Git hosting for AI and ML teams
Version model weights with Git LFS. Commit checkpoints directly from training pipelines via the Python SDK. Let AI agents create branches, commit code, and open pull requests through the MCP server. Unlimited bandwidth on every plan.
Version control designed for ML workflows
From training script to production branch
ML teams need to version large binary files, automate commits from non-interactive environments, and integrate AI agents into their workflow. GitForge provides the APIs and SDKs to do all three.
Direct Commit API
Training pipelines commit model checkpoints without a git binary. The SDK builds a commit object (tree + blobs), sends it over HTTP, and the server stores objects in S3 and updates refs atomically. Python, TypeScript, and Go SDKs all support this.
MCP server for AI agents
AI coding agents (Claude, Cursor, and others) create branches, commit code, and open pull requests through the Model Context Protocol. No shell access or git binary required. The agent operates through structured tool calls.
Git LFS for model weights
Store .pt, .onnx, .safetensors, .h5, and checkpoint files in Git LFS. Models are stored in S3-compatible object storage. Unlimited bandwidth means pulling a 5 GB model for inference or evaluation costs nothing extra.
Webhooks for pipeline triggers
Listen for push, branch, and pull request events. Trigger retraining when data changes, run evaluation when a model is pushed, or deploy when a PR merges to production. HMAC signing and replay protection built in.
CLI and SDK automation
The @gitforge/cli handles repo creation, branch management, and sync in CI scripts. The Python SDK integrates directly into training code. Automatic pagination, typed responses, and error handling across all three SDKs.
An automated ML workflow on GitForge
At the end of a training run, the script uses the Direct Commit API to push the model file to a branch. No git binary installed on the training machine. The SDK handles tree construction, blob upload, and ref updates over HTTP.
A push event fires a webhook to your evaluation service. The payload includes the commit SHA, branch name, and changed file paths. Your service pulls the model and runs benchmarks.
If the model improves on the target metric, an AI agent uses the MCP server to create a pull request from the experiment branch to production. The PR description includes evaluation results and comparison with the current model.
Branch protection requires human approval before merging. The team reviews the metrics, inspects the training configuration, and merges when satisfied. The merge triggers a deployment webhook.
Three ways to commit
Use the SDK from training scripts, the CLI from CI pipelines, or the MCP server from AI agents. All three produce standard Git commits stored in the same repository.
Teams building with AI
ML teams versioning model weights
Store .pt, .onnx, .safetensors, and checkpoint files alongside training code. Git LFS handles the large binaries. Unlimited bandwidth means pulling models for inference or evaluation is free.
AI startups building agentic workflows
Let AI agents create branches, commit generated code, and open pull requests through the MCP server. Human-in-the-loop review via standard PR workflow.
Research teams tracking experiments
Each experiment gets a branch. Training scripts commit results via the SDK. Compare experiments through pull request diffs. Webhooks trigger automated evaluation.
Platform teams building ML infrastructure
Use the TypeScript or Go SDK to build model registries, deployment pipelines, and monitoring systems on top of GitForge. Every operation is an API call.
Frequently Asked Questions
Can Git handle large model files?+
Yes. Git LFS stores large files in S3-compatible object storage while Git tracks lightweight pointer files. On managed plans, storage is included. On BYO storage (Custom plan), there are no file size limits — connect your own S3, R2, or MinIO bucket.
How do AI agents interact with GitForge?+
GitForge provides an MCP (Model Context Protocol) server that exposes tools for branch creation, file commits, pull request management, and more. AI agents call these tools through structured requests — no shell access or git binary needed on the agent side.
Can we automate model versioning from training scripts?+
Yes. The Python SDK provides a Direct Commit API. Your training script imports the SDK, creates a commit with the model file attached, and pushes it to a branch — all over HTTP. No git binary, no working directory, no checkout required.
How does this compare to DVC or MLflow?+
GitForge versions files in Git using the standard Git protocol and Git LFS specification. DVC and MLflow are experiment tracking and metadata management tools. They address different layers of the ML workflow and can be used alongside GitForge — DVC for experiment metadata, GitForge for the actual file storage and version control.
Can webhooks trigger retraining pipelines?+
Yes. Configure a webhook to fire on push events for specific branches or paths. When a new dataset version is pushed, your webhook endpoint receives a signed payload with the commit details. Use this to trigger retraining, evaluation, or deployment pipelines.
Is there a Python SDK?+
Yes. The gitforge Python SDK is available on PyPI. It provides typed clients for repos, commits, branches, pull requests, webhooks, and all other API resources. Builder patterns for commit creation, automatic pagination, and structured error handling are included.
Ready to version your models and automate your ML workflow?
Create a free account, install the Python SDK, and commit your first model checkpoint. MCP server, webhooks, and unlimited bandwidth included.