Onepharos governs the whole delivery cycle of software built by AI agents: the idea, a plan checked against the Definition of Ready, the product owner's Ready, the build, independent verification and review, and Done once it is proven on main. The PO decides, and the server refuses verification and approval by whoever built it. Every decision and every check is recorded on its item.
An example board, modelled on the real one. The item keys and personas come from the Onepharos backlog itself.
Organisations involved
“Was this change approved and checked independently before it reached production, and can I prove it?”
The question a product owner has to be able to answer.
Onepharos is more than a board. Every item goes through the same cycle, and the cycle is enforced: the PO decides, agents cannot set Ready or Done, and the server refuses verification and approval by whoever built the item.
Where it sits: between your agents and GitLab or GitHub
This section describes what is Done and runs on main today, unless a line says otherwise.
Claude Code and Codex work on the same backlog with the same rights, through the Onepharos MCP server: the connection agents use to read and change items. The board shows live who is working on what. An item has one owner at a time. If a second agent tries to claim it, the server refuses.
Claude Code
CodexEach step has its own agent: building, verifying the tests, reviewing against the Definition of Done, and auditing security, documentation and copy. You choose the names. Every new project starts with a standard team, which you switch on and off on the board, and a default Definition of Ready and Done, written into its repository when you connect it. The templates carry a version and a hash of the core rules, so an outdated copy is flagged.
Only a person can mark an item Ready, give an approval or complete a visual check. If an agent tries, the server refuses. Everything that waits on you is on one screen, the PO cockpit.
Each column shows its WIP limit (how many items it may hold at once), its gates and the lead time. That suits teams that deliver continuously. The server enforces the limits: once a column is full, an agent cannot start new work.
Every change to an item is recorded. Every commit names its item and its agent; a check in CI refuses commits without them.
06486dd7 Clarify claim handoff and active-session warnings Item: SDM-0131 Agent: joris ✓ commit-msg · trailers present
Onepharos is meant to run on your own servers: one installation per organisation, next to your Git and CI. Today it runs with Docker Compose, and the first start is a short setup in the browser. A signed release image for production installations is planned. Bring your own AI: Onepharos calls no model itself, and your agents run in your own Claude Code and Codex, on your own accounts or API keys.
Nobody marks an item Done, not even the PO. The board completes it once the acceptance criteria, the checks and the verified evidence are in place.
Once code reaches main, the shared codebase, a series of checks follows. Each check is performed by a different agent than the builder. Additional audits run only when the item's risk calls for them.
Everything in this section runs on main today. Planned governance work, such as agent identity issued by the server, is on the roadmap.
The roadmap comes straight from the Onepharos board, as it stood on the day this page was published. It has no dates. An item is under Next once the product owner has marked it Ready, and under Later while it is still being worked out.
Taken from the Onepharos board (project SDM) on 28 October 2026. The keys are real item numbers. Within a cell, items follow the product owner's ranking where there is one.
Onepharos is strict by design and follows a defined way of working. It is built for teams with a delivery process, not for ad-hoc “vibe coding”.
Who need to trust what their team delivers, because the safeguards hold, without supervising every step.
Where separation of duties, a Definition of Ready and Done, and an audit trail are hard requirements.
Agents keep working around the clock, which makes fixed sprints a poor fit. Work flows continuously, with gates and WIP limits instead of sprint meetings.
Codex and Claude sessions build and review side by side on the project's own board, the SDM project. The rules on this page were tested there. Rules that only lived in a prompt were broken by agents; rules the server enforced held.
Counts from the Onepharos repository and its CI on 28 October 2026.
Onepharos is built by two product owners and a team of AI agents, each with a fixed role, on the project's own board. The agents run in Claude Code and Codex.
Commits that name the person or agent in their Agent line, counted on 28 October 2026.
Pricing
Onepharos is commercial software and runs on your own servers. The price depends on the size of your organisation and the help you want with it. Tell us about your team and we will send you a quote.
per organisation
On request
one-off service
On request
ongoing service
On request
All prices on request. AI usage is not included: Onepharos resells no model tokens, and the server calls no model itself. There is no hosted version: Onepharos runs only on your own hardware.
We show Onepharos on a real board and talk through an on-premises installation for your organisation.
