forge

name

forge

description

Create new OpenClaw agents OR dispatch coding agents to visible terminal sessions. Use when: (1) user asks to create/spin up/forge a new agent, (2) user asks to build/dispatch/delegate a coding task to an agent, (3) user says 'have an agent work on this', (4) user wants parallel coding agents on separate tasks. Triggers: 'create agent', 'new agent', 'spin up agent', 'forge agent', 'build this', 'dispatch', 'send an agent', 'have an agent work on', 'use codex for this'.

Forge

Two capabilities in one skill: Create new OpenClaw agents, or Dispatch coding agents to visible terminal panes.

---

Decision Tree

When the user's request comes in, determine which mode:

Is the user asking to create/spin up a new persistent agent?
  YES → CREATE mode (forge.py)
  NO  →
    Is the user asking to build/code/fix something that should be delegated?
    Or explicitly saying "dispatch", "send an agent", "have an agent work on"?
      YES → DISPATCH mode (forge-pane.sh)
      NO  → Handle normally (not a Forge task)

Signals for CREATE:

• "create agent", "new agent", "spin up agent", "forge agent", "bootstrap agent"
• "clone agent", "make an agent like X"
• Discussing agent names, models, personalities, Telegram bindings

Signals for DISPATCH:

• "build this", "have an agent work on this", "dispatch", "send an agent to"
• Any coding task the main agent should delegate rather than do inline
• "use codex/claude for this" (agent override)
• "do these in parallel" (multi-dispatch)

---

CREATE — New OpenClaw Agent

Creates a fully operational OpenClaw agent with workspace, identity files, config registration, auth profiles, and model routing.

Parameters

Gather these from the user (ask only what's missing):

Parameter	Required	Default	Notes
name	yes	—	Lowercase, becomes agent ID
primary_model	yes	—	Format: provider/model (e.g., openai/gpt-5.4)
fallback_models	no	[]	Comma-separated fallback chain
personality	no	minimal template	Path to custom SOUL.md
telegram_bot_token	no	skip binding	For Telegram integration
clone_from	no	fresh agent	Clone identity from existing workspace

Execution

python3 scripts/forge.py \
  --name <agent-name> \
  --model <primary-model> \
  [--fallbacks <model1>,<model2>] \
  [--clone-from <existing-workspace-path>] \
  [--personality <path-to-soul.md>] \
  [--telegram-token <bot-token>] \
  [--openrouter-key <key>] \
  [--force]

The script path is relative to this skill's install directory. Use the full path based on where the skill is installed.

What It Creates

1. Workspace at ~/clawd-<name>/ with identity files (SOUL.md, IDENTITY.md, MEMORY.md, USER.md, AGENTS.md, TOOLS.md, HEARTBEAT.md)
2. .env with API keys (auto-detected from environment)
3. ~/.openclaw/agents/<name>/agent/ with auth-profiles.json and models.json
4. Updates ~/.openclaw/openclaw.json to register the agent
5. Optionally adds Telegram binding

Post-Create

After forging, tell the user:

• Gateway restart needed: openclaw gateway restart
• If Telegram token provided: send /start to the bot
• The agent is reachable via Telegram or webchat

Model Format

All models use provider/model-name format. Examples:

• openai/gpt-5.4
• anthropic/claude-opus-4-6
• deepseek/deepseek-r1
• google/gemini-2.5-pro

Non-Anthropic models auto-route through OpenRouter.

---

DISPATCH — Send Coding Agents to Terminal Panes

Spawns coding agents in visible tmux terminal panes so the user can watch them work in real-time.

How It Works

1. Detect the best available coding agent (or use the user's override)
2. Create a tmux pane with a descriptive name
3. Open Terminal.app so the user sees it live
4. Run the agent with the given prompt
5. Wait after completion so the user can read the output

Multiple dispatches tile automatically in the same tmux window.

Agent Detection

Run scripts/forge-detect.sh --best to find the recommended agent. Priority order:

1. claude (Claude Code CLI)
2. codex (OpenAI Codex CLI)
3. aider (Aider)
4. opencode (OpenCode)

The user can override with "use codex for this" or the --agent flag.

To see all installed agents: scripts/forge-detect.sh

Single Dispatch

scripts/forge-pane.sh "<pane-name>" "<prompt>" [--agent <name>] [--cwd <path>]

Example — the user says "have an agent add auth to the API":

scripts/forge-pane.sh "add-auth" "Add JWT authentication middleware to the Express API in src/api/" --cwd /path/to/project

Parallel Dispatch

When the user wants multiple tasks done at once, dispatch multiple panes:

scripts/forge-pane.sh "backend-auth" "Add JWT auth to the API" --cwd /path/to/project
scripts/forge-pane.sh "frontend-login" "Build a login page component" --cwd /path/to/project
scripts/forge-pane.sh "test-suite" "Write integration tests for the auth flow" --cwd /path/to/project

All three appear as tiled panes in one Terminal window.

Agent Override

The user can specify which agent to use:

• "use codex for this" → --agent codex
• "dispatch with aider" → --agent aider
• "have claude work on the backend" → --agent claude

Pane Management

scripts/forge-pane.sh list              # Show active panes
scripts/forge-pane.sh kill <pane-name>  # Kill specific pane
scripts/forge-pane.sh kill-all          # Kill entire session

Headless Mode

For background tasks that don't need a visible window:

scripts/forge-pane.sh "<name>" "<prompt>" --no-window

Crafting Good Prompts

When dispatching, write clear, self-contained prompts. The dispatched agent has no context from the current conversation. Include:

• What to build/fix/change
• Where the relevant code lives (file paths)
• How it should work (acceptance criteria)
• Constraints (don't change X, use Y library, etc.)

Bad: "fix the bug" Good: "Fix the TypeError in src/api/auth.ts:45 where req.body is undefined. The body-parser middleware runs in app.ts but the auth route mounts before it. Reorder the middleware chain so body-parser runs first."

---

Examples

Create a new agent

User: "Forge a new agent called scout on deepseek/deepseek-r1"

python3 scripts/forge.py --name scout --model deepseek/deepseek-r1

Create with fallbacks and Telegram

User: "Create an agent called oracle on gpt-5.4 with gemini-2.5-pro fallback, and hook it up to this Telegram bot: 123456:ABC"

python3 scripts/forge.py \
  --name oracle \
  --model openai/gpt-5.4 \
  --fallbacks google/gemini-2.5-pro \
  --telegram-token "123456:ABC"

Clone an existing agent

User: "Clone an agent onto a different model"

python3 scripts/forge.py \
  --name scout-experimental \
  --model deepseek/deepseek-r1 \
  --clone-from ~/clawd-scout

Dispatch a single task

User: "Have an agent add dark mode to the settings page"

scripts/forge-pane.sh "dark-mode" \
  "Add dark mode toggle to src/components/Settings.tsx. Store preference in localStorage. Use the existing Tailwind dark: classes." \
  --cwd /path/to/project

Dispatch with agent override

User: "Use codex to refactor the database layer"

scripts/forge-pane.sh "db-refactor" \
  "Refactor src/db/ to use Drizzle ORM instead of raw SQL queries. Migrate all existing queries." \
  --agent codex --cwd /path/to/project

Parallel dispatch

User: "Split this into three parallel tasks: API, frontend, and tests"

scripts/forge-pane.sh "api" "Build REST endpoints for /users CRUD in src/api/users.ts" --cwd /project
scripts/forge-pane.sh "frontend" "Build the user management UI in src/pages/Users.tsx" --cwd /project
scripts/forge-pane.sh "tests" "Write tests for the users API in tests/users.test.ts" --cwd /project

---

Architecture Reference

See references/agent-anatomy.md for the full file/directory structure of an OpenClaw agent.