werbos

Developer documentation

Developer Docs

How to install, train, nudge, rate, see, and extend werbos — from a fresh clone to a benchmarked, persona-flavored agent.

Start here

Install

Clone, venv, backend, werbos init — a running agent from a fresh box.

Train

How to train

The integrated loop: ingest → split → baseline → train → score → promote.

Extend

Skills

The self-modifying skill registry, autonomy modes, and the auto-curator.

Use

Chat surfaces

Every flag, every slash, private mode, missions, the budget gate.

werbos is a local-first AI agent system. The docs below cover the full loop — from a fresh clone to a fine-tuned, benchmarked, persona- flavored agent that can see screenshots and self-curate its own skill library.

What composes a session

base model        +  adapter stack       +  persona blend         +  user.md          +  skills            +  memory             +  vision
(ollama / mlx)       (~/werbos/             (character + voice +     (~/werbos/          (~/werbos/           (~/werbos/             (--vision-model;
                      adapters/)            mirror slider)            user.md)            skills/ +            memory/)               /see capture
                                            (~/werbos/personas/)                          ~/werbos/skills.db)                        on macOS)

Per-turn the runtime composes the system prompt as: persona → project_context → user.md → skills → history → user message. Each piece is independently swappable; nothing here is a hardcoded constant.

Doc surfaces

The order below is roughly the order you’d touch them on a new machine.

Setup

Voice + knowledge

  • Personas — the character / voice / Tom / named-persona ladder + the mirror slider + user.md hydration
  • Skills — SQLite-backed self-modifying skill registry, 3-way autonomy mode, auto-curator
  • Vision — /see slash + vision-model wiring
  • Chat surfaces — every flag, every slash, private mode, missions, budget gate, schedule

Training + evaluation

  • How to train — the integrated loop: ingest → split → baseline noise → train → score → promote
  • Data pipeline — DATA.1 deep dive: readers, extractors, secret filters, schema
  • Eval suite — SE.9 deep dive: deterministic split, token-F1 scoring, noise floor + 2.5σ promotion math
  • How to nudge — chat-archive labeling + the dogfood loop
  • Rate UI — per-message labels in the SPA
  • Games — the games surface as a corpus pipeline

Background work

  • Background tasks — task store + runner + budget gate + /schedule
  • Autopilot — werbos running its own skills: predict → act → observe → error → correct, a learned routing policy (champion/challenger gated), self-reflection, and a fail-closed leash (allowlist + revert)

Reach

  • Social bridges — Telegram, Slack, WhatsApp via Twilio. One SocialAdapter Protocol; three working platform bridges. All existing werbos surfaces work over any platform unchanged.