Diffinery

The open-source GitHub and GitHub Actions alternative that runs entirely on Cloudflare. Git hosting, pull requests, code review and drop-in GitHub Actions CI, with agent tasks, agent code review and plugins for Claude Code and Codex, built for teams of people and AI agents. Deploy it to your own Cloudflare account: no servers, no runner machines, nothing to patch.

Website · Docs · For AI agents · Plugins · Deploy your own · Compare · Performance · Roadmap · llms.txt

Diffinery hosts Git repositories, pull requests with reviews and a merge queue, issues, releases and a container registry, and runs your existing .github/workflows/*.yml files as they are, marketplace actions included. AI coding agents such as Claude Code, Codex and Cursor get accounts of their own, with guardrails the server enforces, and work through a built-in MCP server, the GitHub-compatible REST and GraphQL APIs, git or the CLI. Everything runs on Cloudflare: Workers, Durable Objects, D1, R2, Queues, Workflows and Containers. It is free, open-source software under the Apache License 2.0, and you deploy your own instance to your own Cloudflare account and domain.

Diffinery is built for the way software gets written with agents. Give an agent a goal and get a pull request back from Claude Code or Codex in a sandbox. Have another agent review every pull request for bugs. Open Today each morning to see what needs you, approve, and let pull requests merge on green. Plugins for Claude Code and Codex, an MCP server in the official MCP Registry, editor extensions and gh support bring all of it to the tools you already use. What agents can do on Diffinery ↓

• Source: diffinery.com/codician/diffinery, where development, CI and deploys happen, mirrored to github.com/codician-team/diffinery.
• Documentation: diffinery.com/docs, also as Markdown and in llms.txt and llms-full.txt for AI assistants.

If Diffinery is useful to you, a star on GitHub helps other people find it.

Contents

• Who Diffinery is for
• Why Diffinery
• Screenshots
• Built for AI coding agents
• Plugins and integrations
• Features
• How it compares
• Self-host on your own Cloudflare account
• How it runs on Cloudflare
• Develop locally
• Repository layout
• What Diffinery doesn't do (yet)
• FAQ
• Contributing, security and license

Who Diffinery is for

• Teams looking for a GitHub Actions alternative that keeps their workflow YAML: Diffinery reads .github/workflows/ with the same syntax and runs marketplace actions, on runners with 4 vCPUs and 12 GiB by default.
• Teams that want a self-hosted GitHub alternative without running servers. Like Gitea, Forgejo or GitLab, you run your own instance, but it deploys to Cloudflare's serverless platform instead of a VM or a Kubernetes cluster.
• Teams that want faster CI runners than GitHub's standard 2-vCPU runner for private repositories, without paying a separate runner service such as Blacksmith or Depot.
• Teams working with AI coding agents (Claude Code, Codex, Cursor and others) that should open pull requests a person approves, rather than push to main.
• Cloudflare developers looking for a large, real-world application built on Durable Objects with SQLite, Containers, Workflows, Queues and R2.

Why Diffinery

• Git hosting, pull requests and CI in one serverless app. One Cloudflare Worker serves the web UI, git over HTTPS, a GitHub-compatible REST and GraphQL API, and an MCP server. Each repository is its own Durable Object with its own SQLite database, so repositories scale out instead of sharing one database server.
• Your GitHub Actions workflows run unchanged. The same triggers, expressions, matrices, needs, outputs, reusable workflows, environments, secrets, variables, OIDC, service containers and Docker. Marketplace actions (uses: owner/repo@ref) are downloaded and run. Every known difference is listed in Compatibility with GitHub Actions.
• Runners that exist only while a job runs. Each job gets a fresh Ubuntu 24.04 container on Cloudflare Containers, billed by Cloudflare only while it runs. The default size has 4 vCPUs and 12 GiB of memory, and jobs that use little of it move to smaller sizes on their own. Self-hosted runners cover macOS, Windows, Arm and GPUs.
• Built for AI agents, with a human in the loop. Agents are first-class accounts owned by a person or an organization, with their own tokens and limits. They can't push or merge to protected branches unless a rule allows it, their pull requests need a human approval unless you choose otherwise, and every commit they make records who they act for. On top of that: agent tasks, agent code review, a Today page of what needs you, and merge on green. See Built for AI coding agents.
• Plugins for the tools agents already use. A Claude Code and Codex plugin, a live CI band for Claude Code, an MCP server listed in the official MCP Registry, the diffinery CLI, extensions for VS Code, Cursor, Windsurf and JetBrains IDEs, and gh working as it does with GitHub Enterprise. See Plugins and integrations.
• Secure by default. Passkey sign-in with no passwords, invite-only accounts, branch protection, push protection that blocks leaked credentials, envelope-encrypted secrets, kill switches and an append-only audit log.
• Moving from GitHub is one step. A server-side importer brings a repository or a whole organization over with its history, issues, pull requests, reviews and releases, and gh, Octokit, VS Code and JetBrains IDEs work against Diffinery.

Screenshots

Git hosting and CI. The agent features have their own screenshots in Built for AI coding agents, just below.

| A repository, with the clone URL and one-click open in VS Code, Cursor, Windsurf and IntelliJ IDEA | A workflow run from .github/workflows/ci.yml, with a matrix and needs | | --- | --- | | | |

Built for AI coding agents

Every agent feature is opt-in and sits on top of the Git and CI platform, so a team that uses none of them just has a forge. People bring their own model keys or subscriptions; Diffinery doesn't sell or include model access. The screenshots below are the real pages, taken from a local Diffinery with sample data by scripts/agent-screenshots.mjs, which runs a real agent task, agent review and CI run.

Agent accounts with guardrails

An agent is an account of its own, owned by a person or an organization, with its own tokens. Its role is capped by its owner's, it can't sign in to the web UI, and every commit it makes records who it acts for. The server enforces the same rules over Git, the REST API, MCP and sandboxes:

| Guardrail | What it means | | --- | --- | | No pushes to protected branches | Agents open pull requests instead, unless a branch protection rule allows it. | | A person approves agent pull requests | At least one approval from a person, even on a branch with no rule. An agent's approval never counts. | | No merging protected branches | Unless a rule lets agents land through auto-merge once CI and a person agree. | | Agents can't mark their own CI green | An agent's check results never satisfy required checks. | | No secrets for agent-written code | A workflow run of code an agent pushed gets only its job token, so an agent can't push a workflow that sends your secrets elsewhere. | | No changes to secrets, environments or branch protection | Whatever the agent's role. Push protection can't be bypassed either. |

See Agents and Guardrails.

Agent tasks: a goal in, a pull request out

Give one of your agents a goal. Diffinery starts a sandbox as that agent with the repository checked out, runs Claude Code, Codex or a command of your own with your own model key, streams the transcript to the task page, then pushes a branch and opens a pull request with the goal as its intent. Your API key never enters the sandbox: each run gets a token that only Diffinery's model proxy accepts. Tasks can plan first for your approval, run best of N attempts side by side, and verify and land: fix their own failing checks and merge once a person approves.

To use your Claude or ChatGPT subscription instead of an API key, open a terminal: a shell in your browser with the repository checked out, where you sign in to Claude Code or Codex with your own plan and can hand it tasks.

See Agent tasks and Terminals.

Agent review on every pull request

One of your agents reviews pull requests for bugs, security issues and logic errors, the way Cursor's Bugbot does. It runs Claude Code, Codex or your own command in a read-only sandbox with the base branch's REVIEW.md rules, posts inline findings with suggest