About Personaxis
We built the spec
AI agents were missing.
A small open standard that says who your agent is. It travels with you, survives a model update, and reads like a document an auditor would accept.
The problem
Most teams ship agents without saying who they are.
System prompts describe what an agent does: its tasks, tools, instructions. They almost never describe who it is. The result is character drift. Agents that start as one thing and quietly become another over a long conversation, a model bump, or a regulatory review.
In practice that looks like:
- 01An agent that opens the session sharp and ends it agreeable.
- 02A model update on Tuesday changes the personality you shipped on Monday.
- 03Compliance asks what values your agent holds; you have nothing to hand them.
- 04Three teammates writing the same persona, none aware of the others.
What we built
The spec, and the tooling around it.
The spec is two files. You write personaxis.md across ten layers in canonical order: identity, character, personality, values & drives, affect, cognition, memory, metacognition, reflexive self-regulation, and how it presents itself. It compiles to PERSONA.md, the document your agent reads. Write it once, compile to any tool.
Personaxis is the tooling around the spec. The `@personaxis/persona.md` CLI is open source today; the hosted platform is on the waitlist.
Open standard
MIT licensed. Fork it, extend it, take it with you. The spec belongs to whoever writes it.
Why it matters →Hide →
Versioned in git like any other infrastructure file. No proprietary serialization, no runtime lock-in.
Measurable
Because the spec is structured, a persona can be tested under pressure and scored, instead of hoped to work.
Why it matters →Hide →
The ten layers give an adversarial test suite something concrete to score the agent against, layer by layer.
Auditable
When a regulator asks what your agent does and why, you hand them a file. Not a deck.
Why it matters →Hide →
Signed manifests, version history, and EU AI Act Annex IV compliance reports.
Who it's for
Founders, developers, and regulated teams.
Knowledge workers
Founders, freelancers, creators
You're tired of AI tools that forget who you are between sessions. You want an agent that knows your work and acts on your behalf without being asked every time.
Developers and engineering teams
Anyone shipping agents into production
You need behavior you can verify, reproduce, and roll back. Your agent has to hold its character across ten thousand conversations, not only the first ten.
Regulated industries
Finance, healthcare, legal, public sector
Auditors will ask what values your agent holds and expect a document. The August 2026 EU AI Act deadline does not move. PERSONA.md is the document.
Define who your agent is.
The catalog is open. Sign up only when you want to publish.