Frame · Report
Director (Proteus)
Frames the task, assigns the work, keeps it on track, reports.
Not: hands-on coding.
What it is
Not a plugin. Not a fork. A layer on top.
A team of AI agents, built from one agent engine. Proteus-kit turns a single AI agent engine (Hermes, by Nous Research) into a small team under clear rules — one plans, one builds, one checks. Open source · built on Hermes (MIT) · kit under Apache-2.0.
| Hermes (engine) | Proteus-kit (layer) | |
|---|---|---|
| What it is | A single AI agent engine (Hermes, by Nous Research), open source under MIT. | A team layer on top: six roles working under clear rules. Open source under Apache-2.0, open-core version 1.6.0. |
| What you install | The Hermes engine. | The Proteus kit from github.com/aska-digital/protean-kit. |
| What you get | One agent that runs. | A team: one plans, one builds, one checks — frame, research, architect, design, build, verify, report. |
| Unit of work | An agent run. | A checked handoff between roles, each stage gated by a done-check. |
| Question it answers | How do I run an agent? | How should a team of agents actually work together? |
The crew
Six roles, each with a lane
Each role keeps to its lane. The boundary lines below are the point: the reviewer is separate from the builder, and you see what is ready and what is still uncertain.
Research
Researcher (Hazen)
Finds facts, prior work, and market info the task needs.
Not: architecture decisions.
Architect
Architect (Leo)
Turns findings into a plan for how the work will be built.
Not: writing code.
Design
Designer (Frida)
Designs what the user sees and touches.
Not: backend logic.
Build
Builder (Mozi)
Writes the code — the only role that does.
Not: strategy prose.
Verify
Reviewer (Shaka)
Checks the result against the bar; hunts bugs, cuts anything extra.
Not: feature expansion.
How it runs
Work moves through fixed stages in order
Work moves through fixed stages in order: frame, research, architect, design, build, verify, report. A done-check sits between stages. Work moves forward only when it meets the bar for that stage (Definition of Done).
Each role keeps to its lane. The reviewer is separate from the builder. You see what is ready and what is still uncertain.
How a run unfolds: Intake (the Director receives the brief and clarifies the goal) → Frame (outcome, constraints, review boundary) → Route (choose the specialist) → Knowledge (load the relevant decisions and sources) → Local helpers (Redact, Gist, Title — uncertain results held for review) → Models (local or hosted to suit the task) → Handoff (save the artifact, owner, verdict, and open questions). Fixed order throughout: lead → research → plan → design → build → check → report.
$ generate.py --settings team.yaml
frame → research → architect → design → build → verify → report
review: check the delivered result, not a claim that it worked
report: show what is ready, what is uncertain, and what comes nextReview: check the delivered result, not a claim that it worked. Report: show what is ready, what is uncertain, and what comes next.
Proteus runs on local models, hosted models, or a mix. Optional privacy helpers prepare text before it leaves your machine. Redact flags sensitive text for review. Gist suggests a topic hint. Title drafts a label.
Technical detail, including setup steps and licenses, lives in the setup guide.
What ships today
What ships today, and what does not
Available now: the open-source kit, templates, working rules, build checks, and the site generator.
Still being built: the conversational setup agent and the one-command installer.
Setup today needs comfort with a terminal and some manual steps. Copying generated files into place is manual for now.
proteus · honesty limit
The public kit has not yet been independently demonstrated to deliver a real product end to end. That line stays until the demonstration exists.
Static statement of the current limit, not a live status feed.
Why this shape
Why the kit works this way
What the kit adds that the engine alone does not. Four shipped, one still being built.
Working rules are committed as files, not kept as habits. The team runs on what is written down.
The generate.py script turns one settings file into the kit materials. If the builder cannot produce it, it does not count as kit.
A small always-on index loads only the knowledge each step needs, instead of everything up front.
Work state is written to disk as it happens, so a crashed run restarts from the last file instead of from scratch.
A conversational setup agent that interviews you and writes the settings file. Not yet shipped. What is still being built.
Operating upgrades the team runs on: supervision with a restart limit, research that stops and asks instead of running silent, the maker never marking their own work, strict independent review, and one fixed order of work.
Keep asking
Questions, answered honestly
Yes, with a compatible local model and enough memory. Small specialist models do narrow jobs; larger models do harder work. Hosted models are another option, and the mix depends on the task, the hardware, and privacy needs. The local helpers (Redact flags sensitive text, Gist suggests a topic, Title drafts a label) do not replace reasoning or human review — an uncertain redaction must be reviewed before text leaves the device.
Think of a workshop notebook. Decisions, sources, and completed work are saved in files, and a new session loads the relevant notes. A way to preserve continuity, not a promise of perfect memory.
Source and licenses
Source and licenses
The kit source is public at github.com/aska-digital/protean-kit. Open-core version 1.6.0. The kit is open source under Apache-2.0.
Pricing summary
Setup, deployment, and training: $9,300 USD, one time. Ongoing service: $540 USD per month. Hardware quoted separately.