hcom is a command-line tool for connecting multiple AI coding agents across terminals, enabling them to message each other, observe each other, and spawn new agents.
Purpose: hcom is mainly used to coordinate multi-agent pipelines, run different AI command-line tools as each other's subagents, or simply avoid manually copying and pasting content between multiple agents. It turns multiple AI CLIs into a collaborative system that can communicate with each other, wait for each other, and trigger each other.
Main capabilities:
- Agents can send messages to each other in real time: whether the other party is in the middle of processing a turn or idle, it can immediately be woken up or receive the message.
- Agents can observe each other: view status, conversation history, file edits, live terminal view, and command history.
- You can subscribe and receive notifications when status changes, file edits, conflicts, or specific events occur, and automatically react.
- You can spawn, fork, resume, and terminate agents in any terminal emulator, and also run headlessly.
- hooks record activity to a local SQLite database and deliver messages from that database. hooks are activated only when an agent is started with hcom, so normal use is unaffected.
Supported tools: claude, codex, opencode, copilot, qoder, grok, pi, omp, agy, cursor, kimi, kilo, gemini.
Usage: prefix the command that starts an agent with hcom, then prompt normally. For example, have claude check the work and send fixes; spawn 3 opencode instances to split tasks and collect the results; have an agent fork itself to investigate a bug and report back; give codex its next task when it is idle. Open the TUI dashboard to manage agents. Agents can be prompted in any CLI tool to keep listening for messages. You can also send messages from any process or script, or chain them with hcom events. Each agent runs in a real terminal and can be viewed, scrolled, and interrupted. Any terminal emulator can be used to spawn agents; kitty, wezterm, tmux, zellij, waveterm, cmux, and herdr also support closing panes via hcom kill. You can configure how a custom terminal opens/closes. Agents on different machines can be connected via an MQTT relay, and tokens should be treated like API/SSH keys. You can safely remove all hcom hooks, then delete the binary.
Multi-agent workflows: built-in and user scripts (located in ~/.hcom/scripts/) supports multi-agent modes, for example:
- hcom run confess: one agent (or a background clone) writes an honest self-assessment, a derived calibrator independently reads the target conversation record, and a judge compares the two reports and sends the verdict back via hcom messages.
- hcom run debate: the judge derives and sets up a debate with existing agents, coordinates rounds in a shared thread, all agents can see each other's arguments, and share the context of workspace files and conversation records.
- hcom run fatcow: a headless agent reads every file under a certain path, subscribes to file edit events to stay up to date, and answers other agents on demand.
- hcom run onidle: waits for an agent to become idle, then inputs text into another agent, or starts a new agent with that text as the prompt. Custom scripts: put *.sh or *.py into ~/.hcom/scripts/, and they will be automatically discovered and override built-in scripts with the same name. You can ask an agent to write a script; hcom run docs --scripts is the authoring guide.
Limitations and notes: the materials do not provide pricing, platform restrictions, performance metrics, or specific effectiveness data. Configuration is located at ~/.hcom/config.toml, with precedence default values < config.toml < environment variables. External environment variables passed to each launched agent can be set via ~/.hcom/env. Detailed commands, flags, and configuration references can be viewed via hcom <command> --help or hcom run docs --cli. For the security policy, see SECURITY.md.
