Skip to content

README.md

Latest commit

 

History

History
134 lines (91 loc) · 6.56 KB

README.md

File metadata and controls

134 lines (91 loc) · 6.56 KB
Raw

Coder

The runtime for agents, your rules, any model, everywhere.

Coder turns any model into an agent you own: an always-on teammate in Slack, GitHub, Linear and your other apps. It does the work and replies in the same thread. Run it in Coder Cloud, or hand it tasks right from your local harness.

Why Coder

  • Always on. Your agents run in the cloud, connected to GitHub, Slack and any other app, each task in its own workspace with only the access you grant.
  • In your harness too. Hand tasks off to subagents from Claude Code, Codex or another harness. Your conversation stays about intent, and only the result comes back.
  • Your agents, your team. Build agents with their own instructions, events and tools, and share them, with their credentials, across your workspace.
  • Built-in review. Review any diff with a read-only agent that checks every finding against your repo.
  • Any model. Claude, Codex or any other model, from your subscriptions, an API endpoint, or your own machine. Pick per task.
  • Fully controlled. One permission model across every engine, live visibility into every task, and steering mid-run.
  • Flows and SDK. Orchestrate whole waves of tasks from one TypeScript file, or drive everything from your own code.

Get started

I want to Start with
Run my agents always on in the cloud, connected to GitHub, Slack and any other app Coder Cloud
Hand tasks off from Claude Code, Codex or another harness In your local harness
Script tasks, reviews and flows from code From code

Coder Cloud

Your agents, always on at coder.wular.ai. Connect GitHub, Slack or any other app, and they work right inside it. No download, no setup:

  1. Sign in at coder.wular.ai.
  2. Connect a platform: GitHub, Slack, Linear or any other app in the catalog.
  3. Mention @coder in an issue, pull request or thread. It does the work and replies in the same thread.
  • Your agents. Instructions, events and tools, set in the dashboard or in .coder/agents/ in your repo.
  • Every app, one dashboard. GitHub, Slack and the rest of the integration catalog.
  • Scoped access. Each task gets a fresh workspace and a platform token limited to that task.
  • A built-in agent. coder answers questions, reviews pull requests, and opens them for requested changes.

Dashboard · Agents

In your local harness

Install.

npm install -g @wular/coder

Add it to your host.

Host Install
Claude Code coder setup-host claude
or, inside Claude Code:
/plugin marketplace add muzam1l/coder
/plugin install coder@coder-plugins
/coder:setup
Codex, Pi, OpenCode and any host that reads ~/.agents/skills coder setup-host agents

coder setup-host adds Coder's MCP server as well; more in setup-host.

Connect a model. Coder uses the engines you're signed in to, or your own model (Models).

npm install -g @openai/codex && codex login
# and/or
npm install -g @anthropic-ai/claude-code && claude auth login

Use it. Ask your host:

Use Coder to explain the directory structure of the workspace.

To make it the default, add this line to your project's AGENTS.md or CLAUDE.md:

Always use Coder for implementation and exploration.

Every task runs under the access you set and stays yours to watch, steer or stop. See them all in the browser:

coder dash

Recommended setup: Claude Code as the host (Fable, low or medium effort) and Codex (sol) as the engine, the best split of performance and cost.

Models · Configuration · Approvals · Connections

From code

Your host drives tasks itself, but everything is scriptable:

coder run "<text>"                     # dispatch a task (--wait to block for the answer)
coder list                             # recent tasks (running + just stopped)
coder result [task-id]                 # status + final answer (--wait blocks until done)
coder task steer <task-id> "<text>"    # continue a task with new instructions
coder task ask <task-id> "<question>"  # ask about a task without interrupting it
coder task stop <task-id>              # interrupt it
coder task watch [task-id]             # stream the live transcript
coder review --pr 42 --post            # read-only review, posted inline

Any [task-id] defaults to the most recent task.

For a repeatable workflow, ask your host to write a flow:

Create a Coder flow that fixes every failing test file: one task per file, gated on its tests passing.

A flow is one TypeScript file that fans out tasks, with verification gates and resume after a crash:

// .coder/flows/fix-tests.ts
import { z } from 'zod';
import { task, gate, pipeline } from '@wular/coder/flow';

const failing = await task('Run `bun test` and list the failing test files.', {
  returns: z.object({ files: z.array(z.string()) }),
});

export default await pipeline(
  failing.data.files,
  file => task(`Fix the failing tests in ${file}.`, { name: file }),
  (r, file) => gate(`bun test ${file}`),
);

CLI · Flows · Review · SDK

Docs

Everything else is at coder.wular.ai/docs.

coder --help lists every command, and coder docs [topic] prints the bundled docs. coder upgrade keeps the CLI and host plugins current.