From forge
How to use Forge to submit headless agent jobs, check status, and retrieve output via MCP tools. Use when the user wants to delegate work to a background agent, run a task on another LLM, or interact with forge.
npx claudepluginhub zate/cc-plugins --plugin forgeThis skill uses the workspace's default tool permissions.
Forge runs tasks on LLM agents in the background via a daemon. Interact through MCP tools.
Designs custom Claude Code agents with strong auto-delegation descriptions, optimal tool access, model selection, and Task delegation patterns. Use for agent creation, troubleshooting memory, or pipelines.
Delegates tasks to external LLMs like Gemini and Qwen with quota management, logging, and error handling. Use for token-heavy tasks exceeding context limits or cheaper processing.
Guides creation and configuration of autonomous agents for Claude Code plugins, covering frontmatter, triggering descriptions, system prompts, tools, teams, permissions, and best practices.
Share bugs, ideas, or general feedback.
Forge runs tasks on LLM agents in the background via a daemon. Interact through MCP tools.
Submit a job. Returns a job ID.
Required: task (string) - the prompt/task description.
Optional:
| Parameter | Type | Description |
|---|---|---|
agent | string | claude (default), gemini, local, lmstudio |
model | string | Agent-specific: sonnet, haiku, opus, gemini-2.5-pro, etc. |
timeout | string | Duration: 30s, 10m (default), 1h |
input_files | string[] | File paths to include as context |
output_format | string | markdown (default), json, yaml, code, raw |
constraints | string[] | Output validation constraints |
tags | string[] | Tags for filtering/organization |
Check job status. Required: job_id (string).
Get completed job result. Required: job_id (string).
Returns: output content, metadata (agent, model, duration, tokens, cost).
List jobs. Optional: status filter (pending, running, completed, failed), limit (default 20).
1. forge_submit → returns job_id
2. forge_status → poll until "completed" or "failed"
3. forge_output → retrieve the result
pending → claimed → dispatched → running → validating → delivering → completed
Any state can transition to failed.
pending, check with forge_status or run /forge:setup.forge_list to monitor.["non-empty", "min-length:100"]).