docs/integrations/claude-code.md
How to use beads with Claude Code.
bd setup claude
This installs:
bd prime --hook-json when a session starts. SessionStart also fires after context compaction, so the same hook refreshes context automatically.CLAUDE.md (skipped if CLAUDE.md is a symlink).By default the hook is written to the project's .claude/settings.json. Variants:
bd setup claude --global # Install to ~/.claude/settings.json instead
bd setup claude --stealth # Stealth mode: flush only, no git operations
bd setup claude --remove # Remove the hook and the CLAUDE.md section
If the beads plugin is enabled, bd setup claude skips writing hooks - the plugin provides its own, and duplicates would run bd prime twice per session.
Add to .claude/settings.json (project) or ~/.claude/settings.json (global):
{
"hooks": {
"SessionStart": [
{
"matcher": "",
"hooks": [
{ "type": "command", "command": "bd prime --hook-json" }
]
}
]
}
}
The --hook-json flag wraps the output in the hook JSON envelope Claude Code expects. No PreCompact hook is needed - SessionStart fires again after compaction.
bd setup claude --check
bd prime injects ~1-2k tokens of contextbd CLI commands directlybd prime refreshes workflow contextbd dolt push syncs changesContext efficiency. MCP tool schemas can add 10-50k tokens to every request; bd prime adds ~1-2k tokens of workflow context - 10-50x less overhead, which means lower cost, lower latency, and better model attention. Prefer CLI + hooks in any environment with shell access; use the MCP server only where the CLI is unavailable, such as Claude Desktop.
Beads doesn't ship or require Claude Skills (.claude/skills/). bd prime already delivers the workflow context, and the workflow fits a simple command set (ready → create → update → close → sync). Skills are also Claude-specific, which would break beads' editor-agnostic approach - the same CLI works in Cursor, Windsurf, and every other shell-capable editor. You can create your own Skills on top of beads, but none are needed.
# Always include description for context
bd create "Fix authentication bug" \
--description="Login fails with special characters in password" \
-t bug -p 1 --json
# Link discovered issues
bd create "Found SQL injection" \
--description="User input not sanitized in query builder" \
--deps discovered-from:bd-42 --json
# Find ready work
bd ready --json
# Start work
bd update bd-42 --claim --json
# Complete work
bd close bd-42 --reason "Fixed in commit abc123" --json
# List open issues
bd list --status open --json
# Show issue details
bd show bd-42 --json
# Check blocked issues
bd blocked --json
# ALWAYS run at session end
bd dolt push
--jsonbd list --json # Parse programmatically
bd create "Task" --json # Get issue ID from output
bd show bd-42 --json # Structured data
# Good
bd create "Fix auth bug" \
--description="Login fails when password contains quotes" \
-t bug -p 1 --json
# Bad - no context for future work
bd create "Fix auth bug" -t bug -p 1 --json
# When you discover issues during work
bd create "Found related bug" \
--deps discovered-from:bd-current --json
# ALWAYS run before ending
bd dolt push
For slash commands and enhanced UX, install the beads plugin:
# In Claude Code
/plugin marketplace add gastownhall/beads
/plugin install beads
# Restart Claude Code
Adds slash commands:
/beads:ready - Show ready work/beads:create - Create issue/beads:show - Show issue/beads:update - Update issue/beads:close - Close issue# Check hook setup
bd setup claude --check
# Manually prime
bd prime
# Force push
bd dolt push
# Check system health
bd doctor
# Initialize beads
bd init --quiet