langgraph-harnessAI coding agent platform for GitLab

Self-hosted, 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.

Get Started

View on GitHub

See it in action

Assign an issue. Get a merge request.

50s

Get started

Clone it, configure it, run it

1. Install2. Configure3. Run

bash

git clone https://github.com/vrajpal-jhala/langgraph-harness.git
cd langgraph-harness
npm install

bash

cp backend/.env.example backend/.env
# add a GitLab PAT, a GitLab OAuth app, and an LLM backend key

bash

npm run dev

Full setup guide →

Live analytics

Every run, watched — not just trusted

A React admin UI shows live runs, queue state, and per-project analytics: reliability, guardrail health, comment acceptance rate, and more. Nothing here is a black box you find out about after the fact.

See the full gallery →

Real production analytics dashboard, Sept 2026 snapshot

Workflows

Four ways to run

MR Review

Webhook

Fires on webhook events — reads the diff, drafts comments, and publishes them, no polling or manual trigger.

Work Item Resolve

Work Item

Assign an issue and the agent implements the fix in a sandboxed checkout, opens a draft MR, and keeps responding to follow-up comments.

Task Resolve

Scheduled

The same sandboxed loop, started from a free-text instruction instead of a GitLab issue — on demand or on a recurring schedule.

Chat

Interactive

A GitLab-aware assistant with full GitLab access, live web fetch, image uploads, and semantic search across past reviews and project history.

Trust

Kept honest

Guardrails & Self-Correction

A run that repeats a tool call, ends on an unfinished-looking question, or skips checking for new feedback gets caught and corrected mid-run, not after the fact.

Quality Control Before Publish

Every drafted comment is screened for low-value noise, and its diff position is verified before and after posting — GitLab's own publish response isn't trusted blindly.

Sub-agents

MR Review can spawn an isolated verifier for a single file's change, so its own investigation doesn't fill the parent run's context — findings come back reported, not just trusted.

Operations

Memory and self-hosting

Durable Memory

The agent flags durable, project-specific facts as it works — conventions, recurring false positives, team decisions — seeding every future run of the same project.

Self-Hostable

A single Docker Compose stack. Bring your own GitLab instance, your own LLM backend (OpenRouter, Gemini, Groq, Ollama, or sglang), and your own data.