All explorations
Exploration
FREN

AI agents · Local-first · RAG

Archipel

Personal local-first multi-agent environment, with specialized agents and human-validated actions.

Status
Personal project in active development
Stack
Python · FastAPI · Pydantic · Uvicorn · SQLite · React · TypeScript · Vite · Playwright · Ollama · ComfyUI · SearXNG · Docker / Compose

01

Context / problem explored

A single general-purpose assistant quickly becomes hard to control once it has to mix very different tasks: development, research, writing, sport, design, or document analysis. Archipel explores a different approach.

  • spreading responsibilities across specialized agents
  • keeping stable identities and domains
  • controlling which data each agent can receive
  • distinguishing conversation, action, and external effect
  • requiring human validation before sensitive operations
  • producing observable evidence rather than letting the model declare on its own that an action succeeded

02

Local architecture

The application core is built in Python around FastAPI. The frontend uses React, TypeScript, and Vite. Conversational inference relies on Ollama locally, image generation can be delegated to a local ComfyUI instance, and an SQLite database holds the persistent data the system needs. The environment can also use a local SearXNG search engine for certain explicitly triggered public searches. The goal isn’t the total absence of network access: some features can query external public services when an action requires it, but those outputs stay explicit, bounded, and controlled.

03

Specialized agents

Archipel currently relies on several specialized roles.

  • Janus — the system’s orchestrator; routes to the right specialist and keeps a clear separation between coordination and domain expertise
  • Vulcain — specialized in development; inspects an authorized folder, prepares limited changes, and works with validation steps before writing
  • Diane — specialized in verification; runs supported test suites and produces a verdict from actually observed results
  • Vénus — specialized in design and visual creation; uses ComfyUI locally to generate images without sending prompts to a cloud AI service
  • Mars — specialized in sport, particularly running and trail running; works only from information that is actually available
  • Minerve — specialized in writing, rewording, translation, and internationalization; also checks sensitive structures like variables, URLs, code blocks, and structured formats
  • Fidès — specialized in cautious legal analysis; separates available facts, open questions, risks, and possible next steps
  • Pluton — one of Archipel’s persistent specialists, with its own domain in the multi-agent architecture

04

Actions and human validation

A central principle of Archipel is never to confuse a model’s response with an action that was actually executed. Sensitive operations are governed by contracts and policies: for example, a code change can be prepared before being applied, and external or irreversible effects must stay explicit and subject to validation when the risk level calls for it. Important technical results, such as test verdicts or certain proof checks, are built from actual program observations rather than freely generated by the model.

05

Privacy and data boundaries

Archipel follows a local-first approach: conversations and persistent data stay mostly local, Ollama handles local inference, and ComfyUI can handle local image generation. Information isn’t automatically shared across all agents — the system aims to define boundaries based on context and specialty. Some network integrations can exist, but they must be explicit, limited to their actual need, and triggered voluntarily.

06

Reliability

The project includes a substantial test suite, complemented by its own type checks, tests, and Playwright flows on the frontend.

  • routing and capability contracts
  • actions and code changes
  • risk and external-effect policies
  • data boundaries
  • memory and context
  • search
  • specialized agents
  • interface and multi-agent flows

07

Current status and limits

Archipel isn’t presented as a finished product. The project is currently evolving significantly, particularly around its visual system and how agents show up in the interface.

  • the visual experience isn’t stabilized yet
  • some capabilities are still being iterated on
  • external integrations depend on third-party services and their configuration
  • security and context-sharing decisions keep being refined
  • the system primarily targets personal local use rather than a public SaaS deployment

Back

All explorations