What changed
Waku Agent has been released as an open-source AI agent harness, emphasizing a local-first architecture and a readable codebase. The project aims to provide a clear implementation of the five key pillars of a robust AI agent: harness, loop, memory, evaluation/LLM-Ops, and an always-on layer. Unlike framework-heavy solutions, Waku Agent exposes these components directly in code, allowing for easier understanding and modification. The memory system is designed to be local, stored in a single SQLite file, and supports semantic, episodic, and procedural memory types with a gating mechanism for deciding what to remember and a pass for consolidation. The core agent loop is approximately 95 lines of Python, designed for step-through debugging. An integrated evaluation system combines deterministic tests with LLM-as-judge capabilities, including a release gate. The 'always-on' layer handles event-driven triggers such as messages, timers, and user approvals, ensuring that events are logged to disk first to prevent data loss upon crashes. The project also offers a local dashboard for visualizing agent activity and supports a wide range of LLM providers, including Anthropic, OpenAI, Gemini, and OpenRouter, with free models available via OpenRouter. A hosted version, agent.waku.one, is also available in preview.
Why it matters for builders
For AI builders, Waku Agent offers a transparent and ownership-centric approach to developing AI agents. The open-source nature and readable code allow developers to deeply understand, customize, and extend agent functionalities without the opacity of proprietary frameworks. The local-first design ensures that sensitive data, such as agent memory, remains under the user's control, which is crucial for privacy-conscious applications. Furthermore, the modular design of the five core pillars makes it easier to experiment with different memory strategies, loop logic, or evaluation methods, fostering innovation and rapid prototyping.
Practical impact
Developers can quickly get started by installing waku-agent via pip and running it from the terminal or launching a dashboard at localhost:7777. For those who prefer to modify the code or contribute, cloning the repository and setting up a development environment is recommended, with instructions provided for installing dependencies and configuring LLM providers. The project includes a quickstart guide and detailed documentation covering architecture, integrations, and evaluation methods. Builders can experiment with agent behavior by issuing commands like "Remember that Alex prefers morning meetings," and then testing its recall with subsequent commands. The ability to connect Waku's memory to other agents like Claude Code and Codex via Waku Memory plugins offers a path for shared knowledge across different AI tools. The waku serve command enables an always-on mode for continuous operation, triggered by external events.
Caveats and source limits
The source material indicates that the hosted version, agent.waku.one, is currently in a preview and invite-only state. While the core Waku Agent harness is MIT licensed, the hosted/ directory is under the Elastic License 2.0, which permits self-hosting but restricts offering it as a service to others. Specific details regarding performance benchmarks, scalability limits for the local-first architecture, or comprehensive pricing for the hosted service are not provided in the source excerpts. The "fresh release" status suggests that the project is actively under development, and certain features or integrations might still be evolving.
Sources
Claim check: 7/7 supported claims - 7 evidence links - 100% avg confidence
- Waku Agent is an open-source AI agent harness.supported - github.com
- Waku Agent features a local-first architecture with memory stored in a single SQLite file.supported - github.com
- The core agent loop in Waku Agent is approximately 95 lines of plain Python.supported - github.com
- Waku Agent includes an integrated evaluation system with deterministic tests and LLM-as-judge capabilities.supported - github.com
- Waku Agent supports multiple LLM providers including Anthropic, OpenAI, Gemini, and OpenRouter.supported - github.com
- A hosted version of Waku Agent, agent.waku.one, is available in preview.supported - github.com
- The core Waku Agent harness code is MIT licensed, while the hosted/ directory is under the Elastic License 2.0.supported - github.com
Caveats
- Single-source caution: verify critical details at the linked source.
Radar score 87/100 - how it was calculated
- Reliability 82: GitHub metadata supports source trust
- Freshness 92: Fresh GitHub activity
- Novelty 77: Fresh GitHub release
- Technical 89: Repository technical metadata
- Developer 96: Developer tooling signals
- Ecosystem 72: Fresh GitHub release
- Confidence 96: Claims have reliable evidence
Discussion
Loading comments...