Your team is running Claude Code, Cursor, Lovable, or a stack of internal agents. The build cost has collapsed; the governance gap has opened. Cultivate OS sits above the build layer, with traceable briefs in, inspected Growth out, and the human still in charge of the decisions that matter.
Four claims. Each is something that holds today, in v0.5, against the agentic builders your team is already using.
Work handed to an AI builder carries the Grow Plan, the Seed it serves, the Soil it sits in, and the evidence trail behind both. The agent doesn't guess context; it inherits it. When the result comes back, every line of Growth can name the Seed it came from.
Assay inspects built Growth against the Seed and Soil it was supposed to honour. Where the agent has wandered (over-built, missed a Constraint, ignored an evidence claim) the Inspection surfaces it before the Harvest. A governance surface for AI-generated code, in the team's language.
Cultivate OS does not ship code and does not compete with Claude Code, Cursor, Lovable, or whatever your team adopts next. It is the layer above. Bring the build tool you have already chosen; the Grow Plan and the Inspection don't change.
Every module ships API + MCP + webhooks in the same canonical shape. No SDK lock-in, no per-module surprises. The Field is queryable; events fire on every transition; the MCP server is the same as the human surface, only typed.
Three layers, not one. Cultivate OS sits above the build layer and beside the operational layer. Nothing is replaced.
The Field, the cycle, the four promises. Where Seeds are chosen, briefs are written, Growth is inspected, Yield is read, Soil is turned. The judgement layer.
Claude Code, Cursor, Lovable, your internal agent, your humans. They consume a Grow Plan and produce Growth. We hand off; we inspect what comes back. We do not own this layer.
Linear, Jira, GitHub, your help-centre, your CRM, your data warehouse. They are sources (Inputs and Gleanings) and destinations (Sheaf artefacts, Yield reads). They stay.
Three shapes per module, identical across the six. Nothing bespoke.
Every object in the Field is addressable. Seeds, Grow Plans, Growth, Sheaves, Yield Reports, Gleanings. Cursor-paginated; idempotency keys on writes; one Holding token.
GET /fields/{id}/seeds?state=selectedThe same Field, typed for AI builders via the Model Context Protocol, read-only today. An agent can read the Field, a Grow Plan, and the evidence trail behind it, so it builds with the full context in hand.
cultivate.fields.read · cultivate.growplans.readTransitions fire as they happen: a Seed selected, a Grow Plan opened, a Sheaf bound, a Harvest scheduled. Signed payloads; retried with backoff; a per-subscription secret.
seed.selected · growplan.opened · sheaf.boundDrawn from closed-beta conversations. Honest answers, not deflections.
Cultivate OS is hosted SaaS in 2026; self-host comes later. What we are not is another delivery tool. Linear, Jira, GitHub Projects stay. The Field sits above them and the AI builders sit beside them. Net add to your stack; nothing displaced.
We integrate, we don't compete. A Grow Plan is a portable brief with its evidence trail attached; any AI builder that can consume it can do the work. We ship a reference handoff for Claude Code; the format is open and documented.
Three controls. Pre-build: the Grow Plan locks in Constraints, Measures and evidence claims the agent must honour. Post-build: Assay inspects Growth against the Seed and Soil it was supposed to honour, flags drift, recommends Pruning. Always-on: Growth Rings is an append-only log of every transition. Nothing is retroactive; every decision is timestamped.
The framework is on the product side. Engineering work continues to run in whatever delivery framework you have: Agile, Shape Up, Continuous Delivery, trunk-based flow. Cultivate OS only insists that every piece of Growth can name the Seed it came from and every Seed can name its Nutrients.
The Field is yours. Recommendations export with their evidence trail today; a full one-click Field export (JSON + Markdown) is on the near-term roadmap. No model training on your Field. Per-Holding region pinning is planned for the EU and UK.
The integration surface (the MCP server, the webhook payloads, the Grow Plan handoff format) is open and documented. The operating model is not currently open-source; we are honest about that. If your evaluation depends on it, talk to us early.
We do not run sales demos. We hand a real Grow Plan to Claude Code, walk the Inspection, and let you decide whether the governance is the shape you need.