cognithor

Cognithor abstracts 19 LLM providers (OpenAI, Anthropic, Google Gemini, Ollama, etc.) behind a unified Python API, allowing agents to switch providers at runtime without code changes. Uses a provider registry pattern with standardized request/response schemas that normalize differences in API signatures, token counting, and streaming behavior across proprietary and open-source models.

Cognithor implements a Model Context Protocol (MCP) tool registry that exposes 145 pre-built tools (web search, file operations, database queries, API calls, etc.) as callable functions within agent workflows. Uses a schema-based function registry pattern where tools are defined with JSON schemas for input validation, and agents invoke them via standardized function-calling APIs supported by OpenAI, Anthropic, and other providers.

Cognithor builds and maintains knowledge graphs that represent entities, relationships, and hierarchies extracted from documents and agent interactions. Agents can traverse knowledge graphs to reason about entity relationships, perform multi-hop reasoning, and answer questions that require understanding connections between concepts, rather than relying solely on semantic similarity.

Cognithor implements a multi-level memory architecture combining short-term context windows, episodic memory (conversation history), semantic memory (vector embeddings), knowledge graphs, and persistent vaults for long-term retention. Uses hierarchical retrieval patterns where agents query appropriate memory tiers based on query type: recent context for immediate relevance, embeddings for semantic similarity, knowledge graphs for relationship reasoning, and vaults for archival data.

Cognithor integrates agents with 18 communication channels (Discord, Telegram, Slack, email, webhooks, etc.) through a unified message routing layer that normalizes channel-specific message formats, user identities, and authentication into a standardized internal message protocol. Agents receive normalized messages regardless of source channel and can respond to any channel without channel-specific code.

Cognithor provides an Agent Packs marketplace where developers can publish, discover, and install pre-configured agent templates that bundle LLM provider selection, memory configuration, tool sets, and channel integrations. Packs are versioned, dependency-managed, and installable via a package manager pattern, allowing rapid agent deployment without manual configuration.

Cognithor is architected as a local-first system where agents run entirely on-premises with no data transmission to external telemetry services or cloud logging. Supports local LLM inference via Ollama integration, local vector databases, and local knowledge storage, enabling GDPR-compliant deployments where sensitive data never leaves the organization's infrastructure.

Cognithor provides an agent orchestration layer that enables autonomous agents to decompose complex tasks into sub-tasks, plan execution sequences, and reason about tool choices using chain-of-thought patterns. Agents can dynamically select from available tools, evaluate outcomes, and adjust strategies based on feedback without explicit human instruction for each step.

Cognithor integrates document analysis capabilities that extract structured knowledge from unstructured documents (PDFs, images, text files) using OCR for scanned documents and semantic parsing for text extraction. Extracted content is automatically indexed into the knowledge vault and made queryable through semantic search, enabling agents to reason over document collections without manual preprocessing.

Cognithor integrates with Obsidian vaults, allowing agents to read from and write to Obsidian notes, leverage Obsidian's graph structure for knowledge relationships, and maintain bidirectional sync between agent memory and Obsidian. This enables users to manage agent knowledge using Obsidian's familiar interface while agents benefit from structured, interconnected knowledge.

Cognithor is released under Apache 2.0 license with a modular, extensible architecture that allows developers to add custom LLM providers, tools, memory backends, and channels through well-defined plugin interfaces. The codebase is designed for community contribution with clear separation of concerns between core orchestration, provider adapters, and integrations.