NullClaw

NullClaw is open-source infrastructure for running an AI assistant. The project repository describes a Zig-based runtime whose model provider, messaging channels, memory, tools, and deployment components can be changed through configuration. It is aimed at people who want to operate and customize their own assistant rather than simply open a hosted chat page.

Models, channels, and actions

The runtime supports hosted model providers, OpenAI-compatible endpoints, and Ollama. Messaging options include a command-line interface, Telegram, Discord, Slack, Signal, Matrix, and other listed channels. You configure the provider and channels the assistant will actually use; the presence of an integration in the repository does not supply an account or credentials for that service.

Built-in tools cover shell commands, file reading and editing, web requests and search, memory, scheduling, and agent delegation. The configuration also includes MCP servers using HTTP or standard-input/output transports. This makes NullClaw relevant to assistants that need to take actions and retain context across messages. The operator must decide which tools and permissions fit the intended job.

Memory and operation

SQLite is the default memory backend, with keyword and vector search described in the README. Other listed backends include Markdown, PostgreSQL, Redis, and several alternatives. The project also documents memory export, import, and maintenance commands.

You can use an interactive agent session or start the gateway, which runs configured channels along with the heartbeat and scheduler. The project’s English documentation index directs new users through installation, configuration, and everyday operation. Homebrew, container, and source-build paths are documented. Source builds require the exact Zig version 0.16.0.

Permissions and maturity

The README describes pairing, channel allowlists, workspace scoping, encrypted secrets, and selectable sandbox backends. These are configurable protections, so an operator should review the security guidance before enabling broader filesystem or network access. Running the assistant locally does not mean that a configured remote model processes data locally.

NullClaw is MIT-licensed. The project explicitly says it is pre-1.0 and gives no stability guarantees for configuration or CLI behavior between releases. It publishes small-binary and low-memory benchmark claims, but those reflect the stated build and benchmark environment. Evaluate the runtime on your intended hardware and account for any model-provider or hosting costs separately.