Generate
Your agent creates the artifact in its own working environment.
ReviewOps for AI-generated artifacts
Be Reviewable.
Review AI-generated artifacts, approve the right changes, and hand clean context back to your agent.
Reviewable gives people a focused place to inspect generated artifacts, make the decisions that matter, and send only those conclusions back to the agent for the next pass.
The demo is a shared, periodically reseeded review showing completed owner decisions — you can comment; deciding stays with the owner.
A clear loop for generated work
People review and decide inside Reviewable. Your agent generates and regenerates outside it, using the handoff as the durable boundary.
Review
People inspect the artifact and leave comments anchored to what they see.
Decide
Owners explicitly accept or discard each proposed change.
Handoff
Approved conclusions become a clean, immutable context snapshot.
Regenerate
Your agent reads that handoff and produces the next artifact outside Reviewable.
A review surface with a hard edge
The product is intentionally narrow: give people real decision control, then make the result safe and useful for their agent.
Anchored decisions
Comments pin to real elements of the artifact; owners accept or discard each one explicitly.
Deterministic handoff
An immutable agent-context/v1 snapshot: decisions in, chat noise out.
Agent-native access
A read-only MCP server your agent connects to, scoped to the reviews you share with it.
No accounts, no lock-in
Capability links, a public demo, and a protocol you can read in full.
The handoff protocol
agent-context/v1 records immutable snapshots of the conclusions an owner has approved, so an agent receives a stable review boundary instead of a live chat transcript.
Each snapshot separates requested changes, constraints, and supporting context into explicit sections that an agent can consume without guessing which discussion is decisive.
Anchors resolve through projection targets, preserving the connection between a conclusion and the real element it addresses.
Version drift is detectable, so an agent can tell when an artifact has changed since the snapshot was made.
feedback_type and severity are reserved fields in v1; they are not a stable signal for consumers to interpret.
Connect your agent
Connecting Claude Code, Codex, or any MCP client gives your agent direct read access to the reviews you choose to share.
It can list them, read approved handoffs, and fetch the artifact to regenerate against — no copy/paste.
Access is scoped by the share links you configure, nothing else.
claude mcp add --transport stdio \
--env REVIEWABLE_MCP_REVIEWS=<paste the demo or your review share URL> \
reviewable -- npx -y reviewable-artifacts-mcpUse reviewer links, not owner links, as agent credentials.