openfabriq

Roadmap

The current version, what comes next, and the directions we are considering.

openfabriq is at version 0.1.0, its first public release.

In 0.1.0

  • The format: factory.yaml, agents, automations, skills and scripts in an openfabriq/ folder; validate, plan with the security changes first, and apply.
  • openfabriq init: Claude Code writes the folder with you, or copies one of the templates.
  • The engine on your computer, or in Docker on a server, with a Claude subscription or an Anthropic API key.
  • Where work starts: GitHub, and Slack, Linear, Sentry and signed webhooks through a relay you deploy.
  • The internal review: reviewed_by names an agent that reviews the change before the pull request exists, at most review_rounds times.
  • The studio: overview and costs, board, runs, agents, workflow, wiki and a chat that only reads.
  • Sign-in with GitHub to the studio and the connector on a server: whoever can write to the factory's repositories enters with their own GitHub account.
  • openfabriq as a connector of Claude: read-only tools for Claude Code, Claude Desktop and claude.ai, served at https://<domain>/mcp by an engine on a server.
  • Your team's GitHub App instead of a token, in trial: openfabriq connect github creates it, and the engine renews a token per repository by itself.
  • The wiki: a map of your code the engine keeps up to date, proposed as a pull request.
  • Network rules: agents and the engine's commands reach only the sites of network, the package registries and GitHub by default.

Next

Planned:

  • Codex: run the agents with OpenAI's Codex, so a team can use a ChatGPT plan.
  • Hooks: Claude Code's hooks, declared in the factory, to run one of its scripts at a given moment of an agent's work, such as formatting the code after each edit. They come from what you applied and show in plan.
  • Accounts outside GitHub: studio accounts, with an invitation list, for people who are not on GitHub. Everyone on your repositories already signs in with their own GitHub account.

Later

Directions we are considering, in no particular order:

  • More agent runtimes and models: other command-line agents besides Claude Code and Codex, with their own models.
  • GitLab: factories for repositories hosted on GitLab, besides GitHub.
  • Microsoft Teams: start work and get the answers in Teams, as in Slack.
  • Secrets bound to a site: a real key that agents and commands never see. They get a stand-in, and the sandbox's proxy puts the real value into outgoing requests, to the sites you name only. Until then, secrets hold test values. See Security.
  • Each agent in its own container: a stronger wall than the sandbox, where an agent no longer sees the machine at all.
  • OpenTelemetry export: runs, durations and costs sent to your own observability tools. Nothing goes to openfabriq.
  • Open benchmarks and evals: reference factories, such as the site migrations, measured version after version (cost and time per ticket, work merged alone, review rounds, a person's interventions), and replayable cases for each kind of agent. You add your own runs if you choose, as a summary of numbers, never your code or prompts.

Have a say

Need one of these, or something missing here? Open an issue with your use case: what teams ask for shapes this page.

On this page