Aurora API / Phase 2 · GitHub Integration
Sync Engine
Milestone 2.3
8 tasks
62%
RepoCrew is project memory for your AI coding agents. It runs discovery, planning and bootstrap, then serves live project context — tasks, decisions, architecture — to Claude Code, Cursor, Codex and Gemini CLI over MCP. Your agents start every session already knowing the codebase, so the token budget goes to writing code, not re-discovering it.
Private beta · Bring your own coding tool · MCP-native
The thesis
Most projects fail the cold-start problem: a new engineer — or AI agent — opens the repo with zero context. The README is stale, decisions live in Slack, the plan lives in someone's head. RepoCrew makes the shared memory the product, so the scaffold stays true as the project evolves.
A streaming discovery chat probes scope, audience, constraints, data and integrations — then distills structured requirements, user flows and tech signals the rest of the pipeline builds on.
The planner converts discovery into phases → milestones → tasks with acceptance criteria — capturing the decisions and risks behind them. You review, iterate, and approve. The approved plan is what gathers context for bootstrap.
One click assembles a complete project bundle: docs, coding rules, an agent crew, deploy configs, and the assistant instructions for whichever tool you use — CLAUDE.md, .cursorrules, AGENTS.md, GEMINI.md — plus a wired .mcp.json.
RepoCrew never runs coding agents server-side. It serves project memory, plan state and decisions over the Model Context Protocol — so your local tools do the building with full context.
Each project gets a suggested agent roster with roles, scoped system prompts and allowed tools. Templates are versioned and pinned per project, so an admin edit never silently mutates older work.
A project-scoped assistant that reads and edits your tasks, memory and plan. Strategic actions are approval-gated.
A git:* namespace and webhook router mean GitLab and Bitbucket are class-file drop-ins, not six-month rewrites.
Every artifact is version-stamped. Telemetry shows which prompts produce bundles humans keep — the signal that's designed to make the planner sharper over time.
Name the workspace and project, pick a default stack.
Chat through the idea; requirements and signals are captured.
Review phases, milestones and tasks. Approve or iterate.
Generate the ten-layer bundle for your coding tool.
Work in your IDE; agents pull context live over MCP.
Tasks, decisions and memory sync back, staying true.
Where it sits
Linear, Jira. Great at tasks.
— blind to code & contextClaude Code, Cursor, Copilot. Great at code.
— re-derive context every sessioncreate-next-app, Yeoman. Great at day-zero.
— irrelevant by day-threeRepoCrew's wedge is the lifecycle handoff: a day-zero scaffold plus a live MCP memory that keeps it true as the project evolves.
Start free
Sign up, run discovery, approve a plan, and hand your coding tool a project it already understands.