Catalerum

catalerum is a self-hostable, automated, fully-integrated LLM assistant, a catalogue of the things in your life.

This is my idea of a highly capable agentic workflow environment.

Goals

• fully integrated automated personal management system
• easy way to do ai native stuff and expose it via mcp servers
• ability to integrate the whole software in coding agents

Recipes

Run just to list every recipe:

Recipe	What it does
just up	Start backing services + the echo llmleaf container and wait for store health.
just dev	Boot the API (migrates + seeds + prints the magic-link).
just llmleaf	Run the llmleaf proxy container with the echo config on :8088 (standalone).
just web	Serve the web workbench with trunk serve on :8080.
just migrate	Migrations run automatically on boot; this documents that path.
just seed	Admin + workspace + magic-link are seeded automatically on boot.
just down	Stop and remove the backing services.
just reset	Wipe volumes and recreate the dev stack.
just test	cargo test across the native workspace.
just check	cargo check --workspace --exclude catalerum-web.
just check-web	cargo check -p catalerum-web --target wasm32-unknown-unknown.
just fmt / just lint	Format / clippy (native + wasm web).
just e2e / just mcp	Placeholders wired up in later milestones.

All-in-one container

For a single-node installation, the all-in-one bake target packages the frontend, API/workers, terminal runner, preview service, llmleaf, Qdrant, and same-origin routing into one image. It exposes only port 8080, uses SQLite and an in-process coordination store, and provides one-time owner setup plus user management. Neo4j is replaced by the relational graph fallback.

docker buildx bake all-in-one --load
docker run --rm -p 8080:8080 \
  -v catalerum-data:/data -v "$PWD/files:/files" -v catalerum-work:/work \
  registry.k3s.s.fionn-router.internal/catalerum/catalerum-all-in-one:dev

See docs/all-in-one.md for volumes, first boot, dynamic llmleaf providers, and the distinction from the existing distributed mode.

OpenTelemetry and Langfuse

Tracing is opt-in under [telemetry] in config/catalerum.toml. Catalerum can export application, HTTP, and GenAI spans to a standard OTLP/HTTP collector, directly to Langfuse, or to both at once. Incoming W3C traceparent headers are honored and exporters flush on graceful process exit.

The OTLP and Langfuse destinations each have an independent LLM content policy:

• metadata-only (default): model, latency, finish state, token/cache usage, cost, and errors; no prompts or generated content.
• all-except-system-prompts: full inputs/outputs with every system-role message removed.
• everything: full request, response, reasoning text, and tool payloads.

For Langfuse, enable [telemetry.langfuse] and inject CATALERUM_TELEMETRY__LANGFUSE__PUBLIC_KEY and CATALERUM_TELEMETRY__LANGFUSE__SECRET_KEY. The default endpoint is Langfuse Cloud EU; set CATALERUM_TELEMETRY__LANGFUSE__ENDPOINT for another region or a self-hosted instance. Generic OTLP is configured under [telemetry.otlp], with optional static headers for authenticated collectors.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE), or
• MIT License (LICENSE-MIT)

at your option.