The coding agent for the fast terminal.
Built for developers who value speed, privacy, and zero framework bloat. Run fully offline with local Ollama models, or stream cloud inference via Groq with sub-10ms startup.
Crafted for pure terminal speed.
Every tool is engineered with safety guards, zero framework overhead, and instant path resolution.
Smart path auto-resolution
Reads complete file contents asynchronously. When the model requests a bare filename like agent.ts, tiny-agent shallowly traverses the repository and auto-resolves to src/agent.ts without failing turns.
Single replacement safety
Exact substring replacement. Fails safely if old_string is missing or appears more than once, preventing silent codebase corruption.
Confirm before action
Read-only inspection executes instantly with dimmed status lines. State changes pause for interactive (y/n) approval with timeout guards.
Offline or instant cloud
Switch instantly between 100% offline Ollama models (Qwen 2.5 Coder, Llama 3.2) with zero telemetry, or stream cloud inference via Groq.
Model Context Protocol
Native MCP client support. Connect any standard MCP server via config.json to give your agent access to databases, GitHub, or local APIs.
Production-grade tooling foundation
Combines Bun's native bundler and runtime with private local models and fast cloud streaming.
Bun Runtime
Native TypeScript execution without compilation steps. Sub-10ms startup and fast compilation to a single binary.
Ollama Engine
Private, local-first inference running directly on your machine. Your code and prompts never leave your local environment.
Groq Cloud
Fast cloud inference via OpenAI-compatible endpoints. Instant tool streaming when you need larger model capabilities.
Interactive terminal simulator
Watch tiny-agent resolve paths, request tool approval, and stream answers without bloated abstraction layers.
What fx doesn't do yet
Both tools share a minimal Unix ethos with zero framework bloat. Here is how their capabilities differ today based on verifiable code.
Local models
Offline inference
No working local-model path today. Relies on external OpenAI-compatible cloud endpoints via curl.
Fully offline execution via native Ollama integration. Tested with local open weights like Qwen 2.5 Coder without API keys or telemetry.
Safety confirmations
Action inspection
Prompts confirmation before executing raw bash commands suggested by the model.
Granular safety policy. State changes (write_file, edit_file, run_shell_command) pause for explicit confirmation with diff previews; read-only operations run instantly. Scriptable with --yolo.
Open & extensible
Standard protocols
Script-level configuration without standard external tool protocols or structured context files.
Model Context Protocol (MCP) client to plug in standard tool servers, automatic AGENTS.md project instruction loading, and modular skills system.
How it works
Three sequential steps to an instant, local-first coding assistant in your terminal.
Interactive first-run wizard saves config to ~/.tiny-agent/config.json, or pass --local / --cloud flags anytime.
$ tiny-agent --local
[tiny-agent] provider=ollama
model=qwen2.5-coder:latestStart an interactive REPL or invoke one-shot commands directly from your shell or CI pipeline.
$ tiny-agent "edit src/config.ts to use port 8080"
[groq/openai/gpt-oss-120b] ❯Autonomous tool loop reads context, auto-resolves bare paths, prints action logs, and streams final answer.
⚡ edit_file src/config.ts
✓ Successfully edited src/config.ts
Updated the port to 8080.Installation
Install globally via Bun or npm to get the tiny-agent executable in your terminal.