GitSeed – Self-hosted control plane for AI coding agents

The open-source platform provides a complete solution for running AI coding agents with zero Git credentials exposure, featuring a three-phase container architecture (prepare → agent → finalize) that ensures untrusted agents never directly access sensitive repositories. GitSeed supports multiple AI models including Claude Code, Codex, Gemini, OpenCode, and others, with the ability to run up to three agents in parallel for 'Second Opinions' during ideation, planning, or review stages.

Security Architecture

GitSeed's security model centers on the principle that coding agents are untrusted components. The platform implements a 3-phase container split where the agent container runs with no privileges, no Docker socket access, and zero Git credentials. All repository commits, blocked path checks, and secret scanning occur in a separate finalize container that mounts workspace directories read-only.

The control plane includes Ed25519-signed PR Governance Receipts that cryptographically attest to locked specifications, models used, security scans performed, and human approvals. These receipts can be verified offline, providing an auditable trail of the AI-assisted development process.

Deployment Options

Developers can get started locally in under five minutes using Docker without requiring cloud accounts or external Git hosts. For production deployments, GitSeed can be installed on Kubernetes using the included Helm chart after pushing images to a cluster registry. The platform offers multiple connection methods including API keys for Anthropic, OpenAI, Gemini, and xAI, local Ollama/vLLM setups, or single-user Personal Mode CLI OAuth.

GitSeed provides both a 5-stage SDLC pipeline (Ideate → Plan → Execute → Review → Merge) for complex features and a ⚡ Fast-Track mode (Execute → Review → Merge) for quick fixes. The platform also includes a persistent SessionDock for live sandbox rescue, allowing developers to drop into any workspace and steer active runs without losing progress.

Model Support

The platform supports a wide range of AI coding models including proprietary options like Claude Code, Codex, Gemini, and Antigravity, as well as open-weight alternatives. GitSeed enforces spend controls through LiteLLM with hard per-run USD caps for cloud APIs or local GPU deployments. Personal Mode strictly binds subscriptions to individual users while team projects can use API keys or local models.

Each stage agent can be independently bound to different CLI runtimes and model backends, with transparent labeling of enforced versus estimated spend controls. The platform maintains runtime availability for various models and allows custom agent implementations through its modular architecture.