Roteiro is a Rust workspace (MSRV 1.96). Install the lean default build from crates.io, or build from source with the feature tiers you want.
# Lean default build — pure Rust, no network call of its own,
# and `roteiro model pull` included (--locked uses the published lockfile)
cargo install roteiro --locked
# Pin a specific release
cargo install roteiro@1.1.0 --locked
# Pick the capabilities you want
cargo install roteiro --features "inference,mcp" --locked
# Everything on — MCP, serving, local models, PDF/OCR/vision.
# Needs a C/C++ toolchain (the inference-local-models/serve/image-vision features build llama.cpp).
cargo install roteiro --all-features --locked
# …or build from source
git clone https://github.com/OffeneDatenmodellierung/Roteiro
cd Roteiro
cargo build --release --all-features| Feature | Adds |
|---|---|
| (default) | Graph build, query, check, render, the roteiro model registry, and roteiro security ingest|list|prefetch|status|run — smallest binary, no C++/cmake toolchain, and no network call unless you consent to a pull or a prefetch --allow-download. |
inference | roteiro infer / duplicates with a pure-Rust hashing embedder. Still fully offline; no download. |
models (default) | The roteiro model registry: list and consent-gated pull of local models. On by default — pull is the prerequisite for working offline, so a stock install has it. |
inference-local-models | Run pulled GGUF embedding & generative models via the shared llama.cpp engine. |
pdf-text · image-ocr · image-vision | Ingest text from PDFs, OCR text from images, and describe images with a local vision model. |
exec-subprocess (default) | Runs an analyzer (semgrep, osv-scanner, cargo audit) as a child process on this host. You install the analyzer; Roteiro never does. security run is sandboxed by default, so a build without exec-boxlite refuses and tells you how to get one; --allow-unsandboxed is how you choose this host instead, and it records isolation=none. Use --no-default-features --features execution for a build that provisions and ingests but cannot execute. |
exec-boxlite | The same analyzers inside a digest-pinned OCI image in a microVM — read-only worktree, no egress. Needs protoc and a provisioned runtime; see the README. |
serve | An OpenAI-compatible /v1 model endpoint over your installed models (llama.cpp). |
remote | The one feature that can send your repository's content off this machine — roteiro remote status|dry-run|call|log against a hosted model, plus spec draft --allow-remote and serve --allow-remote (Ask), which take the same gate (ADR-0019). Off by default and staying off. Compiling it does not enable it: a run needs your ~/.roteiro/config.toml and the invocation, and a committed roteiro.toml may deny it but never grant it. Read the note on the home page before turning it on. |
mcp | The MCP graph server, exposing the graph to AI agents (stdio or HTTP). |
--all-features | Every capability above at once — the largest build. Requires a C/C++ toolchain because the serving and vision features compile llama.cpp. This includes remote, so such a build can call a hosted model once you grant it in your user config and per run; it still cannot without both. |