A governed control plane for AI software factories. Humans define intent and approve consequential steps; agents do bounded work; deterministic code verifies the result before anything reaches a pull request.
A coding agent finishes a task and reports success. That report is not evidence. At one agent it does not matter much; at twenty it is the whole problem — nobody can tell which changes were actually verified, which policies were respected, or why a given diff exists.
Mission Control treats agent completion and verified success as two different things, and refuses to conflate them.
- Governed intake. A Mission carries a specification and an approved Plan before any agent is dispatched. Consequential steps stop at a human gate.
- Bounded execution. Work runs through one execution-only harness contract with explicit capability admission, local or remote-sandboxed.
- Independent verification. Quality Contracts are checked by code that did not produce the change, fail-closed, before a candidate is accepted.
- Evidence, not assertions. Every WorkOrder carries its traces, approvals, and verification records, and the resulting pull request is exact-current.
Requires Node 20, pnpm 9, and a Convex development deployment.
git clone https://github.com/jaydubya818/MissionControl.git
cd MissionControl
corepack enable && pnpm install
cp .env.example .env.local
pnpm exec convex dev --once # creates/links a Convex dev deployment
pnpm run devSet the generated CONVEX_URL as VITE_CONVEX_URL in .env.local, then open
http://localhost:5173.
pnpm run dev:demo # terminal 1
pnpm run convex:seed:demo:force # terminal 2Open http://localhost:5199/v2/command-center and select Software Factory
Demo (sf-demo) to watch a full Mission move through plan, execution,
verification, and acceptance with no external providers involved.
Constitution → Mission → Specification → Plan → WorkOrder → Context
→ Execution → Verification → Pull Request → Human Acceptance → Learning
Each arrow is a gate, not a handoff. Nothing advances on an agent's word.
Active V1 development. The repository carries a deterministic full-system V1 qualification with documented limitations: governed intent, Factory configuration, worker admission, local and bounded live-remote execution, immutable candidates, independent verification, exact pull-request currentness, canonical WorkOrder acceptance, and a human-gated learning continuation all compose end to end.
That is implementation proof. It is not a claim of fleet-scale production operation or general Remote Sandbox certification. Current public client/backend runtime contract: v33.
Per-capability status, evidence, and promotion gates live in the Capability Maturity Ledger.
| Document | What's in it |
|---|---|
| Full overview | The complete design rationale, capability-by-capability implementation detail, governance model, and security boundaries. Start here if you want the whole argument. |
| Architecture | System architecture and repository map. |
| Run the demo | Step-by-step demo walkthrough. |
| Run commands | Every supported run mode and profile. |
| Golden-path proof | A real Codex-to-GitHub pull request, produced and verified through the browser UI. |
Issues and pull requests are welcome — see CONTRIBUTING.md.
Issues labeled good first issue are scoped small and safe to pick up without
asking first.
MIT — see LICENSE.