Telemachus — a private AI lab your team actually hosts

We opened Telemachus — named for the son of Odysseus — a team-oriented, self-hosted, privacy-first platform for hosting AI tools and applications. Local-first, driven primarily by llama-cpp / llama-server, with minimal Python exposure, and open from day one. It is built for teams of 1 to 1000 who want the capability without shipping their documents to somebody else’s datacenter.

What it does out of the box

The generic application set is chat, research, and document translation:

• Streaming chat against a local model, admitted through a per-team concurrency governor and metered against AI quotas.
• Agentic tool use — the agent plans, calls a tool, reads the result, and answers. The same registry serves built-in tools, plugins, external MCP servers, and sandboxed out-of-process tools with declared scopes.
• Translation — live EN↔JA with a team glossary for terminology consistency. The same engine translates whole locale catalogs, which is how the product’s own Japanese UI stays current.
• Teams and RBAC-5 — owner / admin / member / viewer under a team, with an instance operator tier above them. Every API call is authorized through one service; tokens are capped by issuer-permissions ∩ scopes.
• Localized top to bottom — one toggle re-renders the entire UI in Japanese via ICU MessageFormat catalogs, with a fallback chain and source-hash staleness detection.

A platform, not just an app

Beyond the stock apps, Telemachus is an SDK and platform. Deployers extend it with plugins — new tools and integrations, datasets, and localizations — through a stable, documented contract. The first prototype is written in Racket, but frontend and backend are deliberately swappable: the durable surface is the SDK contract, the backend APIs, the security model, and the protocols, not any one implementation.

Since launch it has grown a document repository with an S3-compatible front door — presigned links, object versions, indexed and searchable content — plus PostgreSQL parity alongside SQLite, and federated compute, the pluggable seam for handing work to remote or HPC executors.

Screenshots that can’t drift

Every image in the README is produced by Playwright tours that drive a real server against a throwaway database. CI runs the same tours on every push, so nothing in the documentation is a mockup and nothing quietly stops working.

Source, design docs, and the full feature tour on github.com/IoTone/Telemachus.