Open SWE is LangChain AI's open-source async coding agent framework — the shipping equivalent of the internal coding agents that Stripe (Minions), Ramp (Inspect), and Coinbase (Cloudbot) built for their engineering orgs. Instead of forking an existing agent or starting from scratch, Open SWE composes on top of LangGraph and the Deep Agents framework, giving teams an opinionated architecture with a clean upgrade path for upstream improvements. It targets the specific slot between "chat with an IDE plugin" and "pay per task for a hosted agent": an agent your org owns, customizes, and deploys on its own infrastructure.
The system runs every task in an isolated cloud sandbox — a remote Linux environment with full shell access where the repo is cloned and the agent has free rein inside a contained blast radius. Sandbox providers are pluggable: Modal, Daytona, Runloop, and LangSmith work out of the box, and the customization guide covers plugging in your own. Invocation happens where engineers already live — Slack threads, Linear tickets, and GitHub issues — and the agent reports back with commits and pull requests. Tool curation follows Stripe's principle that a small, focused toolset beats a kitchen sink, so Open SWE layers a tight set of project tools on top of the Deep Agents primitives (read_file, write_file, edit_file, ls, glob, grep, write_todos, and subagent spawning).
Context comes from both AGENTS.md-style project files and source-code retrieval, and orchestration mixes subagents with middleware so teams can inject validation, safety checks, or custom workflows without forking the core. Open SWE is MIT-licensed and positioned explicitly as "the open-source version of the pattern those companies built internally," which matters in 2026 as internal agent platforms become a standard build vs. buy decision. Engineering orgs that want the Stripe/Ramp/Coinbase architecture without a six-month build — and without sending source to a third-party agent service — can stand Open SWE up on their own infrastructure and customize prompts, tools, and review gates to match house style.
