Agent Orchestration

Log in Get started

Features

What the platform does, feature by feature

Each feature below is in the current release, with links to the documentation that describes exactly how it works — including what has only been tested against stand-ins.

01

Task orchestration

Queue, prioritise and schedule coding tasks onto the right machine.

Read the documentation

Four priority levels with aging, so nothing starves
Dependencies between tasks, checked atomically at claim
Concurrency limits per project, worker, organization, agent and provider
Atomic claims and leases: exactly one worker runs a task

02

Worker fleet

Run agents on your laptops, desktops and build servers from one control plane.

Read the documentation

Native workers for Windows, macOS and Linux that connect outbound — no inbound ports
Repository discovery and cloning on each machine
Signed self-updates with automatic rollback
Keeps working through control-plane outages and buffers events

03

Multiple agents

Claude Code, Codex, Gemini CLI, OpenCode and Aider behind one agent-neutral core.

Read the documentation

Each agent runs on its own login by default
Adapters verified against the real CLIs
Policies prefer, allow or block agents per organization, project or task
Same instructions, knowledge and skills for every agent

04

Multiple providers

Keep tasks running on add-on models when an agent reaches its usage limit.

Read the documentation

OpenRouter, Anthropic, OpenAI, Gemini, Groq, DeepSeek, NVIDIA NIM, Bedrock, Vertex, Azure
Local models with Ollama or LM Studio
A built-in model gateway that lets every agent use them
Keys stay in the worker machine’s credential store

05

Automatic recovery

Survive usage limits, exhausted context, crashes, hangs and lost machines.

Read the documentation

Checkpoints in every project; fresh sessions resume from them
Fallback chain: wait, switch agent, provider or model, or ask
Hang detection that looks at output, files and CPU
Tasks of a lost worker are requeued for another one

06

Capabilities

Skills, MCP servers and plugins, governed by permissions and approvals.

Read the documentation

Your own registry per installation — no mandatory marketplace
Organization, project and task scopes
Trust levels, install policies and permission approvals
Plugins run isolated, with only the permissions they declare

07

Verification

A task is complete when your checks pass — not when the agent says so.

Read the documentation

Tests, type checks, lint and build detected from the project
Browser checks that open every page and fail on console errors
Failures go back to the agent to fix, then to a person
Reports separate verified facts from the agent’s claims

08

Git automation

Commit, push or open a pull request — by policy, and never destructively.

Read the documentation

Task branches; your uncommitted work is never committed
GitHub pull requests and GitLab merge requests
Force-push, reset --hard and branch deletion are refused
Your own GitHub App for sync, repositories and scoped tokens

09

Projects and integrations

Projects with one or many repositories, fed by GitHub, GitLab and webhooks.

Read the documentation

Projects spanning several repositories, worked on together
Issues and /agent comments become tasks
Automatic pull request reviews, posted as comments
Signed webhooks from any system, with callbacks

10

AI project manager

Break a large goal into dependent, verifiable tasks — then create them with one click.

Read the documentation

Plans of up to 30 tasks with dependencies, checked for cycles
Plans run in a separate worktree and change nothing
A person decides: nothing is created until approved
Created tasks wait for their dependencies automatically

See it on your own repositories.

Self-host in minutes with Docker Compose, or let us run the control plane.

Request early access Self-host