csaw 0.9.0

Mount, not install. AI workspace configuration from git-backed registries.

csaw

Multi-source AI workspace governance.
Mount layered AI config — company standards, team conventions, personal preferences — into every repo, without copy-paste or repo pollution.

Works with: Claude Code · Cursor · Codex · OpenCode · GitHub Copilot · Antigravity (Google) · Goose

---

Reversible by design

csaw isn't a daemon, a database, or a wrapper around your AI tools. It's a mount tool: every file csaw creates in your project is a symlink to a file in one of your source repos. csaw unmount removes every symlink and restores any originals that were displaced. There's no state to clean up, no migration to undo, no runtime that has to keep working.

Concretely: uninstall csaw entirely and your AI tools keep reading the symlinks until you delete them. Delete the symlinks and you're back to the project's original state. The cost of trying csaw is the cost of running csaw mount once and csaw unmount once.

This is the core promise — csaw fails open. If you stop trusting it, get out without paying a price.

---

Who csaw is for

You have more than one source of AI configuration:

• Engineers in companies with shared standards — you want company standards, team conventions, and personal preferences composed into every repo, kept live, never committed.
• Teams or staff engineers publishing AI policy — you ship rules and skills that lower layers can't silently override, with audit catching drift.
• Individuals composing personal config with team or community sources.
• Contractors and consultants juggling clients — each engagement has its own MCP servers and conventions that must not bleed across projects.

For team or org use, csaw needs one person to maintain each shared source — a small but real role. Everyone else is a passive consumer.

csaw is the wrong tool if: you have one repo, one team, and no personal additions to layer on top. Commit your AGENTS.md and stop here.

Want the full learning path? See the csaw curriculum.

The problem

Multi-stakeholder AI config is a governance problem:

• No source of truth across projects. A team's AGENTS.md gets copy-pasted into every repo. Each copy drifts independently. The "real" version is whoever pushed last.
• No way to enforce policy. Security mandates a rule. A developer overrides it locally. Nobody notices.
• No isolation between contexts. Personal config in ~/.claude/ applies globally to every project — including client repos that shouldn't see it. Experimental skills dropped into a project to try them out get accidentally committed via git. No way to say "this config belongs here, not there."
• No layering. Team has shared rules; you want personal additions on top. Composing them per repo is manual, so nobody bothers.
• No lineage. Fork a team skill, customize for your style, push improvements back? Manual copy-paste, no record of what diverged.
• Cleanup is impossible. Tried an experimental config; now hunting through 3 tool directories and 6 files.

How csaw works

The model

A source is a git repo (or local directory) of AI config — instructions, rules, agents, skills, MCP files. A project mounts files from one or more sources via symlinks. The mount is reversible, live (edits to a mounted file propagate instantly), and invisible to the project's git.

your-registry/                         your-project/
  AGENTS.md                     ──→      AGENTS.md (symlink)
  rules/                                 .claude/rules/
    go-conventions.md           ──→        go-conventions.md (symlink)
  agents/                                .claude/agents/
    code-reviewer.md            ──→        code-reviewer.md (symlink)
  skills/                                .claude/skills/
    code-review/SKILL.md        ──→        code-review/SKILL.md (symlink)
  mcp/                                   .mcp.json (symlink)
    claude-code.json            ──→

Layer multiple sources with priority. Mark files protected so lower-priority sources can't override them. Pin a source to a branch or tag for one project. Fork a file from one source into another. Audit a project to confirm the right context is active and nothing forbidden is mounted.

The full stack

Imagine you're a software engineer in a department of teams in a company with engineering standards. Your AI workspace has four layers, each owned by someone different:

sources
  company    [priority 100, protected]   AGENTS.md, security rules
  department [priority  80, protected]   PR-workflow skill
  team       [priority  50]              Go conventions, code-review
  personal   [priority  10]              debugging, note-capture
                                                │
                                                │  composed into
                                                ▼
  your-project/        ← live symlinks into all four sources
  your-other-project/  ← same composition, also live
  …every repo you work in

What each layer gives you:

• Company (mandatory, protected). Engineering standards that apply to every repo. Marked protected so no lower layer can silently shadow them. Updated centrally — every repo you've mounted in sees the change on the next csaw pull.
• Department (mandatory, protected). Cross-team conventions like a PR-workflow skill. Higher priority than team, also protected. Add it later without rewriting anything below it.
• Team (shared, not protected). Your team's rules and skills, reused across every repo your team owns. Edit once in the team source, every team repo sees it instantly.
• Personal (just you). Your own debugging skills, note-capture preferences, anything personal. Symlinked into every project but hidden from git — never committed, never visible to teammates, never in any repo.

Two properties emerge from stacking these that no single layer would give you on its own:

• Always in sync. Because mounted files are symlinks, every layer updates live across every repo that consumes it. No PR fan-out, no manual copy, no "we forgot to update repo seven."
• Properly governed. Because layers have explicit priority and some files are protected, the company and department mandates can't be silently overridden by a team or personal layer. csaw audit proves it in CI.

These four tiers form one vertical hierarchy — each layer nests inside the one above it (your team is in your department is in your company). They're not a fixed schema. You can also add horizontal sources that cut across the org chart, opted into by responsibility rather than position.

Example: as a staff engineer, you want every repo you touch — regardless of team — to have cost-awareness rules and cost-analysis skills. That concern isn't company-wide, isn't team-specific, isn't purely personal — it's shared with other staff engineers and managers who have similar responsibilities. Add a staff-eng source at whatever priority fits, and only the people who need it mount it.

Other horizontal sources you might add: client (for consultants — must not bleed across projects), community (open-source skill libraries), incident-response (for on-call rotations), security-team (for security-sensitive reviews).

The two roles

• Maintainer — owns a source. Curates its files, defines its profiles, marks protected files, accepts contributions back. One per source.
• Consumer — adds the source (csaw source add), activates a profile (csaw use), pulls updates (csaw pull). Never edits the source they consume.

A single person can play both roles across different sources (you maintain personal, you consume team). For csaw to work in your org, you need exactly one maintainer per shared source — everyone else can be passive.

The three config files

File	Where	Purpose	Edited by
csaw.yml	inside a source registry	Defines that source's profiles and protected files	Source maintainer
.csaw/policy.yml	inside a project repo (committed)	Declares what context the project requires/blocks/audits	Project lead
~/.csaw/config.yml	each developer's machine	Personal: which AI tools you use, which sources you've added	Each user, locally

When each file appears in the walkthrough below, this is what it is. They are three different files with three different lifecycles — don't conflate them.

Why not just commit your AI config to the project?

If a context file belongs to one repo and is safe to commit there, commit it. csaw is for the cases that don't fit:

• Multiple repos sharing the same standards, without copy-paste drift between them.
• Personal additions layered on top of team config, without polluting the project repo for everyone else.
• Client isolation — keeping personal MCP servers, notes integrations, and one client's config out of another client's project.
• Non-overridable mandates — a security or platform team publishing rules that lower layers can't silently shadow, with audit catching drift.
• Cross-tool projection — write the file once, project it into .claude/, .cursor/, .opencode/, .codex/ automatically.

Each section below earns its keep against this question. Mounted files are invisible to the project's git; unmount removes every symlink and restores any originals.

---

Install

uv tool install csaw

Other install methods

# macOS / Linux
brew install --cask NicholasCullenCooper/tap/csaw

# Windows
scoop bucket add csaw https://github.com/NicholasCullenCooper/scoop-bucket
scoop install csaw

# pipx
pipx install csaw

# Go from source
go install github.com/NicholasCullenCooper/csaw/cmd/csaw@latest

macOS note (Homebrew): If you see "Apple could not verify", run:

xattr -d com.apple.quarantine "$(which csaw)"

This is normal for unsigned CLI tools distributed via Homebrew casks.

---

Starting a New Project

You have a brand new repo with no AI config. Create a personal registry:

csaw init ~/my-ai-config

Or scaffold from a curated preset (see csaw init --list-presets):

csaw init ~/team-config --preset team-go        # protected AGENTS.md, Go conventions, code-review agent
csaw init ~/team-config --preset team-frontend  # TS/React conventions, accessibility, code-review agent
csaw init ~/my-config   --preset solo-engineer  # personal scaffold

✔ initialized registry "my-ai-config"

╭─────────────────────────────────────────────────────╮
│  Register as a source?                              │
│  ▸ Yes       No                                     │
╰─────────────────────────────────────────────────────╯

✔ registered source "my-ai-config" with priority 10

This creates a ready-to-use registry:

~/my-ai-config/
  csaw.yml              ← default profile
  .csawignore           ← optional: custom hide patterns (drafts/, archived/, etc.)
  AGENTS.md             ← your coding rules
  rules/                ← always-on standards
  agents/               ← subagent definitions
  skills/
    code-review/SKILL.md
    commit-message/SKILL.md
    experimental/       ← work-in-progress skills (hidden by built-in convention)

Now activate it in your project:

cd ~/my-project
csaw use my-ai-config/default

╭──────────────────────────────────────────────────╮
│                                                  │
│  mounted                                         │
│                                                  │
│  my-ai-config                                    │
│   ✔ AGENTS.md                                    │
│   ✔ .claude/skills/code-review/SKILL.md          │
│   ✔ .claude/skills/commit-message/SKILL.md       │
│                                                  │
│  3 files mounted · 1 tool dirs                   │
│                                                  │
╰──────────────────────────────────────────────────╯

Your project now looks like this:

my-project/
  src/
  package.json
  AGENTS.md                              ← symlink to ~/my-ai-config/AGENTS.md
  .claude/
    skills/
      code-review/SKILL.md               ← symlink
      commit-message/SKILL.md            ← symlink

Open Claude Code (or Cursor, Codex, OpenCode, Antigravity) — it finds the files automatically. Run git status — nothing shows up. The files are hidden via .git/info/exclude.

Walkthroughs

The full scenario-based walkthrough lives in docs/walkthrough.md. Jump straight to the section that matches your situation:

• I Already Have an AGENTS.md — adopt existing AI config into a registry
• Using a Team Source — share config from a team git repo
• Composing Multiple Sources — layer company + team + personal (the canonical case)
• Protected Files — enforce mandates that can't be overridden
• Auditing Active Context — declare and verify project policy
• Experimental Skills — develop skills before promoting
• Pulling Team Updates — keep in sync
• Sharing Your Changes — push edits back to a source
• Forking a Team File — customize without breaking upstream
• Switching Profiles · Clean Removal · Testing a Branch — and more

The Kinds

csaw treats AI workspace artifacts as seven distinct kinds, each with its own conventions and projection target:

Kind	Registry path	Projects to	When loaded
Instructions	AGENTS.md, CLAUDE.md, GEMINI.md, .goosehints	Project root	Every turn — always in context
Rules	rules/*.md	.claude/rules/, .cursor/rules/, .github/instructions/ (Copilot, suffixed .instructions.md)	Every turn — always-on coding standards
Agents	agents/*.md	.claude/agents/, .opencode/agents/, .github/agents/ (Copilot, suffixed .agent.md)	When invoked — specialized subagent personas
Skills	skills/*/SKILL.md	.claude/skills/, .opencode/skills/, .agents/skills/ (Antigravity + fallback), .codex/skills/	When relevant — on-demand procedural workflows
MCP	mcp/*.json	.mcp.json, .cursor/mcp.json, .vscode/mcp.json	Session start — tool/data connectivity
Hooks	hooks/*	.claude/hooks/	Tool lifecycle events
Ignore	ignore/*	.cursorignore, .cody/ignore, .aiderignore, .tongyiignore	Always — context exclusion

csaw does not project settings (they contain API keys/credentials) or memory (session-state, user-private). See docs/planning/projection-roadmap.md.

Agents vs skills. Both are spawnable, both are markdown with frontmatter. The distinction: an agent defines a persona (a subagent with its own tools, scope, and prompt — Claude's .claude/agents/code-reviewer.md); a skill defines a procedure (a step-by-step workflow loaded only when relevant). Use agents when you want a specialist to take over for a focused task; use skills when you want guidance the main agent can pull in mid-task.

Rules vs instructions. Both are always loaded. The distinction is conventional: instructions (AGENTS.md) are the project-level summary every tool reads; rules are split-out always-on standards organized by topic (e.g., rules/go-conventions.md, rules/security.md).

You can mount selectively by kind:

csaw use team/backend --kind agents          # only agent definitions
csaw use team/backend --kind agents,skills   # agents and skills only
csaw use team/backend                        # all kinds

You write files once in your registry. csaw projects them into every tool's native directory. Mounted files are hidden from git via .git/info/exclude. Use csaw show <path> to make one visible, csaw hide <path> to hide it.

csaw inspect groups mounted files by kind within each source so you can see at a glance what's loaded.

---

Configuring Tools

If csaw can't auto-detect any tool directories in your project on first mount, it asks which AI tools you use:

╭──────────────────────────────────────────╮
│  Which AI tools do you use?              │
│                                          │
│  ● Claude Code                           │
│  ● Cursor                                │
│  ○ OpenCode                              │
│  ○ Codex CLI                             │
│  ○ GitHub Copilot (VS Code + CLI)        │
│  ○ Antigravity (Google)                  │
│  ○ Goose                                 │
│                                          │
│  space toggle · enter confirm            │
╰──────────────────────────────────────────╯

This is saved to ~/.csaw/config.yml and applies to all projects. You can also set it directly:

csaw config set tools claude,cursor

Notes on tool coverage:

• GitHub Copilot (v0.7.0+): csaw projects rules to .github/instructions/<name>.instructions.md and agents to .github/agents/<name>.agent.md (suffix rewriting is automatic — you still write rules/security.md in your registry). Projected files are hidden from git by default like every other csaw projection. If your team wants Copilot's .github/ paths committed for PR review, opt in explicitly with csaw show .github/instructions/* and csaw show .github/agents/*. The single-file .github/copilot-instructions.md alias is not yet projected; AGENTS.md at project root covers Copilot's universal-instructions case.
• AGENTS.md coverage (partial — see below): csaw projects AGENTS.md to your project root, which is read by Factory Droid, Pi, Hermes, Cline, Aider (as a secondary), Continue, Amp, Augment, and ~20 other tools. But several of these have native rules/skills/hooks/MCP directories csaw does not yet project to — e.g., .clinerules/, .factory/skills/, .continue/rules/, .augment/rules/, .amp/plugins/, Aider's CONVENTIONS.md. AGENTS.md gets you the universal floor; deep projection for these tools is on the roadmap, prioritized by user demand. See docs/planning/projection-roadmap.md.
• Gemini CLI: Removed in v0.6.0 (Google sunset 2026-06-18). Migrate to Antigravity — it uses .agents/ (csaw's StandardFallback, served automatically) and reads GEMINI.md.
• Trimmed in v0.6.1: Amazon Q, Kiro, OpenHands, Windsurf, and CodeBuddy were removed from active projection to keep the supported set focused. Re-add on user demand — all remain catalogued in the JSON below.

Full per-tool projection details — paths, fallbacks, MCP schemas, and tools we don't yet target — are catalogued in docs/reference/tool-projection.json.

---

Registry Structure

A csaw source is just a git repo with markdown files:

my-ai-config/
  csaw.yml              ← profiles (which files to mount)
  .csawignore           ← optional: custom hide patterns
  AGENTS.md             ← project guidance (the standard)
  rules/                ← always-on coding standards
    go-conventions.md
    testing-standards.md
    experimental/       ← work-in-progress rules (hidden by convention)
  agents/               ← subagent definitions (separate context windows)
    code-reviewer.md
    planner.md
  skills/               ← on-demand reusable workflows
    code-review/
      SKILL.md
    testing/
      SKILL.md
    experimental/       ← work-in-progress skills (hidden by convention)
      new-idea/
        SKILL.md
  mcp/                  ← MCP server configs
    claude-code.json

Every file is standard markdown — usable with or without csaw.

Profiles

Profiles go in csaw.yml. They define which files to mount:

backend:
  description: Go backend development
  include:
    - AGENTS.md
    - rules/go-conventions.md
    - skills/code-review/**
    - skills/testing/**

frontend:
  extends: backend
  include:
    - rules/react-patterns.md
    - skills/react-testing/**

Profiles support glob patterns and inheritance. extends pulls in everything from the parent. A profile can also compose another source's profile by using a source-qualified parent:

backend-with-my-tools:
  extends:
    - team/backend
  include:
    - skills/debugging/**

If this profile lives in your personal source, skills/debugging/** resolves inside personal, while team/backend resolves inside the team source.

---

Full command reference

Commands

Command	What it does
csaw init [dir]	Scaffold a new registry. --adopt imports from existing project; --preset <name> uses a curated starter (run --list-presets).
csaw source add name url	Add a source (auto-clones remote). Accepts full git URL, local path, or host shorthand gh:org/repo[#ref] (also gl:, bb:). --priority n for conflicts.
csaw source remove name	Remove a source.
csaw source clone name dir	Clone remote source locally for contributing.
csaw source list	List configured sources.
csaw profile list	List available named work modes.
csaw profile show name	Show the resolved profile recipe.
csaw use name	Activate a named profile. Replaces previous mount.
csaw mount	Interactive profile picker.
csaw mount profile name	Mechanical equivalent of csaw use name.
csaw mount paths patterns	Advanced: mount registry paths or globs directly.
csaw mount --profile name	Backward-compatible named profile mount.
csaw mount --restore	Re-mount the previous selection.
csaw unmount [patterns]	Remove mounted files, restore originals.
csaw inspect	Full state: sources, mounts, priorities, pins.
csaw audit [path]	Audit active context against .csaw/policy.yml.
csaw audit --init [path]	Write a starter .csaw/policy.yml.
csaw check	Detect broken links, drifted links, and protected content drift.
csaw update	Repair drifted links.
csaw diff path	Diff a mounted file against its source.
csaw pull [source]	Pull latest from remote sources. --stash for dirty state.
csaw push [source] -m "msg"	Commit and push source changes.
csaw pin source@ref	Pin source to a branch/tag for this project.
csaw unpin source	Unpin, return to default branch.
csaw fork source/path	Copy a file into another source. --into target.
csaw promote source/skills/experimental/name	Promote experimental skill to stable.
csaw config set key value	Set config (tools, default_fork_target).
csaw config list	Show configuration.
csaw show / hide path	Control git visibility of mounted files.
csaw status	Quick summary + flags uncommitted edits in source checkouts (so editing a mounted symlink target doesn't go silent).
csaw mcp sync <target> --from <source>	Merge MCP entries from a source into a tool's shared-config file (e.g., Codex's .codex/config.toml) using a bounded section. Dry-run by default; --apply to write; --remove to roll back. See csaw mcp sync walkthrough.

Key Flags

Flag	Commands	What it does
--profile name	mount	Named profile to mount.
--kind list	use, mount	Filter by kind: agents, skills, rules, mcp, instructions (repeatable).
--force	use, mount	Overwrite conflicts, stash originals.
--keep	use, mount	Add to existing mount instead of replacing.
--tools list	use, mount	Target tools (e.g., --tools claude,cursor).
--restore	mount	Re-mount previous selection.
--include-experimental	use, mount	Mount files under any experimental/ path segment (built-in convention; applies to rules, skills, agents, hooks).
--include-ignored	use, mount	Mount files matched by .csawignore patterns.
--strict	audit	Fail on warnings as well as errors.
--json	audit	Emit a machine-readable audit report.
--init	audit	Write a starter .csaw/policy.yml.
--force	audit	Overwrite an existing policy when used with --init.
--adopt	init	Import existing AI config from current project.
--stash	pull	Stash uncommitted changes before pulling.
--priority n	source add	Source priority (higher wins on conflict).
--into source	fork	Target source to fork into.

Contributing

See CONTRIBUTING.md for workflow, validation, and repo standards.

License

MIT