Start here · 2 of 3
Getting started
Everything on this page runs offline, sends no request to a provider and reads no credential you already have.
Getting started
Prerequisites
| Tool | Version | Needed for |
|---|---|---|
| Rust | 1.98, edition 2024 | The workspace |
| Task | v3 | task check and task rust |
| ESS (GitHub) | 0.56.0 in CI | The specification and conformance stages of task check |
| AEP (GitHub) | 0.68.0 in CI | The planning stage of task check |
More columns: swipe horizontally, or focus the table and use the arrow keys.
Rust alone is enough to build, test and run the examples.
Get the source
git clone https://github.com/beyond10x/llm.git
cd llm
cargo test --workspace --locked
The workspace forbids unsafe, and Clippy runs all plus pedantic at deny.
Run a model turn with no network and no credential
cargo run --locked -p llm-core --example embedded
An embedded model turn. (EndTurn; usage None)
The example implements Model in-process and calls it through the same port a real client
implements. usage None is not a mistake: the example reports no counters, so none are shown.
Run a local turn walks through it.
Force one tool call
cargo run --locked -p llm-docs --example forced_tool_call
{"component":"importer","severity":"high"}
call_tool forced one tool on a recorded model and returned the call's arguments.
Call a model with one forced tool shows the program
and how to point it at a Codex login instead.
Explain a route without resolving a secret
cargo run --locked -p llm-routing --example explain
It reads examples/catalog.toml
and prints the selected target, each candidate's provenance and capabilities, and why each other
candidate was rejected. Explain a route shows the output.
Price recorded usage
cargo run --locked -p llm-cost --example quote
It prices fixture observations against a fictional price book and prints an llm.cost/2 report
whose metered estimate is 0.00022 USD. Price recorded usage
explains the report.
Depend on a crate from your own project
Nothing is published to crates.io; every package is publish = false. Depend on a crate from Git
at a release tag, as Loom does:
[dependencies]
llm-core = { git = "https://github.com/beyond10x/llm", tag = "<release>" }
llm-responses = { git = "https://github.com/beyond10x/llm", tag = "<release>" }
A package name is its directory name; its library name uses underscores, so the code says use llm_core::….
Crates lists every package, its library name and its features. The source
is licensed under Apache-2.0 from 0.1.2 on; see
LICENSE.
Next
- Call a local endpoint: send a real turn to a vLLM-compatible server declared in TOML.
- Run the checks: what
task checkproves, and what it does not.