Duet is an AI workspace where agents have a computer, a browser, persistent memory, and shared files. They can research, write, work in connected apps, generate images and video, and publish the tools they build. You can work with them in the web app or connect your own agents through the CLI, REST API, and MCP.
This repository contains the platform's source: the agent runtime and CLI, web app, backend, workspace VM tooling, and public website. It is released under the MIT License.
- Agents that keep their context. Instructions, skills, memory, and files persist between sessions and can be shared across a workspace.
- Work that runs beyond a chat. Agents use a browser, run commands, connect to external tools, and work on schedules or in response to events.
- One interface to models and services. A shared gateway connects text, image, video, and voice models alongside research and data providers.
- A workspace for people and agents. Shared files, conversations, hosted apps, and access controls live together rather than in separate tools.
- Bun 1.3.14
- Node.js 22.20 or newer
- Docker for the container-based test and visual workflows
git clone https://github.com/aomni-com/open-duet.git
cd open-duet
bun installFrom here, choose the part of the system you want to work on:
Each workspace's README explains its boundaries; its package.json owns the
available commands. You can read and modify individual components without
deploying the whole platform.
The hosted Duet service and this source release are separate. Installing the dependencies does not provision a working cloud deployment: you will need your own Convex project, authentication configuration, provider credentials, and workspace VM infrastructure. Features such as billing and email also need their corresponding services.
Start with the backend,
API, and VM image docs.
The source retains hosted-service defaults, domains, and operational scripts;
configure them for your infrastructure before using deployment or release
commands. In particular, the backend's setup shortcut selects the original
Convex project, so configure your own project with the Convex CLI instead.
All .env* files and private deployment credentials are excluded from this
repository. Supply your own configuration; keep environment files out of source
control.
Service defaults live in each package's config.ts, with environment overrides.
The development command loads .env.local and .env.staging; deployment tooling
uses dotenvx-encrypted .env.staging / .env.production
files with private keys in .env.keys. The
secret topology describes which credentials each
service consumes, and the secrets editor owns
the local editing UI.
This is a TypeScript monorepo using Bun workspaces and Turborepo. The web app consumes the public API; Convex owns metadata and policy; the workspace gateway owns access to files and running processes on each VM. The agent runtime is shared by local CLI sessions and managed workspaces.
- api — REST API and runtime release delivery
- web — product UI and browser authentication
- mcp — hosted MCP interface
- model-gateway — model-provider proxy
- edge — public apps and published-file routing
- lander — marketing, blog, and campaign pages
- docs — developer documentation site
- agent — agent execution, memory, and protocol
- duet — CLI, terminal UI, and distribution
- backend — Convex state, identity, and policy
- contract — shared API shapes and capabilities
- gateway — workspace VM runtime and file access
- vm-image — workspace operating-system image
- exe — exe.dev provisioning and infrastructure probes
- sync-tunnel — file-sync transport
- importers — conversation import formats
- service-catalog — provider routes and pricing
- ui, tailwind-config, and typescript-config — shared UI and build configuration
- web-locale — locale and browser analytics policy
- email — transactional email rendering
- seed-drive, official-agents-seed, and skills-registry — starter content and agent skills
- test-harness and fake-oauth-issuer — integration fixtures
- syncthing-darwin-arm64 and syncthing-darwin-x64 — macOS file-sync binaries
The security model explains the platform's trust boundaries. Specs hold design rationale and implementation plans; positioning explains the product's intended audience.
bun run dev # start configured development services
bun run build # build the workspaces
bun run check-types # check TypeScript
bun run lint # lint source
bun run format # format sourcePrefer the owning workspace's commands while iterating. For example,
bun run --cwd apps/lander check-types checks the public site without checking
every service. Root commands and workspace scripts are defined in
package.json.
Run the smallest relevant test first, then broaden coverage as needed.
bun run test is the full repository gate: some suites start real local Convex
deployments, and others use Docker, so it is much more expensive than a focused
check. Live staging journeys require separately configured infrastructure and
credentials. The test harness explains the
integration and staging workflows.
Visual changes have a Docker-based screenshot gate and component playgrounds; see the UI playground workflow. Contributor and coding-agent conventions live in AGENTS.md.
Each application deploys separately. The API build is the sole owner of backend publication and shared runtime delivery; the other applications build their own artifacts. This ordering prevents competing writers and keeps the API compatible with the backend during promotion. The deployment rail owns that orchestration, with backend details in the backend README.
Vercel's function builder uses a different TypeScript environment from ordinary checks. The function apps' explicit compiler settings and streaming exports are deployment requirements; their own configuration files explain the constraints. Environment fanout is a separate operation, and Vercel environment changes take effect on the next deployment.
The Duet skill teaches coding agents how to install and use the CLI. To install it for Codex or Claude Code:
npx --yes skills@1.7.0 add https://github.com/aomni-com/open-duet/tree/main/skills/duet \
--skill duet --agent codex claude-code --global --yesSelect the agents you use; omit --global for a project-local installation.
The skill's installation reference
owns runtime requirements, CLI installation, and login. Its default connection
is the hosted Duet service; see the CLI docs when
configuring another deployment.
Issues and pull requests are welcome. For a substantial change, describe the problem and proposed approach in an issue first. Keep changes focused, follow the owning workspace's conventions, and include the checks relevant to the behavior you changed. Do not include credentials or environment files.
MIT. Third-party components retain their accompanying licenses and notices, including Syncthing's MPL-2.0 license.