PlanWright: Control Plane for AI Coding Agents
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- PlanWright runs inside Claude Desktop and Claude Cowork, synthesizing raw inputs like transcripts, decks, email, and Slack into structured objectives that carry machine-checkable acceptance criteria.
- The product "inverts" planning and acceptance — the agent drafts objectives while humans approve intent, and the system auto-clears mechanical acceptance criteria while routing only judgment-bound items for human sign-off.
- Every lane transition from objective creation through final merge produces an ECDSA P-256-signed, hash-chained audit record, independently verifiable on a public trust page the source describes as "SOC 2 evidence that writes itself."
- Setup is a single command —
claude mcp add planwright --transport http https://mcp.planwright.tools/mcp— with GitHub login handled in-browser and no tokens or config files to hand-edit, per the product page. - Four pricing tiers are listed: Free (up to 3 seats, 30-day retention), Team (4–10 seats, 1-year), Business (11–50 seats, 3-year, custom-domain trust page), and Enterprise (51+ seats, 7-year retention, with SSO/SAML and SOC 2 attestation on the roadmap).
- The page frames compliance timing around the 2026 Trust Services Criteria, which the source says auditors are interpreting to require immutable, tamper-evident trails and documented change-approval controls for AI-driven changes.
Why it matters: If SOC 2 auditors start demanding tamper-evident trails for AI-generated merges, engineering teams that can't produce signed evidence per change could see shipping blocked rather than slowed — and PlanWright is betting that the audit log, not the agent runtime, is the actual moat. The pitch hits two underboted stakeholder pain points at once: developer velocity capped by acceptance review and compliance officers who today get nothing auditable from an agent's PR.
