langchain-ai/open-swe

Open SWE – an open‑source software‑factory built on LangChain’s Deep Agents

What it is – A framework that lets you run autonomous AI agents that can plan, write, test, review and deliver code changes in your own repositories. The agents run inside isolated sandboxes, integrate with GitHub, Slack, Linear and a web dashboard, and can be customized with your own models, tools, and policies.

Core capabilities

• Build – The agent inspects a repo, creates a work plan, edits files, runs targeted validation, commits the changes and opens/updates a PR.
• Review – A read‑only reviewer agent can comment on PRs, learning the repository’s review style from past feedback.
• Operate – Trigger tasks from the dashboard, GitHub issues/PRs, Slack threads or Linear tickets; schedule recurring jobs; monitor CI and automatically rerun flaky jobs.
• Customize – Swap models, sandbox providers (LangSmith, Modal, Daytona, Runloop, E2B, or local), add skills, define repo‑specific instructions, and set delivery policies.
• Safety – Per‑thread sandbox isolation, GitHub‑App/OAuth auth, allow‑lists, credential handling, and optional human approval before pushes.

How it works

1. Deep Agents supplies the low‑level primitives (planning, file ops, shell, sub‑agents). Open SWE builds on this to add software‑engineering tools, prompts and integrations.
2. LangGraph provides durable execution and thread state. Five entry‑point graphs are shipped: Agent (full build), Reviewer, Analyzer (learn review style), Chat (PR Q&A), Scheduler (recurring/CI tasks).
3. Sandboxes run the actual code in isolated Linux environments. The default is LangSmith, but many cloud or local providers are supported via a pluggable interface.
4. Integrations – GitHub (issues, PRs, App), Slack, Linear, and a Vite‑based web dashboard let users start and monitor tasks.

Typical workflow

User → dashboard / GitHub / Slack / Linear → OpenSWE creates a sandbox thread
   → Agent plans & implements change
   → Validation runs in sandbox
   → PR is opened/updated
   → Reviewer agent (optional) comments
   → User replies → same sandbox continues work

Getting started

git clone https://github.com/langchain-ai/open-swe.git
cd open-swe
uv venv && source .venv/bin/activate
uv sync --all-extras   # install Python deps
pnpm install           # install UI deps
make dev   # backend
make web   # dashboard (http://localhost:2024)

Follow the docs/INSTALLATION.md for GitHub‑App setup, sandbox credentials and production deployment.

Project status – Actively developed by LangChain; APIs and UI are still evolving. The repo includes an Open SWE Enhancement Proposals (OE‑P) process for community‑driven changes.

License – MIT.

---

Open SWE turns software‑engineering tasks into repeatable, AI‑driven pipelines that you can host on‑premise or in the cloud, while keeping full control over models, data and execution.