AI SDK — Rust

The world's most comprehensive AI SDK — multi-provider, multi-agent, production-ready. Now implemented in Rust per ADR-011.

Overview

This workspace contains the AI SDK, a comprehensive Rust SDK/ADK for building AI-powered applications and agents. It provides a unified interface across multiple LLM providers, parallel execution, agent orchestration, memory, RAG, web research, workflows, streaming, observability, analytics, and security — with a real, production-oriented implementation (no mocks, no stubs, no placeholders).

• Product Requirements Document:PRD-v1.md
• Engineering Specification (adapted):ENGINEERING-SPEC.md
• Architecture Decision Records:ADRs/
• Design verification report (historical):VERIFICATION_REPORT.md

Crate Layout

Crate	Description
ai-types	Core domain types: messages, content parts, roles, usage, modalities
ai-core	Core traits: Model, Provider, Tool, Client, runtime abstractions
ai-config	Unified configuration: env vars, TOML files, programmatic
ai-errors	Typed error hierarchy
ai-models	Model registry, metadata, capabilities, routing
ai-discovery	Provider-agnostic model/capability discovery with provenance (declared · inferred · probed)
ai-providers	Real adapters: OpenAI, Anthropic, Google Gemini, OpenRouter, Ollama
ai-runtime	Parallel execution: concurrency limits, retries, circuit breaker
ai-stream	Unified streaming events, SSE parsing
ai-tools	Tool framework, built-in tools, skills registry
ai-protocols	MCP client/server, A2A client/server
ai-agents	Agent runtime, sub-agents, patterns, swarms, self-healing
ai-orchestra	Hierarchical task trees: planner/expander/clarifier, mailboxes, registry, watchdogs
ai-web	Web subsystem: crawler, extractor, search (self-hosted)
ai-memory	4-tier memory with pluggable storage
ai-rag	RAG: chunking, ingestion, retrieval, hybrid search
ai-workflows	Workflow engine: sequential/parallel/conditional, checkpoints
ai-observability	Structured logging, spans, chronological event history
ai-analytics	Metrics, cost estimation, aggregation
ai-devtools	Inspector, trace viewer, debugging, doc-truth lint
ai-security	Redaction, PII, SSRF guards, permissions
ai-cache	Caching: TTL, semantic cache interface
ai-storage	Storage backends: KV, document, vector (sqlite adapter)
ai-edge	Edge/WASM build targets, runtime detection
ai-voice	Voice: audio types, VAD, STT/TTS traits + adapters
ai-computer	Real computer control: OmniChrome browser bridge (CDP) + Native Computer Use desktop client
ai-cli	CLI: doctor, providers, models, config, run, inspect, trace, benchmark
ai-sidecar	stdio JSON-RPC gateway binary: configure providers, list/discover models, chat generate/stream
ai-sdk	Facade crate: unified public API

Getting Started

Status: Implemented across 29 Rust crates — a 580+-test offline suite (unit, proof, property, and integration tests) plus credential-gated live suites. CI runs rustfmt, clippy (-D warnings), and workspace tests; the live-gateway job runs only when its gateway secrets are configured (gated per-step on env.AI_SDK_GATEWAY_API_KEY).

# Build the workspace
cargo check --workspace
# Run tests
cargo test --workspace
# Lint (CI enforces this with -D warnings)
cargo clippy --workspace --all-targets --all-features -- -D warnings

Computer Use & Browser Automation

ai-computer drives real engines — nothing is simulated:

• Browser — OmniChromeClient / BrowserTool: OmniChrome Chrome-extension bridge (http://localhost:8765/rpc). Auth: env OMNICHROME_TOKEN, or the token file server/.bridge-token. Navigate, click (xy/selector), type, screenshot, extract Markdown/a11y-tree, evaluate JS, raw CDP.
• Desktop — NativeComputerClient / ComputerTool: Native Computer Use engine (http://localhost:8888/rpc). Auth: env COMPUTERUSE_TOKEN, or %USERPROFILE%\.computeruse\auth.token. Screenshots, OCR text-finding, Set-of-Marks UI tree, mouse/keyboard/paste, window management.

With an engine running, tools execute for real; engine-down is a typed, actionable error. Offline wire proofs run in CI; live smoke paths are credential-gated and never dial out uninvited.

Sidecar JSON-RPC Gateway

ai-sidecar exposes the SDK to non-Rust hosts over stdio NDJSON:

cargo run -p ai-sidecar -- [--config path/to/ai-sdk.toml]

Methods: initialize, configure, provider.list, model.list, model.discover, model.info, chat.generate, chat.stream, stream.cancel. Frames are capped at 16 MiB; oversized frames are rejected while the loop keeps serving.

Debugging Provider Wire Traffic

Set AI_SDK_DEBUG_WIRE=1 to write a redacted outbound-request record (length + sha256, masked credentials) to the temp dir; append =full to include complete bodies.

Quick Links

• Engineering Specification
• PRD v1.2
• Architecture Decision Records
• Changelog
• Contributing
• Security

License

MIT — see LICENSE.