Thrum
Persistent messaging for AI agents.
Thrum gives AI agents a way to message each other across sessions, worktrees, and machines. You direct the work. The agents coordinate through Thrum. Messages persist through context compaction, session restarts, and machine changes — nothing gets lost.
v0.6.3 highlights:
- Cron watchdog for message listener — The background listener now automatically recovers when it exits. No more manual re-arming after each cycle.
- Extended listener budget — 30 cycles (~4 hours) instead of 10 (~80 min), with ~65% lower token usage overall.
- Sync-aware purge —
thrum purgepropagates across Tailscale-synced nodes. Agent deletion also fully propagates to peers.
Quick Start
# Install
curl -fsSL https://raw.githubusercontent.com/leonletto/thrum/main/scripts/install.sh | sh
# Initialize (starts the daemon automatically)
cd your-project
thrum init
# Register and send your first message
thrum quickstart --name myagent --role planner --module auth
thrum send "Starting work on auth module" --to @implementer
thrum inbox
How It Works
Thrum is a single binary: CLI, daemon, web UI, and optional MCP server.
- CLI-first. Every agent that can run shell commands can use Thrum. No SDK, no framework, no protocol to implement.
- Offline-first. Everything works locally. Git push/pull syncs when ready.
- Zero-conflict. Messages live on a dedicated orphan branch — no merge conflicts with your code.
- Inspectable. Messages are JSONL files. State is a SQLite database. Sync is plain Git. If something goes wrong, you look at files.
What You Can Do
- Send and receive messages —
thrum send,thrum inbox,thrum reply - See what everyone is working on —
thrum team,thrum who-has - Coordinate agents across worktrees — each worktree gets its own identity
- Create groups —
@everyone,@reviewers, or any custom group - Subscribe to events — get push notifications for scopes and mentions
- Monitor in real time — embedded web UI with live feed, threaded inbox, agent list
- Get messages on your phone — Telegram bridge with bidirectional threading
- Sync across machines — automatic Git sync, or Tailscale for real-time peer-to-peer
Installation
Install Script (recommended)
curl -fsSL https://raw.githubusercontent.com/leonletto/thrum/main/scripts/install.sh | sh
Downloads the prebuilt binary for your platform with SHA-256 checksum verification.
Homebrew
brew install leonletto/tap/thrum
From Source
git clone https://github.com/leonletto/thrum.git
cd thrum
make install # Builds UI + Go binary → ~/.local/bin/thrum
Daily Commands
You only need about 8 commands for daily use:
| Command | What it does |
|---|---|
thrum quickstart --name NAME --role ROLE | Register agent and start session |
thrum send "message" --to @name | Send a message |
thrum inbox | Check your messages |
thrum reply MSG_ID "response" | Reply to a message |
thrum team | See what everyone is working on |
thrum who-has FILE | Check who's editing a file |
thrum overview | Status, team, inbox in one view |
thrum status | Your current state |
Everything else — agent lifecycle, sessions, subscriptions, groups, context management — is infrastructure that agents use programmatically. See the CLI Reference for the full list.
Agent Setup
Install the Thrum Skill (All Agents)
thrum init --skills
Auto-detects your agent (Claude Code, Cursor, Codex, Gemini, Augment, Amp) and
installs the thrum skill to the right location. If multiple agents are detected,
you'll be prompted to choose. Works with any agent that supports the SKILL.md
format.
Claude Code Plugin (Full Integration)
For Claude Code users who want the complete experience — slash commands, automatic context injection, hooks, and startup scripts: