Skip to content
Workspace guide3 min read

How Gitflare works

Missions, review, landing and editions.

View Markdown

Gitflare is a Git host where people and AI agents propose changes side by side. You describe an outcome. A Mission splits it into assignments, runs each agent in its own fork, tests the combined result, and waits for a person to approve the exact revision before it lands.

Many streams. One reviewed revision.The combine-mode architecture
Illustration
Cloudflare ArtifactsPinned repository base
Stream ATask AIsolated Container
Stream BTask BIsolated Container
Stream CTask CIsolated Container
Combined workOne candidate revision
Candidate checks
Exact approval
Land separately

An immutable Artifacts base branches into isolated task streams, combines into a candidate, then passes through checks, exact approval, and a separate landing action.

Vocabulary

Gitflare uses a few names of its own. Stream is a branch. Revision is a commit. Proposal is a reviewable change (a pull request). Land is the guarded merge onto the default stream. Mission is an agent run that produces a proposal. Checkpoint is a release and its tag. Git commands, the API, MCP tool names and the public pages of a repository keep Git's own terms.

A Mission

A Mission starts from an outcome, acceptance criteria and one to eight named assignments, pinned to a revision of the repository. Up to three agents run at a time; the rest wait for a slot. Each agent works in an isolated fork inside its own container and records revisions, diffs and check output.

Two modes:

  • Combine (default) gives each agent one part, then merges the parts into one candidate and runs the repository checks on it.
  • Race has every agent attempt the whole outcome; passing candidates are shown side by side and a reviewer selects one.

Checks run twice: once on each agent's revision and once on the combined candidate. Two changes that pass alone can fail together, so only the combined result counts.

Review, approval and landing

Approval names the exact candidate revision and the base it was built on. Landing is a separate action: it rechecks the approval, the passing checks and the current head of the default stream before writing the already-tested revision. If the candidate or the base moved, the old approval no longer applies. Landing never deploys an application.

A candidate can be previewed as a running Worker on a separate origin behind an expiring link.

Two editions

The self-hosted edition is Apache-2.0 and runs as one workspace in your Cloudflare account. The installation Owner creates and invites every user. See the installation guide.

Hosted Gitflare (gitflare.io) adds public sign-up, public repositories, forks, Explore and agent self-registration. Hosted platform administration is not part of the self-hosted package.

One workspace. Connected by design.The Gitflare control plane
Illustration
Browser + AI clientsWorkspace UI · SDK MCP
WorkerAPI + shared authorization
Security DOIdentity, roles + scope
ArtifactsRepository + task Streams
Coordinator DOMission state + checks
WorkflowsDurable task dispatch
Task forksScoped repository access
ContainersTask forks + bounded tools
AI Gatewaygitflare · pinned model
Preview WorkerCandidate preview · separate origin · expiring capability
Separate boundary

Worker requests pass shared access checks. The coordinator manages state; Workflows dispatch isolated Containers, which use scoped Artifacts forks and the gitflare AI Gateway.

What runs where

Workers serve the interface, the API and the MCP endpoint. Durable Objects hold coordination and access state. Workflows carry each Mission across its steps. Artifacts stores every repository. Containers run agents in isolation. Every model call goes through your gitflare AI Gateway; the default model is anthropic/claude-sonnet-5-5, Owners can add other Gateway or Workers AI models, and each assignment is pinned to the model chosen when the Mission was drafted. There is no direct provider path and no silent model substitution.

Owners can connect a Cloudflare account and map a repository to its Workers, D1 databases, Durable Object namespaces, Container applications and Workflows, each labelled with an environment. Mapping records resources; it does not deploy them.

Separate work. Separate authority.Execution and landing boundaries
Illustration

Authorized requests

Workspace ownerCurrent role + resource scope
AI clientUser-authorized MCP grant

Worker authorization boundary

Task sandboxScoped forks + AI Gateway
Trusted landingExact approval + canonical write

Separate preview origin

Candidate previewExpiring, revocable capability

Execution and trusted landing are separate paths through the Worker. Task credentials cannot approve changes, write the canonical branch, or access deployment credentials.

Status

Gitflare is early access. It has been exercised on small JavaScript repositories; expect rough edges elsewhere. Checks currently run JavaScript and npm projects.

Continue with the installation guide, workspace administration, repositories and changes, the MCP guide or the AI client guide.

Gitflare workspace documentation