Quick Start
Two paths, same destination: get Anchor into a project and start planning and executing work with disciplined agents.
- For Non-Technical People — plain-language setup using your coding agent (Claude Code or Grok Build).
- For Technical People — clone, CLI flags, endpoints, and the full operator loop.
If you only want the big picture first, skim the one-paragraph version on the intro, then come back here.
For Non-Technical People
You do not need to memorize shell commands. You need:
- A coding agent that supports Anchor skills (Claude Code or Grok Build).
- A copy of the Anchor repository on your machine (or someone who can put it there once).
- The app or project folder you want Anchor to help with.
What Anchor does (in plain English)
Anchor is a process kit, not a new chat model. It teaches your agent to:
- Plan work as short, trackable documents (instead of one endless chat).
- Do one clear job at a time, then check that it really worked.
- Use cheaper / local models for routine steps and save the expensive model for planning and review.
Day to day you mostly type short slash commands into the agent.
Step 1 — Get Anchor on your computer
Ask a technical friend, or open a terminal in a place you keep projects and run:
git clone https://github.com/carefreeinv/anchor.git
cd anchor
Then open that folder in Claude Code or Grok Build.
Step 2 — Make the anchor command available
In the agent chat (still in the Anchor folder), type:
/install-anchor
Follow the prompts. This puts the anchor command on your PATH without needing admin/sudo in the normal case. Details: /install-anchor.
Step 3 — Pick your defaults once
Still in the Anchor checkout, type:
/config
Answer the short survey: which platforms you use (e.g. Claude, Grok), whether you want fleet tooling, model priority, and who should coordinate multi-step work. Your answers are saved so later projects can scaffold without re-deciding everything.
(If /config is not available, ask the agent: “Run ./config.sh and help me through the questions.”)
Step 4 — Turn on Anchor in your project
Open the project you actually work on (your app, website, etc.) in the same agent. Then either:
- Type
/anchorand follow the agent (in a normal project this updates this folder), or - From the Anchor repo, type
/anchor /path/to/your-appso the agent scaffolds that path.
When it finishes, your project has Anchor instructions, plan folders under .plans/, and skills like /draft and /work.
Step 5 — Your daily loop
| You want to… | Type this | What happens |
|---|---|---|
| Capture an idea as a plan | /draft or /draft fix login button | Writes a draft under .plans/drafts/ |
| See drafts | /draft --list | Lists what is waiting |
| Make a draft ready for agents | /draft --promote <name> | Moves it to bugs or features |
| Have the agent do ready work | /work | Picks a fit plan and implements it |
| Sign off finished agent work | /review | You approve or send work back |
| Background workers on a machine | /fleet-watch | Optional; for always-on pullers |
You do not need to understand every folder under .plans/ on day one. Promote drafts when they look clear; run /work when you want implementation; use /review when the agent says work is ready for you.
If something feels stuck
anchornot found → run/install-anchoragain (or open a new terminal after install).- No plans / empty backlog → start with
/draft. - Want the “why” → read The Playbook and Doctrine when you are ready; they are optional for first use.
- Platform-specific install notes → Claude Code, Grok Build.
For Technical People
Operator path: doctrine → defaults → scaffold → plan/execute → optional fleet.
1. Clone and install the CLI
git clone https://github.com/carefreeinv/anchor.git
cd anchor
Register the CLI with /install-anchor (user-local symlink, no sudo), or:
mkdir -p ~/.local/bin && ln -sfn "$(pwd)/bin/anchor" ~/.local/bin/anchor
# ensure ~/.local/bin is on PATH
pip install -e . (or pipx install .) from the repo root also works if you want a packaged anchor command. That install is the scaffolder only — fleet scripts under scripts/ stay copy-paste by design.
Editable install fails with "build backend is missing the 'build_editable' hook"? Old distro pip/setuptools (e.g. Ubuntu 22.04 + setuptools 59) leaking into build isolation. Use a venv (
python3 -m venv .venv && . .venv/bin/activate && pip install -e .),pip install -e . --no-build-isolation, orpipx install ..
2. Read the doctrine
Doctrine (anchor/ANCHOR.md in the repo) is the behavioral contract everything else implements. Skim it before you invent process on top of Anchor.
3. Set defaults
./config.sh
# or /config in Claude Code / Grok Build
Picks default platform(s), fleet tooling, model priority, and preferred orchestrator (who coordinates multi-step work and cross-plan Depends on analysis). Saves under ~/.config/anchor/defaults and prints the exact anchor <project-dir> command to scaffold with those defaults.
Non-interactive example:
./config.sh --platform claude,grok --fleet --language node \
--model-priority nim,claude:sonnet,claude:opus \
--orchestrator claude:opus
Skipping this step is fine — anchor --platform claude,grok (etc.) or the interactive survey still works without saved defaults.
4. Scaffold a project
cd my-app && anchor # current dir + saved defaults
anchor <project-dir> --platform claude,grok # explicit path / platforms
From the Anchor checkout, /anchor <project-path> (path required) runs the CLI with conflict-aware help when the target already has agent config. Inside a scaffolded project, /anchor (no path) locates the local Anchor checkout and conforms this project.
Useful flags — full reference on the CLI page:
| Flag | Purpose |
|---|---|
--fleet | Include fleet/script scaffolding |
--framework <name> | Skip framework detection |
--orchestrator / --set-orchestrator | Preferred planner/coordinator |
--dry-run | Preview without writing |
--check / --diff / --upgrade --yes | Refresh an existing install |
5. Plan and execute
| Command | Role |
|---|---|
/draft | Create, list, load drafts; --promote <slug> → bugs/features |
/work | Execute next (or named) ready plan |
/review | Human sign-off for review-needed/ |
/fleet-watch | Durable multi-tier pullers |
/audit | Security audit → prioritized bug plans |
/deploy | Ship the project with the tooling it already uses |
/optimize | Standards scan → checkbox-picked improvement plans |
Architecture for always-on workers: Fleet workers. Skills map: Skills overview.
6. Endpoints and the full loop (optional)
Point scripts/endpoints.yaml at your OpenAI-compatible endpoints; use scripts/orchestrate.py for plan → execute → critic across them. Serve notes live under Hardware; model quirks stay in anchor_client.py, never in callers.
# examples — see scripts/ and hardware docs for your layout
python scripts/work_once.py --list --tier mid --agent-id worker-1
python scripts/orchestrate.py --help