The open-source workbench for AI coding.
Capture project work as durable Todos, carry it out in one or more inspectable Agent Sessions, and review the evidence before you mark it done. Self-host ArchCode on your machine or server and connect the model providers you choose.
Ideas, ready work, active work, and completed work stay visible in one Project.
- Capture a Todo — record something you want to build, fix, investigate, or improve in an existing Project.
- Shape it when useful — open a dedicated Discussion to clarify the work, and keep an optional Markdown Plan with the Todo.
- Run the work — start one or more Lead Sessions, or create an Automation to start it once later or on a recurring schedule.
- Stay involved — follow the work, add instructions, answer questions, and approve sensitive actions from the Web workbench.
- Review and finish — inspect changes, tool output, tests, and Session history, then decide when the Todo is Done.
Use Run now to create a Todo and start its first Session in one step, without a Discussion or Plan. You can also start an ordinary Session without a Todo.
curl -fsSL https://github.com/boh5/archcode/releases/latest/download/install.sh | shThe installer downloads the correct macOS or Linux binary, verifies its
checksum, and installs it without sudo. Installer-managed copies can later
update from Settings → About & Updates or archcode update. If
~/.local/bin is not on PATH, follow the exact instruction printed by the
installer.
archcodeOpen http://localhost:4096. On first run, the browser setup guides you through connecting a model and optionally protecting the workbench with a password.
Choose an existing project directory. Capture project work as a Todo or start a Session directly. ArchCode works with the files and Git repository already on the machine where it runs.
Need another platform, manual verification, or a remote deployment? See the installation and deployment guides.
A Todo gives each feature, bug, refactor, experiment, or investigation a durable place in its Project. Keep the request, lifecycle, acceptance criteria, references, and optional Markdown Plan together as the work moves from Idea to Ready, In Progress, and Done.
A Todo detail keeps the request, lifecycle, acceptance criteria, references, and related work together.
Register multiple existing workspaces and return to them from the same Web workbench. Start multiple Sessions from a Todo when the work needs separate investigation, implementation, or review without mixing everything into one conversation.
Each Session keeps its own conversation, selected model, working directory, tool activity, approvals, execution state, and history. Follow active work, add instructions, queue the next message, answer questions, approve sensitive actions, or stop an Execution.
The Lead coordinates an Analyst and a Build Agent while each Agent's work remains individually inspectable.
Before accepting the result, inspect file changes, tool output, test results, review summaries, and the complete Session history.
| Agent | Responsibility | Model Profile |
|---|---|---|
| Lead | Owns the outcome, works directly, and coordinates other Agents | principal |
| Discussion | Shapes one Todo and its optional Plan without implementing it | principal |
| Analyst | Deep analysis, planning support, and review | deep |
| Build | Implementation and verification | deep or fast |
| Explore | Fast codebase investigation | fast |
| Librarian | Documentation and external research | fast |
The Lead delegates bounded responsibilities when specialized work is useful. Each Agent identity has its own tools, delegation rules, and authority.
Run ArchCode on a laptop, workstation, Mac mini, home server, or VPS. The same server runtime and Web UI work in every deployment, and active work runs on the ArchCode host rather than in the browser. Closing the page does not cancel an Execution while the process and machine remain running.
ArchCode runs its own Agent loop rather than remotely controlling an existing Claude Code, Codex, or another coding CLI process. Its server owns Agent execution, projects, Sessions, tools, approvals, memory, and durable state. The browser is the control interface, not the place where the Agent runs.
ArchCode separates Agent responsibility from model choice. Connect official AI
SDK providers, custom OpenAI-compatible endpoints, or Responses-compatible
endpoints. Configure principal, deep, and fast Profiles, then use stronger
models where judgment matters and fast or local models for exploration and
routine work. See provider and model configuration.
ArchCode does not claim to make the underlying model smarter. It gives models a self-hosted runtime, specialized responsibilities, and a persistent project workbench.
- Structured file, shell, Git, search, LSP, Web, memory, and MCP tools
- Built-in and project workflow Skill packages for Todo shaping, planning, review, and repeatable working methods
- Automations that start or resume work once or on a recurring schedule
- Optional Goals that keep one Lead focused across multiple Executions and human checkpoints, with a required final review
- Inspectable Markdown memory and context compaction
- Optional Git worktree execution
- GitHub plus live HTTP/STDIO MCP integrations
- Signed direct updates with an idle-only graceful restart
Learn the product vocabulary in workbench concepts, and see GitHub and MCP integrations for the complete integration schema.
- ArchCode is currently designed for a single private operator, not a multi-user team deployment.
- Remote deployments require authentication and HTTPS or a trusted reverse proxy.
- Git worktrees provide Git-level separation, not an operating-system sandbox.
- Registered workspace files stay on the machine running ArchCode, but external model providers receive the context required for model calls.
- ArchCode can continue work while a browser is closed as long as its process and machine remain running. An active Execution does not survive a server process restart in the current release.
- macOS binaries are not currently signed or notarized. Windows is supported through WSL2; native Windows execution is not yet available.
Read the full security and trust boundaries before exposing ArchCode outside a trusted machine or network.
- Installation and platform support
- Local and remote deployment
- Provider and model configuration
- Workbench concepts
- GitHub and MCP integrations
- Security and trust boundaries
- Security vulnerability reporting
- Architecture
- Contributing
- Changelog
Contributions are welcome. Development setup, architecture notes, testing conventions, and pull request guidance live in CONTRIBUTING.md.
ArchCode is available under the MIT License.