Overview
Waku-agent is a local-first personal AI assistant that is implemented as a codebase users can read and modify. It is organized around four pillars: Harness, Loop, Memory, and Eval/LLM-Ops. The project includes a terminal interface and a local browser dashboard that shows messages as they flow through the harness. The agent loop is about 95 lines of plain Python, and the system supports multiple model providers, including Anthropic, OpenAI, Gemini, DeepSeek, MiniMax, Kimi, GLM, OpenRouter, OpenCode Zen, and OpenCode Go. Memory is stored locally in one SQLite file and includes semantic, episodic, and procedural memory, with a gate that decides whether to remember and consolidation that decides what to keep. Eval is built in with deterministic tests and LLM-as-judge side by side, a release gate, and a trace for every turn. Waku Memory is a hosted memory option that can be shared across agents, such as saving in Claude Code and recalling in Waku. The code is MIT licensed except for the hosted/ directory, which is Elastic License 2.0.
Pricing
Free
Detailed plans have not been confirmed in our catalog. Check the official website for current limits and billing terms.
Visit WebsitePrices and limits may change. Confirm the currency, billing period, seat minimum and usage caps on the official website.
Use Cases
- Teams use Waku to run a local-first personal assistant on their laptop.
- Teams use Waku to talk to their assistant in the terminal.
- Teams use Waku to open a browser dashboard cockpit at localhost:7777.
- Teams use Waku to store memory in one SQLite file.
- Teams use Waku to connect hosted Waku Memory across agents like Claude Code and Codex.
- Teams use Waku to use model providers they already pay for.
- Teams use Waku to run deterministic tests and LLM-as-judge evals with a release gate.
- Teams use Waku to connect gateways including terminal, dashboard, voice, Telegram, Discord, and WhatsApp.
Features
Local-first AI agent harness
Code-defined agent loop
Built-in agent memory
Agent evaluation in code
Inspectable code-based architecture
Reviews
No reviews yet. Be the first to share your experience!
FAQ
What does local-first mean in waku-agent?
Local-first means all agent execution, memory storage, and data processing happen on your own machine by default. You don't need to send your data to a third-party server, giving you full privacy and control. You can still integrate with remote AI models if you choose, but the core harness runs locally.
How does the evaluation system work?
waku-agent includes a built-in evaluation harness that lets you define test cases and metrics in code. You can run automated evaluations against your agent to measure performance, catch regressions, and compare different configurations. This helps you iterate with confidence and maintain quality as your agent evolves.
Can I use my own language model with waku-agent?
Yes. waku-agent is model-agnostic and designed to work with any LLM provider or local model. You can plug in your preferred model via a simple interface, whether it's a cloud API like Claude or a locally hosted open-source model. This flexibility ensures you're never locked into a specific vendor.
Is waku-agent really free to use?
Yes, waku-agent is completely free and open source. You can use, modify, and distribute it without any licensing fees. The project is publicly available on GitHub, and we welcome contributions from the community to improve and extend it.
How is waku-agent different from other agent frameworks?
waku-agent emphasizes legibility and ownership. Everything is defined in plain code, so there's no magic or hidden state. It's local-first, so you control your data. It also includes a native eval system, making it a complete package for serious agent development. Many frameworks focus on orchestrating cloud services; waku-agent focuses on giving you a transparent, self-contained foundation.