hcom connects Claude Code, Gemini CLI, and Codex. When one agent runs a command, edits a file, or sends a message, others find out in real-time through hooks, across terminals or across devices.
pip install hcomRun agents with hcom in front:
hcom claude
hcom gemini
hcom codexTell agent in natural language:
send a message to claude
Open the TUI:
hcomAgents have structured access to each other's transcripts, live terminal screens, command history, file edits, and event logs. You describe what you want, the agent runs hcom commands.
send gemini the API decisions from claude's plan
Messages arrive mid-turn (injected after tool calls) or wake idle agents immediately.
agents → hooks → sqlite → hooks → other agents
Hooks capture agent activity into sqlite. Other hooks and PTY push matching events back into agent context. Agents subscribe to what they care about and get notified inline.
when any agent runs git commit do something good
Collision detection is on by default. If 2 agents edit the same file within 20s, both get notified.
Agents can launch other agents into new terminal windows, tabs, or panes.
hcom 3 codex # open 3 codex instances
hcom f <name> # fork agentspawn gemini to write tests for this feature then kill it
Use any terminal emulator. See hcom config terminal --info
Kitty and wezterm work well (agents can auto open and close panes). tmux works in the background.
Bundled scripts that combine launching, messaging, and subscriptions:
clone -- fork the current agent into a new terminal with a task. Result comes back via hcom message.
watcher -- background reviewer that subscribes to an agent's work, reviews the diff every turn boundary, and sends 'lgtm' or flags issues back via hcom.
confess -- honesty self-evaluation based on OpenAI's confessions paper. The target agent writes a confession report, a calibrator generates an independent report from transcript only, a judge compares and returns a verdict. --fork runs in the background.
debate -- a judge sets up a structured debate where agents choose sides with shared context of transcript ranges and workspace files. Rounds, rebuttals, verdict.
Run with hcom run <script>. Create new workflows by telling agent: "read hcom run docs then make a script that does X"
| Agents | When messages arrive | Why |
|---|---|---|
| Claude Code (including subagents) | idle + mid-turn | many hooks |
| Gemini CLI (v0.26.0+) | idle + mid-turn | many hooks |
| Codex | idle + hcom listen |
1 hook |
| Any AI tool that can run shell commands | manual (hcom start, hcom listen) |
no hooks |
run this command:
hcom start
hcom send <message> --from bot-namehcom claude -p 'do task' # detached background (manage via TUI)run 2x task tool and get them to talk to each other in hcom
Connect agents across machines through a private HuggingFace Space (free):
# local
hcom relay hf <HF_TOKEN>
# remote/cloud
pip install hcom && hcom relay hf <HF_TOKEN> && hcom start -hHooks go into ~/ (or HCOM_DIR) on launch or hcom start. If you aren't using hcom, the hooks do nothing.
hcom hooks remove # safely remove only hcom hooks
hcom status # install statusHCOM_DIR=$PWD/.hcom # for sandbox or project localCLI
# hcom CLI Reference
hcom (hook-comms) v0.6.14 - multi-agent communication
Usage:
hcom TUI dashboard
hcom <N> claude|gemini|codex [args] Launch agents (args forwarded to tool)
hcom <command> Run command
Commands:
send Send message to your buddies
listen Block until message or event arrives
list Show agents, status, unread counts
events Query event stream, manage subscriptions
bundle Structured context packages for handoffs
transcript Read another agent's conversation
start Connect to hcom (run inside any AI tool)
stop Disconnect from hcom
kill Terminate agent + close terminal pane
config Get/set global and per-agent settings
run Execute workflow scripts
relay Cross-device communication
archive Query past hcom sessions
reset Archive and clear database
hooks Add or remove hooks
status Installation and diagnostics
term View/inject into agent PTY screens
Identity:
1. Run hcom start to get a name
2. Use --name <name> on all hcom commands
Run 'hcom <command> --help' for details.
## events
Usage:
Query the event stream (messages, status changes, file edits, lifecycle)
Query:
events Last 20 events as JSON
--last N Limit count (default: 20)
--all Include archived sessions
--wait [SEC] Block until match (default: 60s)
--sql EXPR Raw SQL WHERE (ANDed with flags)
Filters (same flag repeated = OR, different flags = AND):
Core:
--agent NAME Agent name
--type TYPE message | status | life
--status VAL listening | active | blocked
--context PATTERN tool:Bash | deliver:X (supports * wildcard)
--action VAL created | started | ready | stopped | batch_launched
Command / file:
--cmd PATTERN Shell command (contains, ^prefix, $suffix, =exact, *glob)
--file PATH File write (*.py for glob, file.py for contains)
--collision Two agents edit same file within 20s
Message:
--from NAME Sender
--mention NAME @mention target
--intent VAL request | inform | ack
--thread NAME Thread name
Time:
--after TIME After timestamp (ISO-8601)
--before TIME Before timestamp (ISO-8601)
Shortcuts:
--idle NAME --agent NAME --status listening
--blocked NAME --agent NAME --status blocked
Subscribe (hcom notification when event matches):
events sub List subscriptions
events sub [filters] Create subscription with filter flags
--once Auto-remove after first match
--for <name> Subscribe for another agent
events unsub <id> Remove subscription
Examples:
events --agent peso --status listening
events --cmd git --agent peso
SQL reference (events_v view):
Base id, timestamp, type, instance
msg_* from, text, scope, sender_kind, delivered_to[], mentions[], intent, thread, reply_to
status_* val, context, detail
life_* action, by, batch_id, reason
type message, status, life
msg_scope broadcast, mentions
msg_sender_kind instance, external, system
status_context tool:X, deliver:X, approval, prompt, exit:X
life_action created, ready, stopped, batch_launched
delivered_to/mentions are JSON arrays — use LIKE '%name%' not = 'name'
Use <> instead of != for SQL negation
## list
Usage:
hcom list All alive agents, read receipts
-v Verbose (directory, session, etc)
--json Verbose JSON (NDJSON, one per line)
hcom list [self|<name>] Single agent details
[field] Print specific field (status, directory, session_id, ...)
--json Output as JSON
--sh Shell exports: eval "$(hcom list self --sh)"
hcom list --stopped [name] Stopped instances (from events)
--all All stopped (default: last 20)
Status icons:
â–¶ active processing, reads messages very soon
â—‰ listening idle, reads messages in <1s
â– blocked needs human approval
â—‹ inactive dead or stale
â—¦ unknown neutral
Tool labels:
[CLAUDE] [GEMINI] [CODEX] hcom-launched (PTY + hooks)
[claude] [gemini] [codex] vanilla (hooks only)
[AD-HOC] manual polling
## send
Usage:
Usage:
send @name -- message text Direct message
send @name1 @name2 -- message Multiple targets
send -- message text Broadcast to all
send @name Message from stdin (pipe or heredoc)
send @name --file <path> Message from file
send @name --base64 <encoded> Message from base64 string
Everything after -- is the message (no quotes needed).
All flags must come before --.
Target matching:
@luna base name (matches luna, api-luna)
@api-luna exact full name
@api- prefix: all with tag 'api'
@luna:BOXE remote agent on another device
Underscore blocks prefix: @luna does NOT match luna_reviewer_1
Envelope:
--intent <type> request | inform | ack
request: expect a response
inform: FYI, no response needed
ack: replying to a request (requires --reply-to)
--reply-to <id> Link to event ID (42 or 42:BOXE)
--thread <name> Group related messages
Sender:
--from <name> External sender identity (alias: -b)
--name <name> Your identity (agent name or UUID)
Inline bundle (attach structured context):
--title <text> Create and attach bundle inline
--description <text> Bundle description (required with --title)
--events <ids> Event IDs/ranges: 1,2,5-10
--files <paths> Comma-separated file paths
--transcript <ranges> Format: 3-14:normal,6:full,22-30:detailed
--extends <id> Parent bundle (optional)
See 'hcom bundle --help' for bundle details
Examples:
hcom send @luna -- Hello there!
hcom send @luna @nova --intent request -- Can you help?
hcom send -- Broadcast message to everyone
echo 'Complex message' | hcom send @luna
hcom send @luna <<'EOF'
Multi-line message with special chars
EOF
## bundle
Usage:
hcom bundle List recent bundles (alias: bundle list)
hcom bundle list List recent bundles
--last N Limit count (default: 20)
--json Output JSON
hcom bundle cat <id> Expand full bundle content
Shows: metadata, files (metadata only), transcript (respects detail level), events
hcom bundle prepare Show recent context, suggest template
--for <agent> Prepare for specific agent (default: self)
--last-transcript N Transcript entries to suggest (default: 20)
--last-events N Events to scan per category (default: 30)
--json Output JSON
Shows suggested transcript ranges, relevant events, files
Outputs ready-to-use bundle create command
TIP: Skip 'bundle create' — use bundle flags directly in 'hcom send'
hcom bundle show <id> Show bundle by id/prefix
--json Output JSON
hcom bundle create "title" Create bundle (positional or --title)
--title <text> Bundle title (alternative to positional)
--description <text> Bundle description (required)
--events 1,2,5-10 Event IDs/ranges, comma-separated (required)
--files a.py,b.py Comma-separated file paths (required)
--transcript RANGES Transcript with detail levels (required)
Format: range:detail (3-14:normal,6:full,22-30:detailed)
normal = truncated | full = complete | detailed = tools+edits
--extends <id> Parent bundle for chaining
--bundle JSON Create from JSON payload
--bundle-file FILE Create from JSON file
--json Output JSON
JSON format:
{
"title": "Bundle Title",
"description": "What happened, decisions, state, next steps",
"refs": {
"events": ["123", "124-130"],
"files": ["src/auth.py", "tests/test_auth.py"],
"transcript": ["10-15:normal", "20:full", "30-35:detailed"]
},
"extends": "bundle:abc123"
}
hcom bundle chain <id> Show bundle lineage
--json Output JSON
## stop
Usage:
hcom stop Disconnect self from hcom
hcom stop <name> Disconnect specific agent
hcom stop <n1> <n2> ... Disconnect multiple
hcom stop tag:<name> Disconnect all with tag
hcom stop all Disconnect all agents
## start
Usage:
hcom start Connect to hcom (from inside any AI session)
hcom start --as <name> Reclaim identity (after compaction/resume/clear)
Inside a sandbox? Prefix all hcom commands with: HCOM_DIR=$PWD/.hcom
## kill
Usage:
hcom kill <name> Kill process (+ close terminal pane)
hcom kill tag:<name> Kill all with tag
hcom kill all Kill all with tracked PIDs
## listen
Usage:
hcom listen [timeout] Block until message arrives
[timeout] Timeout in seconds (alias for --timeout)
--timeout N Timeout in seconds (default: 86400)
--json Output messages as JSON
Filter flags:
Supports all filter flags from 'events' command
(--agent, --type, --status, --file, --cmd, --from, --intent, etc.)
Run 'hcom events --help' for full list
Filters combine with --sql using AND logic
SQL filter mode:
--sql "type='message'" Custom SQL against events_v
--sql stopped:name Preset: wait for agent to stop
--idle NAME Shortcut: wait for agent to go idle
Exit codes:
0 Message received / event matched
1 Timeout or error
Quick unread check: hcom listen 1
## reset
Usage:
hcom reset Archive conversation, clear database
hcom reset all Stop all + clear db + remove hooks + reset config
Sandbox / local mode:
If you can't write to ~/.hcom, set:
export HCOM_DIR="$PWD/.hcom"
Hooks install under $PWD (.claude/.gemini/.codex), state in $HCOM_DIR
To remove local setup:
hcom hooks remove && rm -rf "$HCOM_DIR"
Explicit location:
export HCOM_DIR=/your/path/.hcom
## config
Usage:
hcom config Show all config values
hcom config <key> Get single value
hcom config <key> <value> Set value
hcom config <key> --info Detailed help for a setting (presets, examples)
--json JSON output
--edit Open config in $EDITOR
--reset Reset config to defaults
Per-agent runtime config:
hcom config -i <name> Show agent config
hcom config -i <name> <key> Get value
hcom config -i <name> <key> <val> Set value
-i self Current agent
keys: tag, timeout, hints, subagent_timeout
Global settings:
HCOM_TAG Group tag (agents become tag-*)
HCOM_TERMINAL default | <preset> | "cmd {script}"
HCOM_HINTS Text appended to all messages agent receives
HCOM_SUBAGENT_TIMEOUT Subagent keep-alive seconds after task
HCOM_CLAUDE_ARGS Default claude args (e.g. "--model opus")
HCOM_GEMINI_ARGS Default gemini args
HCOM_CODEX_ARGS Default codex args
HCOM_RELAY Relay server URL (set by 'hcom relay hf')
HCOM_RELAY_TOKEN HuggingFace token (set by 'hcom relay hf')
HCOM_AUTO_APPROVE Auto-approve safe hcom commands (1|0)
HCOM_AUTO_SUBSCRIBE Auto-subscribe presets (e.g. "collision")
HCOM_NAME_EXPORT Export agent name to custom env var
Non-HCOM_* vars in config.env pass through to claude/gemini/codex
e.g. ANTHROPIC_MODEL=opus
Precedence: HCOM defaults < config.env < shell env vars
Each resolves independently
HCOM_DIR: per project/sandbox — must be set in shell (see 'hcom reset --help')
## relay
Usage:
hcom relay Show relay status
hcom relay on Enable cross-device communication
hcom relay off Disable cross-device communication
hcom relay pull Force sync now
hcom relay hf [token] Setup HuggingFace Space relay
--update Update existing Space
Finds or duplicates a private, free HF Space to your account.
Provide HF_TOKEN or run 'huggingface-cli login' first.
Remote agents appear with :SUFFIX (e.g. luna:BOXE).
## transcript
Usage:
hcom transcript <name> View agent's conversation (last 10)
hcom transcript <name> N Show exchange N
hcom transcript <name> N-M Show exchanges N through M
hcom transcript timeline User prompts across all agents by time
--last N Limit to last N exchanges (default: 10)
--full Show complete assistant responses
--detailed Show tool I/O, file edits, errors
--json JSON output
hcom transcript search "pattern" Search hcom-tracked transcripts (rg/grep)
--live Only currently alive agents
--all All transcripts (includes non-hcom sessions)
--limit N Max results (default: 20)
--agent TYPE Filter: claude | gemini | codex
--json JSON output
Tip: Reference ranges in messages instead of copying:
"read my transcript range 7-10 --full"
## archive
Usage:
hcom archive List archived sessions (numbered)
hcom archive <N> Query events from archive (1 = most recent)
hcom archive <N> agents Query agents from archive
hcom archive <name> Query by stable name (prefix match)
--here Filter to archives from current directory
--sql "expr" SQL WHERE filter
--last N Limit events (default: 20)
--json JSON output
## run
Usage:
hcom run List available workflow/launch scripts and more info
hcom run <name> [args] Execute script
hcom run <name> --help Script options
hcom run docs Python API + CLI reference + examples
Docs sections:
hcom run docs --cli CLI reference only
hcom run docs --config Config settings only
hcom run docs --api Python API + scripts guide
User scripts: ~/.hcom/scripts/
## claude
Usage:
hcom [N] claude [args...] Launch N Claude agents (default N=1)
hcom claude Opens new terminal
hcom N claude (N>1) Opens new terminal windows
hcom N claude "initial prompt" Initial prompt (positional)
hcom 3 claude -p "prompt" 3 headless in background
HCOM_TAG=api hcom 2 claude Group tag (creates api-*)
hcom 1 claude --agent <name> .claude/agents/<name>.md
hcom 1 claude --system-prompt "text" System prompt
Environment:
HCOM_TAG Group tag (agents become tag-*)
HCOM_TERMINAL default | <preset> | "cmd {script}"
HCOM_CLAUDE_ARGS Default args (merged with CLI)
HCOM_HINTS Appended to messages received
HCOM_SUBAGENT_TIMEOUT Seconds subagents are keep-alive after task
Resume / Fork:
hcom r <name> Resume stopped agent by name
hcom f <name> Fork agent session (active or stopped)
Run "claude --help" for claude options.
Run "hcom config terminal --info" for terminal presets.
## gemini
Usage:
hcom [N] gemini [args...] Launch N Gemini agents (default N=1)
hcom gemini Opens new terminal
hcom N gemini (N>1) Opens new terminal windows
hcom N gemini -i "initial prompt" Initial prompt (-i flag required)
hcom N gemini --yolo Flags forwarded to gemini
HCOM_TAG=api hcom 2 gemini Group tag (creates api-*)
Environment:
HCOM_TAG Group tag (agents become tag-*)
HCOM_TERMINAL default | <preset> | "cmd {script}"
HCOM_GEMINI_ARGS Default args (merged with CLI)
HCOM_HINTS Appended to messages received
HCOM_GEMINI_SYSTEM_PROMPT Use this for system prompt
Resume:
hcom r <name> Resume stopped agent by name
Gemini does not support session forking (hcom f).
Run "gemini --help" for Gemini CLI options.
Run "hcom config terminal --info" for terminal presets.
## codex
Usage:
hcom [N] codex [args...] Launch N Codex agents (default N=1)
hcom codex Opens new terminal
hcom N codex (N>1) Opens new terminal windows
hcom N codex "initial prompt" Initial prompt (positional)
hcom codex --sandbox danger-full-access Flags forwarded to codex
HCOM_TAG=api hcom 2 codex Group tag (creates api-*)
Environment:
HCOM_TAG Group tag (agents become tag-*)
HCOM_TERMINAL default | <preset> | "cmd {script}"
HCOM_CODEX_ARGS Default args (merged with CLI)
HCOM_HINTS Appended to messages received
HCOM_CODEX_SYSTEM_PROMPT System prompt (env var or config)
Resume / Fork:
hcom r <name> Resume stopped agent by name
hcom f <name> Fork agent session (active or stopped)
Run "codex --help" for Codex options.
Run "hcom config terminal --info" for terminal presets.
## status
Usage:
hcom status Installation status and diagnostics
hcom status --logs Include recent errors and warnings
hcom status --json Machine-readable output
## hooks
Usage:
hcom hooks Show hook status
hcom hooks status Same as above
hcom hooks add [tool] Add hooks (claude | gemini | codex | all)
hcom hooks remove [tool] Remove hooks (claude | gemini | codex | all)
Hooks enable automatic message delivery and status tracking.
Without hooks, use ad-hoc mode (run hcom start inside any AI tool).
Restart the tool after adding hooks to activate.
Remove cleans both global (~/) and HCOM_DIR-local if set.
## term
Usage:
hcom term Screen dump (all PTY instances)
hcom term [name] Screen dump for specific agent
--json Raw JSON output
hcom term inject <name> [text] Inject text into agent PTY
--enter Append \r (submit). Works alone or with text.
hcom term debug on Enable PTY debug logging (all instances)
hcom term debug off Disable PTY debug logging
hcom term debug logs List debug log files
JSON fields: lines[], size[rows,cols], cursor[row,col],
ready, prompt_empty, input_text
Debug toggle; instances detect within ~10s.
Logs: ~/.hcom/.tmp/logs/pty_debug/
Config
# Config Settings Reference
Config is stored in ~/.hcom/config.env (or $HCOM_DIR/config.env).
Commands:
hcom config Show all values
hcom config <key> <val> Set value
hcom config <key> --info Detailed help for a setting
hcom config --edit Open in $EDITOR
Precedence: defaults < config.env < shell environment variables
## HCOM_TAG
HCOM_TAG - Group tag for launched instances
Current value: Use 'hcom config tag' to see current value
Purpose:
Creates named groups of agents that can be addressed together.
When set, launched instances get names like: <tag>-<name>
Usage:
hcom config tag myteam # Set tag
hcom config tag "" # Clear tag
# Or via environment:
HCOM_TAG=myteam hcom 3 claude
Effect:
Without tag: launches create → luna, nova, kira
With tag "dev": launches create → dev-luna, dev-nova, dev-kira
Addressing:
@dev → sends to all agents with tag "dev"
@dev-luna → sends to specific agent
Allowed characters: letters, numbers, hyphens (a-z, A-Z, 0-9, -)
## HCOM_TERMINAL
HCOM_TERMINAL - Terminal for launching new instances
Current value: Use 'hcom config terminal' to see current value
Values:
default Use platform default terminal
<preset> Use a named preset (see list below)
<command> Custom command with {script} placeholder
Available presets:
default Platform default (Terminal.app / wt / gnome-terminal)
Terminal.app macOS Terminal
iTerm macOS iTerm2
Ghostty Fast GPU-accelerated terminal
kitty Auto: split if inside, tab if reachable, else new window
kitty-window Always new kitty OS window
wezterm Auto: split if inside, tab if reachable, else new window
wezterm-window Always new WezTerm OS window
alacritty Minimal GPU-accelerated terminal
ttab Open in new tab (npm install -g ttab)
tmux
tmux-split Split current tmux pane horizontally
wezterm-tab New tab in WezTerm (requires wezterm CLI)
wezterm-split
kitty-tab New tab in kitty (requires kitten CLI)
kitty-split Split pane in kitty (requires kitten CLI)
custom Custom command (see below)
------------------------------------------------------------------------
CUSTOM TERMINAL SETUP
------------------------------------------------------------------------
To use a terminal not in the presets list, set a custom command.
The command MUST include {script} where the launch script path goes.
How it works:
1. hcom creates a bash script with the claude/gemini/codex command
2. Your terminal command is executed with {script} replaced by script path
3. The terminal runs the script, which starts the AI tool
Examples (what the presets are):
ghostty: open -na Ghostty.app --args -e bash {script}
kitty: kitty {script}
alacritty: alacritty -e bash {script}
gnome-terminal: gnome-terminal --window -- bash {script}
wezterm: wezterm start -- bash {script}
Testing your command:
1. Set the terminal: hcom config terminal "your-command {script}"
2. Launch a test: hcom 1 claude
3. If it fails, check that:
- The terminal binary/app exists
- {script} is in the right position
Reset to default:
hcom config terminal default
## HCOM_HINTS
HCOM_HINTS - Text injected with all messages
Current value: Use 'hcom config hints' to see current value
Purpose:
Appends text to every message received by launched agents.
Useful for persistent instructions or context.
Usage:
hcom config hints "Always respond in JSON format"
hcom config hints "" # Clear hints
Example:
hcom config hints "You are part of team-alpha. Coordinate with @team-alpha members."
Notes:
- Hints are appended to message content, not system prompt
- Each agent can have different hints (set via hcom config -i <name> hints)
- Global hints apply to all new launches
## HCOM_SUBAGENT_TIMEOUT
HCOM_SUBAGENT_TIMEOUT - Timeout for Claude subagents (seconds)
Current value: Use 'hcom config subagent_timeout' to see current value
Default: 30
Purpose:
How long Claude waits for a subagent (Task tool) to complete.
Shorter than main timeout since subagents should be quick.
Usage:
hcom config subagent_timeout 60 # 1 minute
hcom config subagent_timeout 30 # 30 seconds (default)
Notes:
- Only applies to Claude Code's Task tool spawned agents
- Parent agent blocks until subagent completes or times out
- Increase for complex subagent tasks
## HCOM_CLAUDE_ARGS
HCOM_CLAUDE_ARGS - Default args passed to claude on launch
Example: hcom config claude_args "--model opus"
Clear: hcom config claude_args ""
Merged with launch-time cli args (launch args win on conflict).
## HCOM_GEMINI_ARGS
HCOM_GEMINI_ARGS - Default args passed to gemini on launch
Example: hcom config gemini_args "--model gemini-2.5-flash"
Clear: hcom config gemini_args ""
Merged with launch-time cli args (launch args win on conflict).
## HCOM_CODEX_ARGS
HCOM_CODEX_ARGS - Default args passed to codex on launch
Example: hcom config codex_args "--search"
Clear: hcom config codex_args ""
Merged with launch-time cli args (launch args win on conflict).
## HCOM_RELAY
HCOM_RELAY - Relay server URL
Set automatically by 'hcom relay hf'.
Custom server: implement POST /push/{device_id}, GET /poll, GET /version
See: https://huggingface.co/spaces/aannoo/hcom-relay/blob/main/app.py
## HCOM_RELAY_TOKEN
HCOM_RELAY_TOKEN - HuggingFace token for private Space auth
Set automatically by 'hcom relay hf'
Or optional authentication token for custom server.
## HCOM_AUTO_APPROVE
HCOM_AUTO_APPROVE - Auto-approve safe hcom commands
Current value: Use 'hcom config auto_approve' to see current value
Purpose:
When enabled, Claude/Gemini/Codex auto-approve "safe" hcom commands
without requiring user confirmation.
Usage:
hcom config auto_approve 1 # Enable auto-approve
hcom config auto_approve 0 # Disable (require approval)
Safe commands (auto-approved when enabled):
send, start, list, events, listen, relay, config,
transcript, archive, status, help, --help, --version
Always require approval:
- hcom reset (archives and clears database)
- hcom stop (stops instances)
- hcom <N> claude (launches new instances)
Values: 1, true, yes, on (enabled) | 0, false, no, off, "" (disabled)
## HCOM_AUTO_SUBSCRIBE
HCOM_AUTO_SUBSCRIBE - Auto-subscribe event presets for new instances
Current value: Use 'hcom config auto_subscribe' to see current value
Default: collision
Purpose:
Comma-separated list of event subscriptions automatically added
when an instance registers with 'hcom start'.
Usage:
hcom config auto_subscribe "collision,created"
hcom config auto_subscribe "" # No auto-subscribe
Available presets:
collision - Alert when agents edit same file (within 20s window)
created - Notify when new instances join
stopped - Notify when instances leave
blocked - Notify when any instance is blocked (needs approval)
Notes:
- Instances can add/remove subscriptions at runtime
- See 'hcom events --help' for subscription management
## HCOM_NAME_EXPORT
HCOM_NAME_EXPORT - Export instance name to custom env var
Current value: Use 'hcom config name_export' to see current value
Purpose:
When set, launched instances will have their name exported to
the specified environment variable. Useful for scripts that need
to reference the current instance name.
Usage:
hcom config name_export "MY_AGENT_NAME" # Export to MY_AGENT_NAME
hcom config name_export "" # Disable export
Example:
# Set export variable
hcom config name_export "HCOM_NAME"
# Now launched instances have:
# HCOM_NAME=luna (or whatever name was generated)
# Scripts can use it:
# hcom send "@$HCOM_NAME completed task"
Notes:
- Only affects hcom-launched instances (hcom N claude/gemini/codex)
- Variable name must be a valid shell identifier
- Works alongside HCOM_PROCESS_ID (always set) for identity
Python API
# hcom Python API Reference
## hcom.session()
Get an identity-bound session for hcom operations.
Args:
name: Instance name. Auto-detects from environment if None.
external: If True, creates external sender (no instance required).
Returns:
Session object with messaging and event methods.
Raises:
HcomError: If name required but not provided or not found.
Examples:
s = hcom.session() # auto-detect
s = hcom.session(name="luna") # explicit instance
s = hcom.session(name="ci", external=True) # external sender
## hcom.instances()
List active instances or get one by name.
Args:
name: Specific instance name, or None for all.
Returns:
If name: dict with keys name, session_id, status, directory, parent_name, tool
If None: list of such dicts
Raises:
HcomError: If name specified but not found.
Examples:
all_instances = hcom.instances()
nova = hcom.instances(name="nova")
## hcom.launch()
Launch AI tool instances.
Two calling modes:
Single launch:
hcom.launch(3, tag="worker", prompt="do task")
hcom.launch(1, tool="gemini", prompt="review code")
Group launch (heterogeneous agents, shared batch):
hcom.launch([
{"tag": "confessor", "prompt": "...", "background": True},
{"tool": "gemini", "tag": "calibrator", "prompt": "..."},
{"tool": "gemini", "tag": "judge", "prompt": "..."},
])
Args:
count: Number of instances (int) or list of spec dicts.
tool: One of 'claude', 'gemini', 'codex' (single mode).
tag: Group tag (single mode).
prompt: Initial prompt (single mode).
system_prompt: System prompt override (single mode).
background: If True, run headless (single mode).
claude_args: Additional Claude CLI args (single mode).
resume: Session ID to resume from (single mode).
fork: If True with resume, fork instead of continue (single mode).
tool_args: Additional tool-specific args (single mode).
cwd: Working directory (single mode).
wait: If True (default), block until all instances are ready or timeout.
timeout: Max seconds to wait when wait=True.
batch_id: Share a batch ID across multiple launch() calls for wait-for-all.
name: Explicit instance name (single mode). Used for resume to reuse
the stopped instance's name. Requires count=1 and name must not
be in use by an active instance.
Spec dict keys (group mode):
count (int): Number of instances, default 1.
tool (str): 'claude', 'gemini', or 'codex', default 'claude'.
tag, prompt, system_prompt, background, claude_args,
resume, fork, tool_args, cwd: Same as single mode args.
Returns:
Single mode: Dict with tool, batch_id, launched, failed, background,
log_files, handles, errors, launch_status (if wait=True).
Group mode: Dict with batch_id, results (list of per-spec dicts),
total_launched, total_failed, launch_status (if wait=True).
Raises:
HcomError: On invalid tool, hook setup failure, or launch failure.
## hcom.bundle()
Manage bundles for context handoff and review workflows.
Bundles package conversation transcript ranges, event IDs, and file paths
into referenceable context units for handoffs between agents.
Args:
action: One of 'list', 'show', 'create', 'chain'.
title: Title for new bundle.
description: Description for new bundle.
events: List of event IDs/ranges for new bundle (e.g., ["123-125", "130"]).
files: List of file paths for new bundle.
transcript: List of transcript ranges for new bundle (e.g., ["5-10", "15"]).
extends: Parent bundle ID for chaining related work.
data: Full bundle dict (alternative to separate fields).
bundle_id: ID for show/chain actions.
last: Limit for list action.
Returns:
list (for 'list', 'chain'): List of bundle dicts.
dict (for 'show'): Bundle details.
str (for 'create'): New bundle ID.
Examples:
# Create a bundle
bundle_id = hcom.bundle("create",
title="Code review: auth module",
description="Implementation complete, ready for review",
events=["123-125", "130"],
files=["auth.py", "tests/test_auth.py"],
transcript=["10-15"]
)
# List recent bundles
bundles = hcom.bundle("list", last=10)
# Get bundle details
details = hcom.bundle("show", bundle_id="abc123")
# Get bundle chain (all related bundles)
chain = hcom.bundle("chain", bundle_id="abc123")
## Session
Identity-bound session for hcom operations.
Provides messaging, events, and transcript access tied to a specific
instance identity. Data is fetched fresh on each call (no caching).
Create via hcom.session():
s = hcom.session() # auto-detect
s = hcom.session(name="luna") # explicit
s = hcom.session(name="bot", external=True) # external sender
### Session.name
Instance name (e.g., 'luna' or 'worker-0').
### Session.info
Fresh instance info from database.
Returns:
Dict with keys:
name (str): Full instance name (may include tag prefix).
session_id (str): Claude session ID for transcript binding.
connected (bool): True if instance exists in DB, False if external.
directory (str): Working directory path.
status (str): Current status ('active', 'listening', 'inactive').
transcript_path (str): Path to transcript file.
parent_name (str): Parent instance name (for subagents).
tool (str): Tool type ('claude', 'gemini', 'codex').
Raises:
HcomError: If instance no longer exists.
### Session.send
Send message to instances.
Args:
message: Message text. Use @name or @prefix- for targeting.
to: Target name (auto-prepends @name if not in message).
intent: One of 'request', 'inform', 'ack'.
reply_to: Event ID to reply to (required for intent='ack').
thread: Thread name for grouping related messages.
bundle: Bundle dict to create and attach. If provided, creates bundle event
and appends bundle summary to message.
Returns:
List of instance names that received the message.
Examples:
s.send("@nova hello")
s.send("@worker- start task", thread="batch-1", intent="request")
s.send("received", to="luna", intent="ack", reply_to="42")
# With bundle
s.send("@reviewer check this", bundle={
"title": "Code review",
"description": "Auth module complete",
"refs": {
"events": ["123-125"],
"files": ["auth.py"],
"transcript": ["10-15"]
}
})
### Session.messages
Get messages for this instance.
Args:
unread: If True, only messages delivered to this instance (mentions or
broadcasts). If False, returns all messages in the system.
last: Maximum number of messages to return (most recent first).
Returns:
List of dicts with keys:
ts (str): ISO timestamp when message was sent.
from (str): Sender's display name.
text (str): Message text content.
mentions (list[str]): Instance names mentioned in message.
delivered_to (list[str]): Instance names message was delivered to.
intent (str, optional): Message intent ('request', 'inform', 'ack').
thread (str, optional): Thread name for grouping messages.
reply_to (int, optional): Event ID this message replies to.
### Session.events
Query the event stream.
Args:
sql: SQL WHERE clause filter (e.g., "msg_from='nova'").
params: Parameters for SQL placeholders (?).
last: Maximum events to return.
Returns:
List of dicts with keys: ts, type, instance, data
SQL fields:
Common: id, timestamp, type, instance
Message: msg_from, msg_text, msg_thread, msg_intent,
msg_reply_to, msg_mentions, msg_delivered_to, msg_bundle_id
Status: status_val, status_context, status_detail
Lifecycle: life_action, life_by, life_batch_id
Bundle: bundle_id, bundle_title, bundle_description, bundle_extends,
bundle_events, bundle_files, bundle_transcript, bundle_created_by
Examples:
s.events(sql="type='message'")
s.events(sql="msg_from=?", params=["nova"])
s.events(sql="msg_thread='task-1'", last=50)
### Session.wait
Block until an event matches the SQL condition.
Args:
sql: SQL WHERE clause to match.
params: Parameters for SQL placeholders (?).
timeout: Seconds to wait before returning None.
Returns:
Matching event dict, or None if timeout.
Examples:
event = s.wait("msg_from='nova'", timeout=60)
event = s.wait("msg_thread=?", params=["task-1"], timeout=120)
### Session.subscribe
Create a push subscription for events.
When matching events occur, a notification is sent via hcom.
Args:
sql: SQL WHERE clause to match events.
params: Parameters for SQL placeholders (?).
once: If True, subscription auto-removes after first match.
Returns:
Subscription ID (e.g., 'sub-a1b2').
Raises:
HcomError: If called from external session (can't receive notifications).
Examples:
sub_id = s.subscribe("msg_thread='task-1'")
sub_id = s.subscribe("msg_from=?", params=["nova"], once=True)
### Session.subscriptions
List all active event subscriptions.
Returns:
List of dicts with keys: id, sql, caller, once
### Session.unsubscribe
Remove an event subscription.
Args:
sub_id: Subscription ID (with or without 'sub-' prefix).
Returns:
True if removed, False if not found.
### Session.transcript
Get conversation transcript for an instance or timeline across all instances.
Args:
agent: Instance name, or "timeline" for timeline mode (all instances).
last: Number of recent exchanges to return.
full: If True, include truncated tool output (use detailed for full output).
range: Exchange range like "5-10" (1-indexed, inclusive). Only valid for specific agent.
detailed: If True, include full tool calls, results, file edits, errors.
Returns:
If agent is name: List of exchange dicts with keys: user, assistant, position, timestamp
If agent is "timeline": List of entry dicts with keys: instance, position,
user, action, timestamp, files, command
Examples:
s.transcript("nova") # nova's transcript
s.transcript("nova", last=5) # nova's last 5
s.transcript("nova", range="1-10") # nova's exchanges 1-10
s.transcript("timeline", last=20) # timeline across all agents
s.transcript("timeline", detailed=True) # detailed timeline
### Session.stop
Stop this instance's hcom participation.
The instance will no longer receive messages or appear in listings.
## hcom.HcomError
Exception raised for hcom errors.
# Creating Custom Scripts
## Location
User scripts (shadow bundled by name):
~/.hcom/scripts
Bundled scripts (reference examples):
<hcom-package>/scripts/bundled
File types:
*.py Python scripts (executable with python3)
*.sh Shell scripts (executable with bash)
## Script Structure Template
# --- python ---
#!/usr/bin/env python3
"""Brief one-line description shown in hcom run list."""
import argparse
import sys
import hcom
def main():
parser = argparse.ArgumentParser(description="...")
parser.add_argument('--target', help='target instance')
parser.add_argument('--name', help='instance identity (optional)')
args = parser.parse_args()
s = hcom.session(name=args.name) if args.name else hcom.session()
s.send("@target hello")
return 0
if __name__ == '__main__':
sys.exit(main())
# ---
## Common Workflow Patterns
Launch agents:
instances = hcom.launch(count=3, tool='claude', tag='worker',
prompt='your task here')
Send messages:
s = hcom.session()
s.send("@luna check this") # to specific instance
s.send("@group- broadcast to group") # to all in group
s.send("message", intent="request") # with envelope
Subscribe to events:
s.subscribe("instance='luna' AND type='status'")
s.wait() # block until event matches
Read transcripts:
exchanges = s.transcript('luna', last=5, detailed=True) # specific agent
timeline = s.transcript('timeline', last=20) # all agents (timeline)
Check messages:
for msg in s.messages():
print(msg['text'])
## Identity Handling
Auto-detect (when run from hcom instance):
s = hcom.session()
Explicit (for testing or specific workflows):
s = hcom.session(name="watcher-luna")
External (script not run by an instance):
s = hcom.session(name="ci-bot", external=True)
Support --name flag in your script:
parser.add_argument('--name', help='instance identity')
s = hcom.session(name=args.name) if args.name else hcom.session()
## Reference Examples
View bundled script sources as working examples:
hcom run clone --source # Spawn clone for a task, result sent back automatically via hcom. (1 fork)
hcom run confess --source # Honesty self-evaluation based on OpenAI's confessions paper. (3 agents)
hcom run debate --source # PRO/CON debaters (fresh or existing agents) + judge evaluate a topic in shared hcom thread. (2+ agents)
hcom run ensemble --source # Ensemble Refinement - multiple agents implement, see each other's work, refine iterativly with a judge. (3-6 agents)
hcom run glue --source # Background glue that watches transcript timeline and connects dots across agents. (1 agent)
hcom run watcher --source # Background code reviewer that subscribes to activity, sends review back via hcom (1 agent)
## Best Practices
- Use hcom API for all hcom operations (not CLI commands in subprocess)
- Support --name flag for identity (allows external callers)
- Scripts can call other scripts: hcom.launch() or subprocess to hcom run
# Examples
View workflow script sources:
hcom run clone --source
hcom run confess --source
hcom run debate --source
hcom run ensemble --source
hcom run glue --source
hcom run watcher --source
Build
# Prerequisites: Rust 1.85+, Python 3.11+
git clone https://github.com/aannoo/hcom.git
cd hcom
pip install -e . --no-build-isolation
# Put rust binary on PATH
ln -sf $(pwd)/src/native/target/release/hcom ~/.local/bin/hcom
# Build (+ copy binary + restart daemon)
./build.sh
# Testing
pytest test/
cargo -C src/native test
# platform wheel in target/wheels/
pip install maturin
maturin build --releaseMIT