Connect coding agents

Connect a coding agent that speaks the Agent Client Protocol (ACP), such as Claude Code. An agent sends it a task. The coding agent works in a git worktree, asks before each change, and commits on a branch of its own.

This page runs Claude Code end to end. Requires Node.js, git, and Claude Code signed in or an Anthropic API key.

wardrik init offers steps 1 and 3 when it finds Claude Code: it shows the install command, and writes the harness and a coding policy. Answer it without a terminal with --set coding.claude=true.

1. Install the ACP adapter

npm install -g @agentclientprotocol/claude-agent-acp
claude-agent-acp --version

2. Make a repository

mkdir -p ~/acp-demo && cd ~/acp-demo
git init -b main
printf 'defmodule Demo do\n  def hello, do: "Hello"\nend\n' > demo.ex
printf '# Demo\n\nOne module, demo.ex. Keep changes small.\n' > AGENTS.md
git add . && git commit -m "First"

3. Configure the harness

~/.config/wardrik/config.yaml

harnesses:
  claude:
    command: claude-agent-acp
    env: {ANTHROPIC_API_KEY: "${ANTHROPIC_API_KEY}"}
    meta: {claudeCode: {options: {settingSources: []}}}
    options: {mode: default}
    context: [CLAUDE.md, AGENTS.md]
checkouts:
  demo: ~/acp-demo

systemctl --user restart wardrik
wardrik doctor            # checks the command and each checkout

• settingSources: [] stops Claude Code from loading ~/.claude/settings.json and the repository's .claude/, so neither one can allow a tool.
• mode: default makes Claude Code ask before each change. Its own default, auto, may not ask.
• Without env, Claude Code uses its own sign-in from ~/.claude.
• See all settings in config.yaml › harnesses.

4. Define an agent

~/.config/wardrik/agents/builder.md

---
description: Sends coding work to Claude Code.
tools: [claude]
policy: coding
---
You send coding work to the claude tool, then say what it did.

~/.config/wardrik/policies/coding.yaml

allow: ["claude/execute(command = mix test)"]
deny: [claude/delete]
default: {read_only: allow, else: ask}

• Each harness is a tool of the same name.
• Each request from Claude Code is a call named by its ACP kind: claude/read, claude/edit, claude/execute, claude/delete, claude/fetch, and so on.
• A rule matches a request's input, such as command, and its paths relative to the repository: claude/edit(lib/*).

5. Ask for a change

wardrik reload
wardrik ask builder "Use claude on the demo checkout, work add-greet: add greet/1 to demo.ex, returning \"Hello, NAME!\", then commit it."

→ claude {"checkout":"demo","task":"Add greet/1 ...","work":"add-greet"}

╭─ claude__edit ─────────────────────────────────────────╮
│ {                                                      │
│   "file_path": ".../demo.ex",                          │
│   "new_string": "...def greet(name), do: ...",         │
│   ...                                                  │
╰────────────────────────────────────────────────────────╯
Approve claude__edit? [y/n, Enter leaves it]

Type y to approve each request. Type n to refuse one, with a reason Claude Code receives.

• wardrik approval goes through the requests that wait, showing each one's arguments.
• In Slack, a notice shows each request with Approve and Deny.

6. See the result

wardrik agent show builder          # its ID, agt_...
cd ~/acp-demo
git log --oneline main..wardrik/AGENT_ID/_/add-greet

• The branch is wardrik/<agent ID>/<conversation>/<work>, from the checkout's base. main is unchanged.
• The worktree is in ~/.local/share/wardrik/worktrees/claude/demo/.
• Asking again with the same work name continues the same session and branch.

7. Merge it

git merge --squash wardrik/AGENT_ID/_/add-greet && git commit -m "Add greet/1"

A sweep runs each day. It removes each worktree whose work is in its checkout's base, and the branch with it. Merged, squashed and cherry-picked work all count. Work with changes or untracked files stays.

Run without asking

A coding agent can work with no person in the loop, in one of two ways.

Wardrik decides each request. Claude Code still asks, and the policy answers:

~/.config/wardrik/policies/autonomous.yaml

allow: [claude/*]
deny: [claude/delete, "claude/execute(command = git push*)"]

~/.config/wardrik/agents/builder.md

---
tools: [claude]
policy: autonomous
---

• No request asks a person. Deny rules still refuse, and every decision is in the agent's events.
• After Claude Code fetches a web page, gated requests ask again. List under allow_after.outside the ones that may still run. See Outside content.
• The work stays on its branch until you merge it.

The coding agent decides each request. Set its own mode:

~/.config/wardrik/config.yaml

harnesses:
  claude-open:
    command: claude-agent-acp
    meta: {claudeCode: {options: {settingSources: []}}}
    options: {mode: bypassPermissions}

• Claude Code sends no requests, so no policy sees its commands. The worktree, the branch, the timeout and the cost are its only limits.
• acceptEdits allows edits and still asks before commands.
• Claude Code refuses bypassPermissions as root outside a sandbox.

Start work from a remote branch

checkouts:
  core: {path: ~/src/core, base: origin/main, fetch: true}

• base is where each new branch starts, whatever branch the checkout is on.
• fetch: true runs git fetch before each new branch and before each sweep.

Which orgs can use a harness

A harness is available to the deployment's org. List it under orgs for other orgs:

orgs:
  globex:
    harnesses: [claude]

A harness runs on the daemon's machine, with its checkouts. Give it only to orgs you trust with that machine.