Overload reviews GitHub pull requests with AI models you control: a local model on your own Mac (served on Metal by DwarfStar or llama.cpp) or any hosted OpenAI-compatible model. You decide which agents review which files, every finding is checked against the diff before it is shown, and reviews are posted back to the pull request as inline comments.
Status: pre-release. Local and webhook reviews, scheduled prompts, the GitHub App and posting reviews to real pull requests work and are tested.
| Where | Models | Guide |
|---|---|---|
| Your Mac | Local or hosted | docs/install-mac.md |
| A Linux server (DigitalOcean, AWS, …) | Hosted | docs/install-linux.md |
| Kubernetes | Hosted | docs/install-kubernetes.md |
| Kubernetes behind Cloudflare Tunnel and Access (no public IP) | Hosted | docs/cloudflare-kubernetes.md |
| Receiving webhooks on a Mac or home server | docs/cloudflare-tunnel.md |
On a Mac, the short version is:
curl -fsSL https://raw.githubusercontent.com/daltoniam/overload/main/install.sh | shThis installs overload (and Postgres, with Homebrew), starts both as login
services and prints the UI address (http://127.0.0.1:8082) and password.
1. Agents and a workflow. An agent is a model plus its instructions. A workflow has a main agent and up to eight sub-agents, each limited to the files it is for. An optional planner sends files to the right sub-agent whatever their path, and an optional verifier drops weak findings before they are posted.
2. Preview the routing. Paste a pull request's file list (or pick a real pull request) to see who would review each file, how many model calls that takes and roughly how long, before any model runs.
3. Reviews run on every pull request. Bind a repository's pull request events to the workflow. Each run records the findings, which agent reported them, what the verifier dropped and how much each agent cost.
4. Findings are posted to GitHub. With posting turned on, the GitHub App leaves the review as inline comments on the changed lines.
The dashboard tracks runs across all repositories:
overload review --repo owner/name --pr 123This uses your default model and gh auth token (or GITHUB_TOKEN), prints
the findings and saves the run so you can open it in the UI. Nothing is
posted. --workflow NAME runs a saved workflow, and --concurrency N
reviews N files at once when the model server has parallel slots.
Overload talks to any OpenAI-compatible server. On a 64 GB M1 Max, Qwen3.8 Flash Next on DwarfStar found all five bugs in our planted-bug test file in four of four runs, at 8 to 12 minutes per file. See docs/install-mac.md for local and hosted setups and deploy/README.md for measurements.
Everything in the UI is also available from the CLI as JSON
(overload agents|workflows|bindings|repositories|schedules list|apply), so
a coding agent can configure overload too. docs/workflows.md
describes every workflow setting.
Scheduled agents can also call tools on MCP servers such as Switchboard, for jobs like keeping an issue queue up to date every morning. See docs/tools.md.
Requires Go 1.26.5+ and Postgres (overload install or
docker compose up -d postgres).
export DATABASE_URL=postgres://overload:overload@localhost:5432/overload?sslmode=disable
make ci # generate, gofmt, vet, lint, test, build
make install-test # macOS: install under a test label and check it end to end
make kind-test # disposable kind cluster: app, Postgres, restartsOVERLOAD_DEV_SEED=1 overload dev-seed adds disabled example configuration
and one sample run for exploring the UI. PLAN.md holds the design
and roadmap; AGENTS.md the conventions for coding agents.