surface an agent talks to. Every capability is one operation, served identically over MCP and HTTP. Add Artifact as an MCP in your coding agent and it works a real task end to end. Try status, show FEA-0001, drift FEA-0001, work_packet FEA-0001, verify FEA-0001. Add --json to see the payload.
Claude Code, Cursor, and Codex are supported hosts today. All model inference runs on your own subscription; the server stays deterministic.
Artifact is live for Builders in Residence and beta customers: a private instance, connected read-only to your repo, serving Claude Code, Cursor, and Codex over MCP.
More seats, cohort by cohort. The same hosts, the same loop, for teams beyond early access.
Drift on every change, team workflows, and deeper verification of what ships.
Certified partners delivering custom software for small businesses on Artifact.
The next chapter is yours
Explore a different way to make software.
We're inviting early adopters into the conversation.
