Quick Start

Quick Start

The shortest path from zero to a running TinyAgents graph, plus the crate layout you will navigate as you go deeper. TinyAgents is a recursive language-model (RLM) harness for Rust — see Recursion and RLM for what that means once you are up and running.

Install From crates.io

Add the crate to your project. The default build is offline (no hosted providers, no network):

[dependencies]
tinyagents = "0.1"

Enable hosted providers

Hosted model providers live behind the openai Cargo feature. Enable it to use OpenAI and OpenAI-compatible endpoints:

[dependencies]
tinyagents = { version = "0.1", features = ["openai"] }

Then export your key (and optionally point at a different model or base URL):

export OPENAI_API_KEY=...
export OPENAI_MODEL=gpt-4.1-mini          # optional
export OPENAI_BASE_URL=https://api.openai.com/v1   # optional

The same feature powers other providers (Anthropic, Ollama, DeepSeek, Groq, xAI, OpenRouter, Together, Mistral, and compatible endpoints) through provider specs and helper constructors. See Providers.

• crates.io: https://crates.io/crates/tinyagents
• docs.rs: https://docs.rs/tinyagents

Clone, Test, And Run

To work against the source or run the bundled examples, clone the canonical repository:

git clone https://github.com/tinyhumansai/tinyagents.git
cd tinyagents
cargo test

Useful local checks (these mirror CI):

cargo fmt --check
cargo clippy --all-targets -- -D warnings
cargo build --all-targets
cargo test

Run the local graph example

The basic graph example needs no provider credentials and runs fully offline:

cargo run --example basic_graph

It threads typed Rust state through a small two-node graph, routes conditionally after the agent node, and exits when the state no longer needs the tool:

flowchart TD
    Start((START)) --> Agent[agent]
    Agent -->|needs_tool| Tool[tool]
    Tool --> Agent
    Agent -->|done| End((END))

Loading

Run an OpenAI-backed example

OpenAI-backed examples need the openai feature and OPENAI_API_KEY:

export OPENAI_API_KEY=...
cargo run --features openai --example openai_chat

Other offline examples include complex_graph, durable_graph, and agent_loop_tools; other hosted examples include openai_tools, openai_structured, openai_graph_agent, orchestrator_subagents, openai_self_blueprint, and rag_blueprint. See Examples.

Crate Layout: The Five Surfaces

TinyAgents is organized around five surfaces. Knowing where each lives makes the source easy to navigate:

• src/harness/ — Harness: provider-neutral model calls, typed tools, middleware, structured output, streaming, usage/cost, retry/limits, cache, memory/embeddings, sub-agents, steering, summarization, and testkit doubles.
• src/graph/ — Graph runtime: durable typed state graphs with START/ END, nodes, edges, conditional routing, commands, Send fanout, reducers/channels, checkpoints, interrupts, subgraphs, and topology export.
• src/registry/ — Registry: the named capability catalog (models, tools, agents, graphs, stores, middleware, policy) that .rag/.ragsh bind by name.
• src/language/ — Expressive language .rag: declarative, side-effect-free blueprints that compile (lexer → parser → compiler) into the same runtime.
• src/repl/ — REPL language .ragsh: imperative, capability-bound interactive orchestration — the RLM/CodeAct loop surface.

Supporting paths:

• examples/ — runnable examples (offline and OpenAI-backed).
• docs/spec/ — contributor-facing system specification.
• wiki/ — this GitHub wiki source.

Next Steps

• Recursion and RLM — the execution model that makes TinyAgents an RLM harness, with the concrete recursive surfaces and research lineage.
• Examples — work through the runnable examples, including the self-authoring .rag blueprint demos.
• Architecture — how the five surfaces compose end to end.