A vendor-neutral control plane for autonomous AI agents.
MindRail is an early-stage, source-available project exploring a simple problem: increasingly capable agents can execute long tasks, but they still lack a shared external system that tells them what matters now, coordinates work across sessions and runtimes, applies delegated permission policy, preserves state, and decides when a human actually needs to be interrupted.
MindRail aims to provide that control plane without making any single model vendor, agent runtime, cloud provider, or chat session the system of record.
Status: public early development. A deterministic in-memory control-plane vertical slice is implemented and verified. Durable persistence, permission execution, deployed HTTP/MCP transports, cloud runtime, and real agent integrations are still in development.
Human goal
│
▼
MindRail control plane
│
├── context / task / policy
▼
Agent execution
│
▼
Evidence + checkpoint
│
▼
MindRail
│
├── continue
├── re-plan
├── review
└── ask human only when policy requires it
The design separates:
- declarative control — versioned policy, roles, workflows, and project configuration;
- runtime coordination — goals, tasks, sessions, leases, checkpoints, and decisions;
- agent protocols — a small vendor-neutral contract exposed through adapters such as HTTP or MCP;
- integrations — GitHub, dashboards/projections, notifications, and other external systems;
- optional reasoning providers — useful for planning or review, but not trusted to mint authority by themselves.
- Agents are clients, not the system of record. A chat or CLI session may disappear without losing the project state.
- Policy is explicit. Permission decisions should be deterministic and auditable by default.
- State is external to prompts. Agents receive the minimum context needed for the current work.
- Reference implementation is replaceable. Cloudflare is a deployment target for the reference implementation, not part of the protocol definition.
- Evidence over claims. A test or platform behavior is not considered verified until it actually ran.
- YAGNI. New abstractions enter the repository when a real slice needs them.
The repository now contains:
- canonical JSON Schema domain contracts and generated TypeScript bindings;
- accepted runtime/concurrency and protocol v0.1 semantics;
- a deterministic in-memory reference runtime for the first Goal → Task → Lease → Checkpoint → completion loop;
- Lease recovery and fencing, protocol command idempotency, retry, and Task/Goal cancellation tests;
- a replaceable Cloudflare persistence reference design;
- reproducible TypeScript/Vitest quality tooling and GitHub Actions verification.
See docs/CURRENT_STATE.md for the authoritative boundary between implemented and planned capabilities.
AGENTS.md— execution contract for coding agents.docs/00_PROJECT_INDEX.md— context routing index.docs/CURRENT_STATE.md— factual current state.docs/adr/— binding architecture decisions once accepted.docs/architecture/— architecture derived from accepted ADRs.docs/roadmap/— planned work.CONTRIBUTING.md— contribution workflow.
The repository baseline targets Node.js 24 and pins pnpm through package.json.
pnpm install --frozen-lockfile
pnpm check
pnpm test:coverageThe quality gates have been executed successfully on GitHub-hosted Ubuntu with Node 24 and the pinned pnpm version. Exact evidence, test counts, coverage, and remaining limitations are recorded in docs/CURRENT_STATE.md.
MindRail is distributed under Business Source License 1.1 (BUSL-1.1) with project-specific parameters and a future Change License. During the BSL period it is source-available, not OSI Open Source.
The Additional Use Grant is intended to permit internal/self-hosted production use while keeping competing hosted/managed MindRail services outside the grant. BSL does not provide a blanket prohibition on redistribution. The MindRail-specific license parameters have not received professional legal review and should not be treated as legally validated for material commercial enforcement.