@weave_protocol/adapter-claudecode

🛡️ @weave_protocol/adapter-claudecode

WARD.md enforcement for Claude Code, via Claude Code's native hook system.

AGENTS.md tells your agent what to do. SKILL.md tells your agent how to do it. WARD.md tells your agent what it can't. This package makes Claude Code listen.

Part of the Weave Protocol security suite. First in a planned series of cross-platform harness adapters (Claude Code → Antigravity → MDASH).

What it does

When Claude Code is about to use a tool (Bash, Edit, Write, Read, WebFetch, etc.), this adapter intercepts via Claude Code's PreToolUse hook, reads your project's WARD.md, and blocks the call if it violates the declared policy.

You: "Delete all the SSH keys"
Claude Code: [about to run Bash with `rm -rf ~/.ssh`]
       ↓
   PreToolUse hook fires
       ↓
   weave-claude-code reads ./WARD.md
       ↓
   checkFilesystem('delete', '~/.ssh/**') → DENY
       ↓
   Returns { "decision": "block", "reason": "🛡️ WARD: ..." }
       ↓
Claude Code refuses, shows reason

No code changes needed in your project. Just a WARD.md file and a one-time hook install.

Quick start

# Install once
npm install -g @weave_protocol/adapter-claudecode

# Wire it into Claude Code's hook config
weave-claude-code init

# Drop a WARD.md in your project root
npx @weave_protocol/ward init

# Done. Open Claude Code in that project — every tool call is now gated by WARD.

Verify with:

weave-claude-code status

How WARD.md is resolved

When the hook fires, it looks for a WARD.md in this order:

• $WEAVE_WARD_PATH — explicit env var override
• <project>/WARD.md — project-level (CWD from the hook payload)
• <project>/.weave/WARD.md — alternate project location
• ~/.claude/WARD.md — user-global fallback (applies to every Claude Code session)

The first one that exists wins. If none exists, the hook silently passes through — zero impact on workflows that haven't adopted WARD yet.

Tool mapping

Claude Code's tools are mapped to generic capability names so a single WARD.md works across platforms:

Claude Code tool	WARD capability	Path arg	Implicit fs op
Bash	shell_exec	(heuristic scan of command)	varies
Edit / MultiEdit / Write	file_write	file_path	write
Read	file_read	file_path	read
Grep / LS	file_read / file_list	path	read / list
Glob	file_list	path	list
WebFetch	http_request	(uses url)	—
WebSearch	web_search	—	—
Task	subagent	—	—
TodoWrite	todo_write	—	—
NotebookEdit	notebook_edit	notebook_path	write

Your WARD.md can use either name — Bash or shell_exec, Edit or file_write. The stricter rule wins.

For Bash, the adapter also runs a light heuristic scan of the command string to catch obvious violations (URLs to denied domains, references to sensitive paths like ~/.ssh, ~/.aws, /etc/).

Commands

weave-claude-code init [--matcher=X] [--fail-closed]
    Install the WARD pre-tool-use hook.
    --matcher=Bash|Edit|...     Restrict to specific tools (default: all)
    --fail-closed               Block on policy errors (default: fail-open)

weave-claude-code disable
    Remove the WARD hook from ~/.claude/settings.json
    (your other hooks are not touched)

weave-claude-code status
    Show hook installation state + active WARD policy for current dir

weave-claude-code test <tool> [--input=JSON]
    Dry-run a tool call against your WARD.md without invoking Claude Code

weave-claude-code help

Examples

Test whether rm -rf ~/.ssh would be allowed:

weave-claude-code test Bash --input='{"command":"rm -rf ~/.ssh"}'

Test a file write to /etc/:

weave-claude-code test Write --input='{"file_path":"/etc/passwd","content":"..."}'

Test a fetch to a disallowed domain:

weave-claude-code test WebFetch --input='{"url":"https://evil.example.com/exfil"}'

Fail-open vs fail-closed

By default, the hook is fail-open: if WARD.md is missing, malformed, or evaluation throws, the tool call is allowed and a warning is written to stderr. This is the right default for adoption — broken policy files shouldn't break your workflow.

For higher-stakes contexts, install with --fail-closed:

weave-claude-code init --fail-closed

Then a missing or broken policy blocks the call.

What gets installed

weave-claude-code init adds an entry to ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "weave-claude-code hook pre-tool-use # WEAVE_WARD_HOOK",
            "timeout": 5
          }
        ]
      }
    ]
  }
}

The # WEAVE_WARD_HOOK marker is how disable finds and removes only our entry without affecting other hooks you've added manually.

A backup of your existing settings.json is written to settings.json.weave-backup before any modification.

A sample WARD.md for Claude Code

---
ward: "1.0"
agent: my-coding-agent
---

# WARD.md

## Filesystem

allow:
  - read: /Users/me/projects/**
  - write: /Users/me/projects/**
deny:
  - read: ~/.ssh/**
  - read: ~/.aws/**
  - write: /etc/**
default: deny

## Network

allow:
  - url: "https://api.github.com/**"
  - url: "https://registry.npmjs.org/**"
default: deny

## Capabilities

allow:
  - file_read
  - file_write
  - file_list
requireApproval:
  - shell_exec
  - http_request
deny:
  - subagent
default: deny

## Behavioral Limits

maxIterations: 50
maxCostUSD: 10.00

With this loaded, Claude Code is allowed to read/write inside ~/projects but cannot touch SSH/AWS keys, will require approval before running shell commands, and is hard-blocked from spawning subagents.

Programmatic use

import {
  resolveWardForCwd,
  evaluateCall,
  installHook,
} from '@weave_protocol/adapter-claudecode';

const resolved = resolveWardForCwd(process.cwd());
if (resolved) {
  const decision = evaluateCall(resolved.policy, 'Bash', { command: 'rm -rf /' });
  console.log(decision.decision); // "deny"
}

License

Apache 2.0 — see LICENSE.