cognithor: Run a Local Autonomous AI Agent

Overview

Cognithor is a local-first AI agent system that plans tasks and carries them out through tools such as file operations, shell commands, web research, browser automation, and desktop control. Its Planner, Gatekeeper, and Executor architecture separates task planning from policy checks and action execution.

It is aimed at people who want a self-hosted assistant with persistent memory and multiple ways to interact, rather than a cloud-only chatbot. Local models can run without API keys, while cloud LLM providers are optional. The project describes itself as beta and experimental, so it is better suited to hands-on evaluation than deployments requiring stable production behavior.

Key Features

• Connects to local and cloud LLM providers, including Ollama, LM Studio, OpenAI, Anthropic, and Gemini.
• Supports terminal, web, REST, and messaging interfaces such as Telegram, Discord, and Slack.
• Uses a six-tier memory system with hybrid search across text, vectors, a knowledge graph, and document structure.
• Routes planned actions through a deterministic Gatekeeper with risk levels and approval checks before execution.
• Provides MCP tools for filesystem, shell, web, browser, documents, media, memory, and an Obsidian-compatible knowledge vault.
• Includes workflow execution, multi-agent features, and signed audit receipts for tracking agent activity.
• Offers encryption and privacy controls, with local operation available and cloud providers optional.

Use Cases

• A privacy-conscious individual can run a personal assistant locally for research, note organization, and recurring tasks.
• A developer can prototype tool-using agents and workflows across local and hosted models from one Python system.
• A team can connect an assistant to messaging channels for internal automation, while evaluating its beta channel integrations.
• A knowledge worker can search personal notes and documents using persistent memory and the Obsidian-compatible vault.

Getting Started

Requires Python 3.12 or later. For a basic local install:

pip install cognithor
ollama pull qwen3:8b
cognithor

A cloud LLM provider can be configured instead of Ollama. See the README for setup options, model configuration, and channel integrations.

Alternatives

• taOS: taOS centers on a web desktop for agents, local models, and optional distributed compute, while Cognithor is a Python assistant operating system with messaging integrations.
• OpenHuman: OpenHuman is a Rust agent harness spanning desktop, browser, terminal, and library interfaces, while Cognithor is a Python operating system for a locally run assistant.
• QwenPaw: QwenPaw focuses on a personal assistant connected to chat apps and usable locally or in the cloud, while Cognithor is a Python platform for configurable assistant automation.

Considerations

• The project identifies itself as pre-v1.0 beta and experimental. APIs may change, and the README does not recommend it for production customer-facing deployments.
• Local use requires a compatible model runtime and enough hardware for the selected model. Cloud providers are optional, but require their own credentials.
• Security controls and tests are described by the project, but its security is not independently audited. Test the system and its permissions in your own environment before relying on it.
• The README notes that tests do not cover real-world deployment conditions, long-running stability, multi-user load, or actual LLM response quality.