Open SWE is an open-source asynchronous coding agent built by LangChain, officially described as an "open-source software factory built on top of Deep Agents." It turns engineering work into a repeatable system: investigate the codebase, implement changes, verify changes, and deliver a Pull Request; at the same time, it can also review PRs, learn repository-specific review preferences, monitor CI, and respond to feedback. It is open source and can be deployed on users' own infrastructure.
Purpose and Positioning Open SWE is aimed at team-based software engineering workflows, automating steps such as code investigation, modification, verification, PR submission, and code review. It supports asynchronous execution: users can initiate work from the dashboard, GitHub issue and PR conversations, and Slack; Linear supports issue comment triggering and replies through a configured Linear MCP connection. Cloud coding follow-up operations reuse the context and sandbox of the same thread, and independent threads can run in parallel.
Main Capabilities Build: investigate the repository, edit code, run targeted verification, and open or update a Pull Request. Parallel: use subagents for research and independent work. Review: support on-demand or opt-in automatic review, publish findings to GitHub, and learn from historical feedback. Investigate: use read-only PR chat to understand changes without implementing changes. Operations: schedule recurring tasks, and monitor opted-in PRs via /baby-sit, diagnose failures, and rerun only flaky tasks supported by evidence. Customization: choose models, reasoning effort, instructions, skills, integrations, and sandbox providers.
Technical Architecture Deep Agents provides planning, filesystem, shell, skills, and subagent primitives; LangGraph provides durable execution and thread state; Open SWE adds engineering tools, integrations, authorization, and a user interface on top. Graph entry points are declared in langgraph.json, and an architecture manifest is provided. Cloud coding runs in persistent, per-thread isolated Linux sandboxes, with tools provided by workspace scripts or snapshots. When the coding sandbox is unreachable, it will not be silently replaced. LangSmith is the default sandbox and tracing provider, and other providers and local execution are configurable. PR chat does not require a sandbox.
Usage Getting-started paths include:
- Team deployment: set up the backend, dashboard, GitHub and Slack apps, and model credentials. Production standalone Agent Server Deployment requires a license key.
- Local development: complete the setup of dependencies, credentials, database, hot reload, and webhook-only tunneling in order.
- Desktop (experimental): works against local repositories. Packaged app releases target macOS; source builds also support Windows and Linux.
- Using the CLI: connect a local directory to an agent on a deployment. Commands execute locally as the user, without sandbox isolation. (Specific commands and installation steps are not fully listed in the provided materials and are not provided for now.)
Control and Security GitHub access: coding sandboxes typically receive installation-level GitHub App access; some workflows use a narrower repository scope. Workspace repository binding controls routing and preloaded checkouts, but is not a separate credential boundary. Integrations: MCP connections layer instance-level, workspace-specific, and personal tools, and their scope and credentials can be configured in the customization guide. Approvals: workflow file approval prompts are used to intercept detected Git pushes, not all possible shell or API writes. Reviewers are asked not to commit or push; PR chat excludes change-making tools. Sandboxes have powerful tools and may have network access. It is recommended to use least-privilege credentials, restrict repositories and integrations, and adjust the approval policy according to the deployment. Local execution does not provide cloud sandbox isolation.
Applicable scenarios and limitations Suitable for teams that want to deploy asynchronous coding agents in their own infrastructure and automate GitHub/Slack/Linear workflows. In terms of limitations: the project is in active development and may have incompatible changes; the official project does not currently accept issues or external contributions; correctness, stability, and compatibility are not guaranteed; production standalone Agent Server deployment requires a license key; local execution has no cloud sandbox isolation; other details (such as specific pricing, supported platform details, and complete installation commands) are not provided for now in the provided materials. The project uses the MIT license.
