Diffinery

Diffinery

Hosts Git repositories, pull requests, code review and GitHub Actions workflows on your own Cloudflare account, with guarded accounts for AI coding agents.

Made by codician-team on GitHub. These notes are Mindket's, written from its README; the project's own files and license are in its repository.

About

What it is: Diffinery is an open-source alternative to GitHub and GitHub Actions that runs entirely on Cloudflare. It covers Git hosting, issues, pull requests, code review and a merge queue, plus releases and a container registry. It also runs your existing workflow files and gives AI coding agents accounts of their own.

Who it suits: It suits teams that want GitHub-style tooling on their own Cloudflare account, especially teams that work with AI coding agents. Teams that do not want to run on Cloudflare can skip it.

Industry | Who uses it | How they use it | Example
Software development | Teams on GitHub Actions | Keep their workflow YAML and run it on Diffinery's runners | A marketplace action running on a runner with 4 vCPUs
Engineering with AI agents | Teams using Claude Code, Codex or Cursor | Give each agent an account and have a person approve its pull requests | An agent opens a pull request instead of pushing to main
Platform engineering | Teams that want to self-host | Run their own instance without managing servers | An instance deployed to a Cloudflare account instead of a VM
Cloud development | Cloudflare developers | Study a large application built on Durable Objects, Containers, Workflows, Queues and R2 | Durable Objects backed by SQLite

Where to use it: Diffinery runs on your own Cloudflare account, and people and agents reach it from the tools they already use.

• Web app: the interface where you review pull requests and open the Today page
• Claude Code and Codex: plugins for both bring Diffinery into them
• Cursor: it can work as an agent with an account of its own
• Editors and the command line: editor extensions, gh support and a CLI

Use case | Who | What they get back
Replace GitHub Actions CI | Teams with workflow YAML | Existing workflow files run unchanged, and marketplace actions work too
Hand a goal to an agent | Teams using AI agents | A pull request from Claude Code or Codex, built in a sandbox
Agent code review | Reviewers | A second agent checks every pull request for bugs
Faster CI with code on GitHub | Teams with private repositories | GitHub runners option runs the jobs on Diffinery

How to set it up: You deploy your own instance, so you need a Cloudflare account first.

• A Cloudflare account
• A domain for your instance

1. Get a Cloudflare account and pick the domain for your instance.
2. Open the self-hosting guide in the project's docs.
3. Deploy the instance into your Cloudflare account, on the domain you chose.

How to use it: Start each morning on the Today page, which shows what needs you. Approve what is ready and let pull requests merge when checks pass. Agents reach it through the MCP server, the REST and GraphQL APIs, plain git or the CLI.

Tool | What it does
Today | Shows what needs you each morning
Agent tasks | Takes a goal and returns a pull request built by Claude Code or Codex in a sandbox
Agent code review | Has another agent review each pull request for bugs
Merge queue | Merges approved pull requests when checks pass
Built-in MCP server | Lets agents work in Diffinery through MCP
REST and GraphQL APIs | GitHub-compatible interfaces for tools and agents
Container registry | Stores container images next to your code

Example: You give an agent a goal in plain words. Claude Code or Codex works on it in a sandbox and opens a pull request. Another agent reviews it for bugs, you approve it from Today, and it merges once checks pass.

For newcomers and experts: If you are new, think of Diffinery as GitHub that you run yourself on Cloudflare. It hosts code, pull requests and CI, and your workflow files keep their syntax. A first small thing to try is giving one agent a goal and reading the pull request it opens.

Under the hood, a single Worker handles the web UI, git over HTTPS, the REST and GraphQL API and the MCP server. Every repository gets a Durable Object of its own, backed by a separate SQLite database, so each one scales on its own and none share a database server. The rest uses D1, R2, Queues, Workflows and Containers. The trade-off is Cloudflare lock-in, where Gitea, Forgejo or GitLab run on a VM or a Kubernetes cluster.

You give an agent a goal
The agent opens a pull request from a sandbox
Another agent reviews it for bugs
Your workflow files run as checks
You approve it on Today and it merges on green

Good to know:

• License: Apache License 2.0, free and open source
• Platform: it runs only on Cloudflare, with your own account and domain
• Runners: 4 vCPUs and 12 GiB by default
• Agents: they get their own accounts, with guardrails the server enforces
• Gaps: the README lists what Diffinery does not do yet, so check it before you move over

Title: vCPUs per default CI runner
Diffinery: 4 vCPUs
GitHub standard, private repos: 2 vCPUs

On GitHub, 7 November 2026

• Stars: 82
• Forks: 3
• Watching: 2
• Open issues and pull requests: 1
• Language: TypeScript
• License: Apache-2.0
• Topics: agentic-coding, ai-agents, ci-cd, cloudflare, cloudflare-containers, cloudflare-workers, code-review, continuous-integration, devops, durable-objects, git, git-server

Links

• Code: https://github.com/codician-team/diffinery
• Website: https://diffinery.com/
• Releases: https://github.com/codician-team/diffinery/releases
• Issues: https://github.com/codician-team/diffinery/issues
• Star history: https://www.star-history.com/#codician-team/diffinery&Date

Get the code

git clone https://github.com/codician-team/diffinery.git

Then follow the setup in its README. Nothing in this kit runs by itself.