hatch3r

Crack the egg. Hatch better agents.

hatch3r is an open-source CLI and editor plugin (Claude Code + Cursor) that installs a tool-agnostic agentic coding setup into any repository. Audited each release across 24 governance domains and generated for 3 platform adapters (Claude Code, Cursor, GitHub Copilot). One command gives you the full set of agents, skills, rules, commands, hooks, and MCP integrations -- optimized for your coding tool of choice (live counts in governance/inventory.json ). Selective init installs only what you need based on your project type and team size.

v1.9.0 scope cut: As of 1.9.0 hatch3r supports only Claude Code, Cursor, and GitHub Copilot. Twelve adapters were removed in a hard cut; canonical content is now read from the bundled npm package (no .agents/ materialization in user repos), and the manifest moved to .hatch3r/hatch.json. See CHANGELOG.md for the full breaking-change list and migration notes.

Quick Start

Requires Node.js 22+.

npx hatch3r init --default  # recommended: zero-prompt setup with the standard profile
npx hatch3r init            # interactive: customize profile, tools, and CLI tools (6 prompts for GitHub greenfield; +1 for Azure DevOps, +1 for `custom` preset, +1 for workspace mode)

--default generates a working standard-profile setup with no questions — the fastest path to a configured repo. The interactive init detects your repo, infers your project context (greenfield/brownfield, solo/team), then walks platform → repo identity → content profile (minimal/standard/full/custom) → tools → CLI-tools picker, and generates everything. MCP is not prompted — opt in with --mcp or npx hatch3r mcp setup later. The platform (GitHub, Azure DevOps, or GitLab) is auto-detected from your git remote either way. Run into issues? See Troubleshooting.

Already using Cursor? Carry your existing rules across with npx hatch3r init --import cursor — they land under .hatch3r/overrides/rules/ (.md + .mdc) with per-file conflict reporting. See Migrating from another tool.

What You Get

| Category | Count | Highlights | |----------|-------|-----------| | Agents | 29 | Code reviewer, lint-fixer, dependency auditor, implementer (sub-agentic), fixer, researcher, architect, DevOps, handoff loader / preparer, 9 content-quality specialists (UI/UX/security/reliability/testability/scalability/performance/maintainability/enhancability), and more | | Skills | 53 | Bug fix, feature implementation, issue workflow, release, incident response, context health, cost tracking, handoff prepare / resume, recipes, API spec, CI pipeline, migration, customization, board lifecycle, ad-hoc orchestration scaffold, 5 standalone CLI-tool skills (ripgrep, jq, gh, fd, fzf) + a 24-tool cli-toolbox, and more | | Rules | 70 | Code standards, testing, API design, observability, theming, i18n, security patterns, agent orchestration, fan-out discipline, right-sizing, deep context analysis, handoff readiness, mobile + backend stack rules, and more | | Commands | 31 | Board management, planning (feature, bug, refactor, test), workflow, quick-change, bug-pipeline, revision, debug, healthcheck, security-audit, onboard, benchmark, handoff (prepare/resume/list/complete/prune), and more | | CLI tools | 29 across 3 tiers | Tier-1 default (ripgrep, fd, jq, yq, gh, delta, bat, sd, ast-grep, zstd); tier-2 conditional (Playwright, duckdb, qsv, taplo, glab, az-devops, Docker, llm, fzf, lazygit, difftastic); tier-3 opt-in (RTK, Stagehand, aichat, mods, Comby, miller, csvkit, Podman) -- emitted as per-tool canonical skills + a decision-tree overview | | MCP Servers | 10 (opt-in) | Playwright, Context7, Filesystem, GitHub, Brave Search, Sentry, Postgres, Linear, Azure DevOps, GitLab -- pure opt-in since 2.0.0: init --mcp or npx hatch3r mcp setup (interactive init does not prompt for MCP; features.mcp defaults to false) | | Platforms | 3 | GitHub, Azure DevOps, GitLab -- auto-detected from git remote |

Supported Tools (3 Adapters)

| Tool | Output | |------|--------| | Cursor | .mdc rules, agents, skills, commands, MCP config | | GitHub Copilot | instructions, prompts, GitHub agents | | Claude Code | CLAUDE.md, skills, .mcp.json |

Platform is auto-detected from your git remote during hatch3r init. All board commands, agents, rules, and skills adapt to your selected platform.

How It Works

.hatch3r/                              <- hatch3r footprint in your repo
  ├── hatch.json                       <- Manifest
  ├── overrides/                       <- User-authored canonical overrides (escape hatch)
  ├── learnings/                       <- /learn-captured project knowledge
  ├── handoffs/                        <- Cross-session handoff bundles
  └── mcp/mcp.json                     <- Resolved MCP server config

.claude/                               <- Generated (Claude Code adapter) + CLAUDE.md at repo root
.cursor/                               <- Generated (Cursor adapter)
.github/copilot-instructions.md        <- Generated (Copilot adapter, plus .github/instructions, .github/prompts, .github/agents)
.worktreeinclude                       <- Generated (worktree isolation)

Canonical content (agents, skills, rules, commands, hooks) lives inside the bundled npm package -- adapters read from there directly, so end-user repos no longer contain a .agents/ mirror. The only hatch3r-managed directory in your repo is .hatch3r/. hatch3r can also manage multiple git repos from a single workspace root -- see the Workspace guide.

Where canonical content lives: to inspect the exact agent/rule/skill files a generated output was produced from, look inside the installed package at node_modules/hatch3r/dist/content/ (the agents/, skills/, rules/, commands/, hooks/, mcp/ subtree). A global install resolves under your npm global prefix (npm root -g → hatch3r/dist/content/). This bundled tree is read-only and is the single source of truth for canonical inputs.

Workflow

hatch3r provides a full project lifecycle, from setup to release:

1. Initialize -- npx hatch3r init detects your repo and platform, asks about profile, tools, and CLI tools, generates agents/skills/rules/commands. For headless CI, pass --yes (add --mcp to also configure MCP servers).
2. Set up the board -- hatch3r-board-init creates or connects a Projects V2 board with status fields, label taxonomy, and config writeback.
3. Define work -- Create a todo.md at the project root (one item per line).
4. Fill the board -- hatch3r-board-fill parses todo.md, classifies items, groups into epics, builds a dependency DAG, and marks issues status:ready.
5. Groom the backlog -- hatch3r-board-groom surfaces stale items, priority imbalances, and decomposition candidates.
6. Pick up work -- hatch3r-board-pickup auto-selects the next issue by dependency order and priority, creates a branch, delegates implementation, and opens a PR.
7. Review cycle -- Reviewer + fixer agents loop (max 3 iterations) until clean, then testability and security specialists run final checks.
8. Release -- hatch3r-release determines the semver bump, generates a changelog, tags, and publishes.

After init: For greenfield, run hatch3r-project-spec then hatch3r-roadmap. For brownfield, run hatch3r-codebase-map. For a single feature, run hatch3r-feature-plan. For small changes, run hatch3r-quick-change.

Commands

CLI Commands

npx hatch3r init          # Interactive setup (or --default for zero prompts)
npx hatch3r setup [dir]   # Scaffold a new project (mkdir + git init) then run init
npx hatch3r config        # Reconfigure tools, MCP servers, features, and platform
npx hatch3r sync          # Re-generate from canonical state
npx hatch3r update        # Pull latest templates (safe merge)
npx hatch3r status        # Check sync status between canonical and generated files
npx hatch3r validate      # Validate bundled canonical content + on-disk adapter outputs
npx hatch3r verify        # Drift check on adapter outputs (non-zero exit on drift)
npx hatch3r clean         # Remove generated files (optional --reinit)
npx hatch3r worktree-setup <path>   # Set up gitignored files in a worktree
npx hatch3r worktree-cleanup <path> # Clean up worktree-specific files
npx hatch3r cli-tools     # Manage CLI tools (picker / list / install / detect)
npx hatch3r mcp           # Manage MCP servers (setup / list / remove / env-check)
npx hatch3r add <pack>    # Install a community pack (coming soon)

hatch3r cli-tools and hatch3r mcp are side-door entry points for users who skipped a section during init. cli-tools defaults to the picker (list, install, detect are the other subcommands); mcp requires a subcommand (setup, list, remove <id>, env-check).

Every non-stub command accepts --format <human|json> and --quiet; mutating commands add --dry-run. --format json emits exactly one JSON document on stdout and is an exit-2 usage error on a prompting invocation.

Agent Commands

Invoked inside your coding tool (e.g., as Cursor commands). All are prefixed hatch3r-.

• Board management: board-fill, board-pickup (lifecycle helpers board-init, board-groom, board-refresh, board-shared ship as skills)
• Planning: project-spec, codebase-map, roadmap, feature-plan, bug-plan, refactor-plan, migration-plan, test-plan, api-spec
• Workflow: workflow, quick-change, revision, debug, onboard, benchmark, hooks, learn, pr-resolve, handoff
• Operations: healthcheck, security-audit, report (helpers dep-audit, release, context-health, cost-tracking, recipe ship as skills)
• Customization: create (single hatch3r-customize skill covers agent/command/skill/rule customization)

See the CLI Commands reference and Agent Commands reference for full details.

CLI Tools

Since 1.7.5, hatch3r ships a first-class CLI-tools surface as the token-efficient alternative to MCP. The picker runs during init (3 tiers grouped, tier-1 default-on, tier-2 conditional on detected project signals, tier-3 opt-in). Detection probes each tool via command -v / where with a 2s timeout; the installer prints copy-paste commands grouped by package manager and never executes on your behalf. 5 essentials (ripgrep, jq, gh, fd, fzf) ship as standalone skills; the remaining 24 tools live in a single hatch3r-cli-toolbox skill, emitted to all 3 adapters. Manage at any time via npx hatch3r cli-tools [list|install|detect]. See CLI Tools for the full 29-tool table and the trade-off discussion vs MCP.

MCP Configuration

Since 1.7.5, MCP is opt-in; since 2.0.0 interactive npx hatch3r init no longer offers an MCP prompt. Without an opt-in, init skips MCP entirely -- no .env.mcp, no mcp.json, no servers in the manifest, and features.mcp stays false. When you opt in (init --mcp on any init path, or npx hatch3r mcp setup afterwards), hatch3r writes a gitignored .env.mcp with the required environment variables and MCP config to the tool-appropriate location (.cursor/mcp.json, .mcp.json, .vscode/mcp.json).

• VS Code / Copilot: secrets pass via the env object in .vscode/mcp.json.
• Cursor / Claude Code / others: source the file first: set -a && source .env.mcp && set +a && cursor .

Manage MCP at any time via npx hatch3r mcp setup | list | remove <id> | env-check. CI note: interactive init no longer offers MCP, and npx hatch3r init --yes does not configure it by default -- opt in via npx hatch3r mcp setup or init --mcp. See MCP Setup for per-server details and PAT scope guidance.

Content Profiles

During hatch3r init, you choose a content profile:

| Profile | What's included | Best for | |---------|----------------|----------| | Minimal | Core agents and workflows only (core tag) | Quick setup, minimal footprint | | Standard (recommended) | Full lifecycle; drops AI feature engineering + performance capability clusters -- pick Full if you need those | Most projects | | Full | Everything including board management and all audits | Large teams, full coverage | | Custom | Choose exactly what you need | Fine-grained control |

After init, use hatch3r config to add or remove individual content items.

Sub-Agentic Architecture

A four-phase pipeline (Research, Implement, Review Loop with reviewer + fixer at max 3 iterations, Final Quality with testing + security) drives implementation. The implementer agent receives a single sub-issue and returns code + tests; the fixer agent applies targeted fixes from reviewer findings; the issue-workflow skill runs an 8-step flow with parallel sub-agent delegation for epics. Tooling hierarchy: project docs > codebase search > library docs (Context7) > web research. See Agent Teams for delegation patterns; hatch3r is complementary to rule-distribution tools like Ruler — it owns the full generation pipeline rather than distributing a single instruction file.

Why hatch3r vs just AGENTS.md?

AGENTS.md (Linux Foundation AAIF spec, 60K+ repos as of January 2026) is the greatest-common-denominator markdown standard for agent instructions; it is consumed by 20+ tools including Cursor, Copilot, Codex, and Gemini CLI. hatch3r is complementary: AGENTS.md describes one file's content; hatch3r owns the entire generation pipeline that emits tool-native configurations across 5 artifact classes (rules, skills, commands, hooks, MCP servers) for 3 supported platforms (Claude Code, Cursor, GitHub Copilot). Three measurable differences:

• Scope: AGENTS.md is one flat instruction file per repo; hatch3r generates platform-specific structured output (.mdc rules with frontmatter scoping for Cursor, CLAUDE.md with managed blocks for Claude Code, .github/instructions/ + .github/prompts/ for Copilot) plus board commands, MCP server configs, and event-driven hooks.
• Currency: AGENTS.md content is hand-edited per project; hatch3r ships canonical content (29 agents + 66 rules + 53 skills + 31 commands + 7 hooks + 10 MCP servers — see governance/inventory.json) audited weekly across 24 governance domains.
• Adoption path: AGENTS.md remains the spec hatch3r-emitted Cursor / Claude / Copilot configurations align with — the 1.9.0 hard-cut withdrew direct AGENTS.md emission per CONSTITUTION §6 Decision #12, but AAIF spec evolution feeds per-adapter feature work for the 3 supported adapters. Use AGENTS.md alone when one flat file suffices for your project; use hatch3r when you need the full content + tooling stack.

Customization

hatch3r separates managed from custom files:

• hatch3r-* files are managed and fully replaced on update; files without the prefix are your customizations and are never touched.
• All hatch3r-generated markdown uses managed blocks (<!-- HATCH3R:BEGIN --> / <!-- HATCH3R:END -->); content outside the markers is preserved.
• User-authored canonical overrides live under .hatch3r/overrides/ (escape hatch); adapters prefer overrides over bundled content.
• Configure preferred AI models per agent via hatch.json (global default + per-agent overrides), agent frontmatter, or .hatch3r/agents/{id}.customize.yaml. Resolution order: customization file > manifest per-agent > agent frontmatter > manifest default. See Model Selection.

Editor Plugin

hatch3r ships plugin manifests for both Claude Code (.claude-plugin/marketplace.json) and Cursor (.cursor-plugin/plugin.json). Install it straight from this repo as a marketplace — in Claude Code, run /plugin marketplace add hatch3r/hatch3r then /plugin install hatch3r@hatch3r — for access to all rules, skills, agents, and commands without running init.

Documentation

Full documentation is at docs.hatch3r.com.

• Vision -- Framework north-star vision and principles
• MCP Setup -- Connecting MCP servers and managing secrets
• Adapter Capability Matrix -- Per-tool support and output paths
• Agentic Process -- Visual diagrams of init flow, board workflow, and agent orchestration
• Troubleshooting -- Common issues and solutions
• Changelog -- Release history

License

MIT