Coderium Agent Framework
CAF turns a ticket into a pull request with AI agents. You decide every merge.
- Ticket
- Plan
- Implement
- Validate
- Reviewer
- PR
- Merge
- Free
- Open source
- Self-hosted
From the terminal to the agent floor
Two apps: caf-initiator sets up your repo with one command, caf-orchestrator runs and shows every agent on your own server.
One command sets up your repo
caf-init detects your stack and drafts the agents, commands and skills for you to review.
- scaffold, curate, docs, export
- Deterministic, no AI calls
- Dry-run first
AI agents that work on their own, but never merge on their own
Spec-driven
The Planner writes a spec with numbered acceptance criteria first. A blocking question stops the run and is asked on the ticket, and an approved spec is never edited by an agent.
Plan → Implement → Verify
Every task runs through Plan → Implement → Verify. A failing gate retries once, then stops with a Draft PR and hands off to a human.
Every phase leaves a file in your repo
Each phase's output is stored as a Markdown file in the repo — not chat context that disappears once the session ends.
Starts when a ticket is ready
The Orchestrator is triggered automatically when a ticket becomes ready — a Linear workflow state or a GitHub Issue label. Jira support is planned.
Checkers never edit what they check
The Validator checks every acceptance criterion and the Reviewer checks the diff against the spec. They write a report and don't change the code they are checking — a failure goes back to the implementation agent.
Human decides the merge
The agent chain runs by itself up to an open pull request, and can wait for you to approve the spec first. Merging is always an explicit human decision — there is no auto-merge.
Two apps, one workflow
CAF Initiator
Automatically detects your project's stack, then generates the knowledge base your agents need — five agents, five slash commands and nine skills. Deterministic: it never calls an AI model.
- Auto-detects your stack: framework, package manager, monorepo or single repo
- Generates .claude/agents/, slash commands and the CAF skills, including the spec skills
- Leaves a TODO instead of guessing — /caf-complete-drafts helps you fill them in
# in your repo root: preview what would be written
npx -p caf-initiator caf-init scaffold --dry-run
# generate the drafts
npx -p caf-initiator caf-init scaffold
CAF Orchestrator
Fastify + BullMQ + Redis, running on your own small VPS. Receives webhooks from your tracker, then spawns headless Claude Code agents for each phase.
- Triggered automatically from Linear or GitHub Issues (Jira planned)
- Asks on the ticket when the spec has an open question, and can wait for /caf-approve-spec
- Gate failures open a Draft PR you can resume with /caf-retry-pipeline
- Self-hosted — runs on your own VPS, not a Coderium SaaS
# run: web server, worker & view dashboard
pnpm dev && pnpm dev:worker
# or run TUI (terminal UI)
pnpm tui
The agent chain runs itself — merging never does
Once a ticket is ready, the agent chain runs on its own up to an open pull request. It pauses for a person only on an open question, an optional spec approval, or a gate that keeps failing — and Merge is always a human decision.
- Agent — runs on its own
- Human — needs a person
Same way of working.
Self-install
Free and open source.
- Your team installs CAF on its own.
- Runs on your own server.
Installed by Coderium
Starts with a free audit, then a written quote.
- Free 30-60 minute audit to see whether AI actually helps your team.
- Written quote with a fixed scope.
- We install and tailor it on your server.
Monthly support
Optional, after installation.
- Your choice once installation is done.
- Limited hours per month.
Design partner program
For the first 3 clients we offer special pricing. In return, you let us write a case study from the installation.
Frequently asked questions
CAF (Coderium Agent Framework) is an AI agent orchestration framework that takes a ticket from a written spec to a pull request through Plan → Implement → Verify, with strict governance — automated quality gates along the way, and a human decision before anything is merged.