𒉺 Tablette

Tablette is a tiny AI-Factory prototype made during the IGN 2026 Territorial Data Factory Hackathon. AI-Factories are 'LLM-based applicative infrastructures '. The name (French word for tablet) is a reference to the Si.427 tablet and is a the first try of a multi-step marhackathon (a marathon of hackathons). Next step : the 2026 French National Assembly Hackathon.

Tablette seeks to imagine the next-generation digital systems capable of creating agile feedback loops between regulatory and legislative outputs on the one hand, and the reality of environmental issues on the ground on the other.

Clermont example.

Quick demo

make serve

Open http://localhost:8080, type a prompt like "make a simple python server", and click Run.

What happens:

1. The frontend sends your prompt to the sdbxr MCP server
2. The MCP server spawns opencode inside a bwrap sandbox (--share-net)
3. Opencode calls the LLM, writes code to a temp workspace
4. The frontend streams logs live as they arrive
5. When done, the preview server starts under bwrap on port 8000
6. The result appears in the iframe on the right — fully sandboxed

Architecture

Browser ──► FastAPI ──► SdbxrClient ──► sdbxr_mcp.sh ──► bwrap ──► opencode
 │ │
 └── SSE (logs) ◄───────────────────────────────────────┘
 │ │
 └── iframe preview ◄────────── bwrap ──► server:8000 ──┘

Tags

• v0.1 — Initial cartographic AI factory. Contains prototype sandbox wrappers (dagger/, jai/, kata/, sbx/, sdbxr/) explored during v0.1 development. These directories were removed in subsequent commits to keep the tree focused; see the v0.1 tag for the full archive of prior investigations.
• v0.2 — Current. Working sdbxr MCP pipeline with web UI, live log streaming, and iframe preview.

See also

IGN MCP AI projects :

• mcp-carto
• geocontext