ACTIVE · shipping since 2024

ABOUT COGNITHOR.

A local-first agent operating system built for people who want AI on their own terms. Open-source. Offline-capable. Privacy by architecture.

▶ Read the Manifesto ▶ View the Code

▸ WHAT THIS IS

WHAT COGNITHOR IS.

Cognithor is an agent operating system. It runs entirely on your machine, uses the local LLM of your choice, and manages the full stack a serious agent needs — memory, tool execution, channel adapters, safety gates, and a persistent knowledge graph.

The project ships 145 tools across twelve MCP modules, eighteen channel adapters (Telegram, Slack, Discord, iMessage, Matrix and more), nineteen supported LLM providers (local via Ollama / LM Studio / vLLM / llama.cpp, cloud via OpenAI / Anthropic / Google / Groq / Mistral), and a six-tier memory stack that distinguishes identity, core, knowledge, episodic, entities, and tactical storage.

Every request walks through the PGE Trinity: a Planner that decomposes intent into tool calls, a Gatekeeper that classifies risk against a four-level policy (green / yellow / orange / red), and an Executor that runs approved calls in a sandbox. Yellow risks require human-in-the-loop approval. Red risks escalate to a pending-review queue.

The Evolution Engine runs in idle time. It re-indexes new content, summarizes long episodes, updates the entity graph, and surfaces meta-learnings back to the user. No cloud roundtrip. No telemetry. No subscription pricing.

It is not a chatbot. It is a person's operating system for AI.

▸ HISTORY

PROJECT TIMELINE.

2023EXPLORATION
Open-weight LLMs became genuinely useful on consumer hardware. Qwen, Llama, Mistral proved that a solo developer could host a serious model at home. The question was no longer whether, but how to build an operating system around them.
2024INCEPTION
First commits. The PGE (Planner / Gatekeeper / Executor) architecture was established. A six-layer memory stack was built. Eighteen channel adapters went online. The project was originally named Jarvis.
2025EXPANSION
145 tools across 12 MCP modules, 19 LLM providers, a Flutter Command Center, the Evolution Engine with autonomous learning and idle-loop meta-learning. The project was renamed Cognithor to make room for paid agent packs.
2026AGENT PACKS
First paid pack shipped: Reddit Lead Hunter Pro. The pack infrastructure landed — MDX schemas, signed distribution, Gumroad checkout. A 70/30 creator marketplace was announced for Q4.

▸ THE MISSION

THE MISSION.

“Make the most capable AI assistant in the world run entirely on one person's computer.”

— the project, 2024

▸ TRUST CONTRACT

PROJECT COMMITMENTS.

A public contract between the project and its users. These twelve lines are the non-negotiables — the things that would be broken first if the project ever turned into a SaaS or a venture play.

THE PROJECT WILL NEVER

✗Ship telemetry or anonymized tracking
✗Require a cloud roundtrip to run
✗Lock you into a subscription to keep using what you built
✗Train models on your vault, your chats, or your memory
✗Take venture capital that would force a vendor lock-in
✗Hide breaking changes, quietly regress features, or break backward compatibility without notice

THE PROJECT WILL ALWAYS

✓Ship the core under Apache 2.0
✓Work fully offline with a supported local LLM
✓Keep every tool call auditable and gateable
✓Let you export, back up, or delete every byte the project stores about you
✓Run every commit through the full test suite in CI
✓Document breaking changes in advance with a migration guide

▸ RECEIPTS

LIVE NUMBERS.

12,847

COMMITS

since 2024

145

MCP TOOLS

14 modules

13,247

TESTS

unit + integration

89.4%

COVERAGE

lines / branches

187k

LINES OF CODE

Python + Dart + TS

Every commit is visible. Every test runs in CI. Every release is signed.

▸ SUSTAINMENT

HOW COGNITHOR IS SUSTAINED.

Transparent revenue model. No venture capital, no ads, no telemetry data sold to third parties. The core is free; packs fund the maintainer; creators get the majority split.

CORE · FREE

The Cognithor agent OS is Apache 2.0. Clone it, fork it, ship your own distribution. You never pay to run it and never will.

PACKS · PAID

Specialized agent packs are paid one-time purchases. They fund the maintainer, the test infrastructure, and the documentation effort.

CREATORS · 70/30

Q4 2026: third-party creators will publish their own packs with a 70/30 revenue share. The project becomes sustainable without venture capital.

▸ CREDITS

MAINTAINER & COMMUNITY.

Cognithor is maintained by Alexander Söllner, an independent developer based in Germany, and contributors across the open-source community.