Agent Configurations
Recommended: Install the Thrum plugin instead of manual agent definitions. The plugin provides 10 slash commands, automatic context hooks, and 8 resource docs — all in a single install.
Claude Code agent definitions teach Claude how to use Thrum effectively. These
.md files with YAML frontmatter ship in toolkit/agents/ and load into your
project's .claude/agents/ directory. For Beads task tracking, we recommend
installing the Beads plugin instead.
Install the agent configs
Copy the agent definitions to your project:
mkdir -p .claude/agents
cp toolkit/agents/thrum-agent.md .claude/agents/
cp toolkit/agents/message-listener.md .claude/agents/
Claude Code automatically detects and loads these files when you start a session.
Available agents
thrum-agent
Comprehensive guide for Thrum messaging. Teaches Claude how to register agents, send/receive messages, and coordinate with teammates via the CLI.
Use when:
- Coordinating multi-agent workflows
- Working across worktrees or machines
- Requesting code reviews or assigning tasks
- Broadcasting status updates
Key capabilities:
- Agent registration with roles and intents
- Direct messaging and broadcasts
- MCP server integration with async notifications
- Session management and heartbeats
Beads Plugin (recommended)
For Beads issue tracking, install the Beads plugin instead of using a local agent file. The plugin provides richer functionality:
- 30+ slash commands (
/beads:ready,/beads:create,/beads:close,/beads:sync, etc.) - 15+ resource files covering dependencies, workflows, troubleshooting, and more
- Hooks that auto-run
bd primeon session start for workflow context - Session protocol with CLI reference and resource links
Install the plugin in Claude Code:
/install-plugin beads
Or visit the Beads project for details.
message-listener
Lightweight background listener that watches for incoming Thrum messages so you
don't have to manually check your inbox. Runs on Haiku for cost efficiency
(~$0.00003/cycle). Uses thrum wait for efficient blocking instead of polling
loops — returns immediately when messages arrive, covers ~90 minutes across 6
cycles.
Use when:
- You're running multiple agents and want to know when they message you
- Working long sessions where agents on other worktrees may send updates
- You want incoming messages surfaced without manually running
thrum inbox
Key capabilities:
- Blocking wait via
thrum wait --timeout 8m(6 cycles max, filters by agent identity) - Immediate return on message arrival
- Time-based filtering with
--afterflag (negative value = "N ago"; e.g.,-1sincludes messages sent up to 1 second ago) - CLI-only (no MCP tools — sub-agents can't access MCP)
Configure the message listener
The message-listener runs as a background task so your main agent session stays focused on work while messages are watched for you.
Launch it at session start:
// In Claude Code with Thrum MCP configured
Task({
subagent_type: "message-listener",
model: "haiku",
run_in_background: true,
prompt:
"Listen for Thrum messages. WAIT_CMD=cd /path/to/repo && thrum wait --timeout 8m --after -15s",
});
Wait command flags:
--timeout 8m— Block up to 8 minutes per cycle--after -15s— Include messages sent up to 15 seconds ago (negative = "N ago"; covers re-arm gap)
The listener uses thrum wait which blocks until a message arrives or the
timeout expires — no polling loops needed. Each cycle is a single Bash call.
Re-arm the listener after processing messages to continue listening.
Customize for your project
You can edit these agent definitions to match your team's workflows. Add project-specific commands, adjust priorities, or include custom context.
For the agent file format, see Claude Code agent documentation.
Next Steps
- Claude Code Plugin — the recommended approach: install the plugin to get 10 slash commands and automatic hooks instead
- Codex Plugin — the Codex skill bundle for users who prefer skill-guided CLI workflows
- Agent Coordination — multi-agent messaging patterns that these agent definitions enable
- Workflow Templates — pre-built skill pipelines for the full research → plan → implement cycle