OpenLoom

Don't trust the agent's word — trust the file system.

OpenLoom is a webhook-driven task harness for OpenCode. It gives OpenCode a JSON task API, a session monitor, webhook fan-out, a web dashboard, and a file-inbox storage runner — so you can dispatch work from anywhere, even a phone.

OpenLoom does not replace OpenCode. It fills the gaps in its HTTP API.

Install

Requires Python 3.11+ and a running OpenCode server.

uv tool install openloom

That's it. Server deps, python-docx, requests — all bundled. No --with, no [extras].

Quick start

1. Start OpenCode

opencode serve

2. Configure (optional)

Defaults work out of the box with a local OpenCode on port 4096. For custom setups:

openloom init

This writes ~/.openloom/config.yaml. Edit it, or use env vars:

export OPENLOOM_OPENCODE_URL=http://127.0.0.1:4096
export OPENLOOM_OPENCODE_USERNAME=opencode
export OPENLOOM_OPENCODE_PASSWORD=your-password

3. Start OpenLoom

openloom serve

Web dashboard at http://127.0.0.1:8967. Verbose logging: openloom serve -v.

Webhook API

Dispatch a task from any HTTP client, CI pipeline, or messaging bot:

curl -X POST http://127.0.0.1:8967/api/tasks \
  -H "Content-Type: application/json" \
  -d '{"name": "fix-typo", "workspace": "/path/to/project", "goal": "Fix the typo in README.md"}'

Endpoints:

Method	Path	Description
POST	/api/tasks	Create and dispatch a task
GET	/api/tasks	List all tasks
GET	/api/tasks/{id}	Get task status
POST	/api/tasks/{id}/abort	Abort a stuck session
DELETE	/api/tasks/{id}	Delete a task
GET	/api/tasks/{id}/events	SSE stream of task events

Task payload

{
  "name": "unique-task-name",
  "workspace": "/absolute/path/to/project",
  "goal": "Natural-language instruction for OpenCode"
}

Webhook notifications

OpenLoom fires webhooks on task lifecycle changes. Configure in config.yaml:

bus:
  webhook_url: https://your-service.com/hooks/openloom
  webhook_secret: "shared-secret"

Event payload:

{
  "event": "TASK_UPDATED",
  "task_name": "fix-typo",
  "task_id": "abc123",
  "status": "completed",
  "timestamp": 1719676800,
  "timestamp_iso": "2024-06-29T12:00:00+00:00",
  "recent_activity": [...]
}

Storage Runner

Drop files into an inbox directory and OpenLoom dispatches them as tasks automatically.

Setup

Add a storage level to config.yaml:

levels:
  storage:
    path: ./inbox           # directory to watch
    poll_interval_s: 5      # check every 5 seconds
    class: myapp.MyConnector  # optional connector class

Supported formats

Extension	Format
.yaml / .yml	Task spec (name, workspace, goal)
.md	Markdown — first heading as name, body as goal
.txt	Plain text
.docx	Word documents (auto-detected)

Connectors

Storage connectors fetch task files from external sources. A connector is a Python class with a download(path) -> bytes | None method. Put it in ./connectors/:

# connectors/mysource.py
import requests
from openloom.levels.storage.base import Connector

class MyConnector(Connector):
    def __init__(self):
        self._session = requests.Session()
        self._session.proxies = {"http": None, "https": None}  # bypass system proxy

    def download(self, path: str) -> bytes | None:
        url = f"https://my-server/files/{path}"
        r = self._session.get(url, timeout=30)
        r.raise_for_status()
        return r.content

Configure in config.yaml:

levels:
  storage:
    class: mysource.MyConnector

Dashboard

openloom serve serves a Svelte dashboard at http://127.0.0.1:8967:

• Dashboard — live task overview
• Tasks — create, browse, abort, delete tasks
• Activity — event stream with session messages
• Session drawer — inspect OpenCode sessions in real time

Configuration

Run openloom init for a starter ~/.openloom/config.yaml. Key sections:

opencode_url: "http://127.0.0.1:4096"
server:
  port: 8967
bus:
  webhook_url: ""
  webhook_secret: ""
levels:
  storage:
    path: ./inbox
    poll_interval_s: 5

Full reference: docs/configuration.md.

Requirements

• Python 3.11+
• OpenCode running with HTTP API enabled