langgraph-harness

<div align="center"> <img src="docs/public/logo.svg" width="72" alt="langgraph-harness logo" /> # langgraph-harness **Self-hosted AI coding agent platform for GitLab, built with LangGraph** Reviews merge requests, resolves work items and tasks autonomously, and chats with full project context — remembering what matters so every run builds on the last. Battle-tested in production since mid-2026 — **133 releases**, 1k+ reviews, 67% comment acceptance rate (Sept 29 2026, [see snapshot](#screenshots)). [Development history →](CHANGELOG.md) every failure and fix, documented. [](LICENSE) [](package.json) [](https://vrajpal-jhala.github.io/langgraph-harness/) [Docs](https://vrajpal-jhala.github.io/langgraph-harness/) · [Getting Started](https://vrajpal-jhala.github.io/langgraph-harness/guide/getting-started) · [Features](https://vrajpal-jhala.github.io/langgraph-harness/features) · [Screenshots](https://vrajpal-jhala.github.io/langgraph-harness/screenshots) · [Screencasts](https://vrajpal-jhala.github.io/langgraph-harness/screencasts) <img src="docs/public/hero.gif" width="900" alt="Demo: agent resolving an issue end-to-end" /> </div> > Not affiliated with LangChain. See [DISCLAIMER.md](DISCLAIMER.md). --- ## What it does Four harnesses, one shared foundation: - 🔍 **MR Review** — fires on webhook events, reads the diff, drafts comments, publishes them. No polling, no manual trigger. - 🛠️ **Work Item Resolve** — assign a GitLab issue or task and get an agent that runs inside its own Kata Containers VM (a dedicated guest kernel per sandbox, not just syscall interception), opens a draft MR, and keeps responding to follow-up comments on the same branch. - 📋 **Task Resolve** — the same sandboxed loop, started from a free-text instruction instead of a GitLab issue — on demand or on a recurring schedule. - 💬 **Chat** — an interactive, GitLab-aware assistant with real tool access: GitLab data, a headless browser, its own review/chat history; sensitive tool calls pause for explicit human approval before running. Shared foundation: every drafted comment is screened before it posts; a run that misbehaves (repeats a call, loops, ends on a question, skips a check) gets caught and corrected mid-run — backed by persistent checkpoints, so a run resumes instead of restarting from scratch. Full breakdown: [Features](https://vrajpal-jhala.github.io/langgraph-harness/features) · [Architecture](https://vrajpal-jhala.github.io/langgraph-harness/architecture) ## Screenshots <div align="center"> <img src="docs/public/screenshots/analytics-overview.png" width="720" alt="Real production analytics, Sept 29 2026 snapshot: 1k+ reviews, 67% comment acceptance rate" /> <img src="docs/public/screenshots/analytics-detail.png" width="720" alt="Real production analytics detail, Sept 29 2026 snapshot: reliability, efficiency, and guardrail health" /> </div> More in the [screenshots gallery](https://vrajpal-jhala.github.io/langgraph-harness/screenshots). ## Roadmap - [x] Automatic MR reviews, issue-to-draft-MR, on-demand tasks, and a GitLab-aware chat - [x] Isolated sandbox for code-changing runs, with human approval before Chat takes sensitive actions - [x] Project memory that carries across reviews, and crashed runs that resume instead of restarting - [x] Recurring scheduled tasks - [ ] GitHub support: assign an issue to the bot, get a draft PR ([#35](https://github.com/vrajpal-jhala/langgraph-harness/issues/35)) - [ ] Project memory that keeps itself up to date, with visibility into what it learned ([#2](https://github.com/vrajpal-jhala/langgraph-harness/issues/2)) - [ ] A self-hosted memory engine that links related facts and retrieves them by relevance ([#1](https://github.com/vrajpal-jhala/langgraph-harness/issues/1)) - [ ] Web search, so agents can look things up instead of only fetching a known page ([#24](https://github.com/vrajpal-jhala/langgraph-harness/issues/24)) Everything else is tracked in [open issues](https://github.com/vrajpal-jhala/langgraph-harness/issues). ## Stack - **Frontend** — React admin UI (Dashboard, Threads, Chat, Workflows) for monitoring runs and chatting directly with the agent - **Backend** — Elysia API server running a LangGraph agent with persistent checkpoints - **Agent** — Multi-provider LLM (OpenRouter, Gemini, Groq, Ollama, or sglang) with GitLab MCP tools and skill-based workflows - **Queue** — BullMQ; debounced re-reviews, capped concurrency, live queue state in the UI - **Memory** — context engineering for GitLab: durable, project-scoped facts learned across reviews; personal memory in Chat ## Quick Start **Prerequisites:** Node.js 22+, Docker, a GitLab PAT (`api` scope), a GitLab OAuth app, and an OpenRouter/Gemini/Groq key or a local Ollama/sglang instance. ```bash npm install cp backend/.env.example backend/.env # fill in GITLAB_PAT, GITLAB_OAUTH_CLIENT_ID/SECRET, SESSION_SECRET, SECRETS_ENCRYPTION_KEY, # ADMIN_GITLAB_USERNAMES, and one of OPENROUTER_API_KEY / GEMINI_API_KEY / GROQ_API_KEY / OLLAMA_BASE_URL / SGLANG_BASE_URL npm run dev ``` Frontend at `http://localhost:5173`, API at `http://localhost:3698`. Full walkthrough — GitLab OAuth app setup, webhook config, per-repo `.harness.yml`: [Getting Started](https://vrajpal-jhala.github.io/langgraph-harness/guide/getting-started). ## Deployment & CI/CD Self-hosted via Docker Compose; production deploys are automated through GitLab CI on push. See [Deployment](https://vrajpal-jhala.github.io/langgraph-harness/deployment) for server setup, and [sglang Deployment](https://vrajpal-jhala.github.io/langgraph-harness/sglang-deployment) if self-hosting the LLM backend. ## The Story Curious how this got built? [The Story So Far](https://vrajpal-jhala.github.io/langgraph-harness/journey). --- "LangGraph" is a trademark of LangChain, Inc., used here under nominative fair use to describe the framework this project is built on. This project is independent and not affiliated with LangChain, Inc. — see [DISCLAIMER.md](DISCLAIMER.md). Licensed under the [MIT License](LICENSE).