TinyClaw

Tiny footprint. Maximum capability. 100% Rust.
Ultra-efficient AI assistant with streaming TUI, parallel tools, and tiered builds.

~3.5MB binary · <10ms startup · 2,235 tests · 22+ providers · streaming TUI · parallel tools

---

What is TinyClaw?

TinyClaw is a high-performance AI assistant built in Rust, forked from ZeroClaw with significant new capabilities:

Feature	ZeroClaw	TinyClaw
Interface	CLI only	CLI + ratatui TUI with markdown rendering
Streaming	None	Real-time SSE token streaming
Token Tracking	Stubbed	Live cost tracking across all providers
Tool Execution	Sequential	Parallel via tokio::spawn
Sessions	Lost on exit	Auto-saved to disk, resumable
TUI Commands	None	/help, /cost, /clear, /model, /sessions, /export
Build Tiers	Monolithic	3-tier feature flags (tiny/standard/full)

Quick Start

git clone https://github.com/suislanchez/tinyclaw.git
cd tinyclaw
cargo build --release

# Setup
tinyclaw onboard --api-key sk-... --provider openrouter

# Launch TUI (recommended)
tinyclaw tui

# Or single message
tinyclaw agent -m "Hello!"

# Interactive CLI
tinyclaw agent

Build Tiers

Choose the right build for your hardware:

# Tiny (~3.5MB) — CLI agent only, minimal deps
cargo build --release --no-default-features --features tiny

# Standard (~3.7MB) — adds TUI with ratatui
cargo build --release --no-default-features --features standard

# Full (~4.6MB) — everything: gateway, daemon, channels, OTel, skillforge
cargo build --release  # default

Tier	Size	Includes
tiny	~3.5MB	CLI agent, providers, tools, memory, security
standard	~3.7MB	+ ratatui TUI with streaming & markdown
full	~4.6MB	+ gateway, daemon, channels, OTel, skillforge, tunnel

TUI Features

The TUI provides a rich terminal interface:

• Real-time streaming — tokens appear as they're generated via SSE
• Markdown rendering — bold, italic, code blocks, headings, lists
• Live cost tracking — token count, request count, estimated USD in status bar
• Session persistence — conversations auto-save and can be resumed
• Slash commands:

Command	Description
/help	Show available commands
/cost	Detailed token usage breakdown
/clear	Clear history (keeps system prompt)
/model	Show current model
/sessions	List saved sessions
/session	Show current session ID
/export	Export conversation as markdown
/quit	Exit

Architecture

Every subsystem is a trait — swap implementations with a config change.

Subsystem	Trait	Ships with
AI Models	Provider	22+ providers (OpenRouter, Anthropic, OpenAI, Ollama, Groq, Mistral, xAI, DeepSeek, etc.)
Channels	Channel	CLI, Telegram, Discord, Slack, iMessage, Matrix, WhatsApp, Email
Memory	Memory	SQLite (hybrid FTS5 + vector search), Markdown
Tools	Tool	shell, file_read, file_write, memory (store/recall/forget), browser, composio
Observability	Observer	Noop, Log, OpenTelemetry
Runtime	RuntimeAdapter	Native (Mac/Linux/Pi)
Security	SecurityPolicy	Pairing, sandbox, allowlists, rate limits, encrypted secrets

Token Tracking

All providers (OpenRouter, OpenAI, Anthropic, Compatible) report token usage to a shared UsageTracker with atomic counters. The TUI status bar shows live metrics:

[1,247 tokens, 3 reqs, ~$0.0142]

Parallel Tool Execution

When the LLM requests multiple tools in one response, TinyClaw executes them concurrently via tokio::spawn instead of sequentially. Results are collected in order.

Streaming

OpenRouter supports real-time SSE streaming. The Provider trait includes chat_with_history_stream with a default fallback to non-streaming. The ReliableProvider wrapper tries streaming providers first and falls back automatically.

Configuration

Config: ~/.tinyclaw/config.toml (created by onboard)

api_key = "sk-..."
default_provider = "openrouter"
default_model = "anthropic/claude-sonnet-4-20250514"
default_temperature = 0.7

[memory]
backend = "sqlite"
auto_save = true

[autonomy]
level = "supervised"
workspace_only = true
allowed_commands = ["git", "npm", "cargo", "ls", "cat", "grep"]

[runtime]
kind = "native"

[browser]
enabled = false
allowed_domains = ["docs.rs"]

Supported Providers

OpenRouter, Anthropic, OpenAI, Ollama, Gemini, Venice, Groq, Mistral, xAI/Grok, DeepSeek, Together AI, Fireworks AI, Perplexity, Cohere, GitHub Copilot, Moonshot, MiniMax, Bedrock, Cloudflare AI, Vercel AI, and any OpenAI-compatible endpoint via custom:https://your-api.com.

Commands

Command	Description
tinyclaw tui	Launch TUI interface
tinyclaw agent -m "..."	Single message mode
tinyclaw agent	Interactive CLI mode
tinyclaw onboard	Setup wizard
tinyclaw status	System status + build tier
tinyclaw gateway	Start webhook server
tinyclaw daemon	Autonomous runtime
tinyclaw doctor	System diagnostics

Development

cargo build              # Dev build
cargo build --release    # Release build
cargo test               # 2,235 tests
cargo clippy             # Lint
cargo fmt                # Format

License

MIT — see LICENSE

Credits

Forked from ZeroClaw by @theonlyhennygod.

---

TinyClaw — Tiny footprint. Maximum capability. Deploy anywhere.