A funnel that compounds context.
Each step turns a conversation into a more concrete artifact — and every artifact feeds the next.
Pressure-test the idea
A streaming discovery chat probes scope, audience, constraints, data and integrations — and records a framework choice.
- Seeds from onboarding
- Surfaces one-way-door decisions early
- Outputs structured requirements + signals
Approve a real plan
Discovery becomes phases → milestones → tasks with acceptance criteria, plus the decisions and risks behind them.
- Priority, dependencies, percent-complete
- Decisions & challenges as a reasoning trail
- You approve; your agents update tasks live over MCP
Generate the bundle
A ten-layer bundle assembled in three stages — Copy, Compose, Adapt — tailored to your coding tool.
- Deterministic floor, one LLM pass on top
- CLAUDE.md · .cursorrules · AGENTS.md · GEMINI.md
- Every file version-stamped
Everything else the brain does.
MCP bridge
An HTTP server that serves project memory and tools to your local Claude Code, Cursor, Codex or Gemini — read tools plus approval-gated writes.
The crew
A coordinator and specialists per project, each with a scoped prompt and allowed tools. Templates versioned and pinned per project.
In-app Planner
A project-scoped assistant at ⌘J that reads and edits tasks, memory and plan — sharing the exact tool registry your MCP client uses.
Live memory
Typed knowledge — facts, decisions, constraints, patterns — compiled into runtime context packs. Relevant signal, not noise.
Provider-neutral Git
A git:* namespace and webhook router. GitHub-native today; GitLab and Bitbucket are class-file drop-ins.
Self-improving
Version-stamped artifacts feed telemetry — which prompts produce bundles humans keep. The foundation for a planner that improves as the platform sees more outcomes.
Guardrails
Every response is validated before it reaches the UI, so you don't get silent data errors. Approval gates keep agents on a short leash.
CLI & admin
A repocrew CLI for the terminal, and an admin panel where prompts, templates and scaffolds are edited and version-pinned.
Architecture for an agent-native, multi-provider future.
RepoCrew makes a few deliberate calls that compound over time — the kind of decisions that are cheap on day one and brutal to retrofit later.
| Dimension | Typical PM + Git integration | RepoCrew |
|---|---|---|
| Agent surface | REST / GraphQL — built for humans wiring integrations | MCP-first — git:create_issue is a callable agent verb |
| Issue link syntax | Proprietary prefixes (e.g. TEAM-123) |
GitHub-native Closes #7 — nothing new to learn |
| Tasks vs issues | One concept, conflated | Tasks (planning) and issues (collaboration), cleanly separate |
| Provider model | Per-provider integrations with different shapes | One git:* namespace, providers are drop-ins |
| Code requirement | Shaped around code teams | Git is optional — serves research, content and ops too |
| Improvement loop | Static templates and rules | Telemetry-driven planner, built to self-improve |
The honest version: mature tools have more polish and feature surface today. RepoCrew's bet is architectural — GitHub-native syntax, MCP-first, and provider-neutrality are the dimensions that compound if the future is agent-native and multi-provider.