Local Cognitive GitHub
100% open source

Build AI systems
that stay yours.

A local-first workspace for designing, running, and supervising agent pipelines. Connect any model. Shape behavior with visual state machines. Keep the runtime under your control.

Explore on GitHub See how it works

macOSWindows desktop builds coming soon

local-cognitive / workflow
Runtime ready
WORKFLOW Research & publish
82%
ENTRYNew task
AGENTResearchQwen 3.5 · Local
AGENTChallengeClaude · Endpoint
HUMAN REVIEWApprove resultGuard: score ≥ 0.8
TERMINALPublish
parallel
needs revision

A SYSTEM, NOT A WRAPPER

Your agents need more than a prompt.

Local Cognitive gives them state, tools, roles, transitions, and boundaries — then makes the entire process visible to you.

01

LOCAL BY DEFAULT

Your machine is the runtime.

Sessions, task state, workflow definitions, and long-term memory stay on your device. Connect LM Studio or Ollama and keep the whole loop local.

02

MODEL AGNOSTIC

Local, cloud,
or both.

Assign the right provider to every agent without rebuilding your pipeline.

03

MCP NATIVE

Use it from the tools you already work in.

$ npm run mcp:stdio
✓ local_ai_chat
✓ local_ai_code
✓ local_ai_runtime_status

VISUAL ORCHESTRATION

Design behavior,
not prompt chains.

FSMs bring a game-development idea to agent work: explicit states and guarded transitions. Every step can run a model, call a tool, wait for a person, retry, branch, or stop.

  • 01 Persist state after every node
  • 02 Route by outcome and guard conditions
  • 03 Put human review exactly where it matters
LOGIC DRIVERDefault task workflow
v1 · valid
ENTRYTodo
AGENTExecutelocal model
AGENTCritiquecloud endpoint
REVIEWVerifyhuman gate
ENDDone
score < 0.8 approved
Autosaved locally

ONE RUNTIME, ANY MODEL

Let every agent use its best mind.

Mix private local inference with proprietary endpoints in the same workflow. The orchestration layer stays yours.

◎Ollamalocal
LMLM Studiolocal
AIOpenAIendpoint
ANAnthropicendpoint
GGeminiendpoint
+Compatible APIsbring your own
01

Support

Build the strongest version of an idea using a fast local model.

qwen3.5:9b LOCAL
02

Attack

Find weak assumptions with a different provider and context window.

claude-sonnet CLOUD
03

Judge

Compare both sides, produce a decision, then route the next state.

gemma3:4b LOCAL

ONE COHERENT WORKSPACE

From conversation to execution.

A

Chat workspace

General chat, code mode, and structured hypothesis debates in persistent sessions.

B

Task board

Create, prioritize, and follow work from Todo through In Progress to Done.

C

Tools & memory

Connect filesystem and Notion tools, Telegram access, and local long-term memory.

D

MCP access

Expose your local runtime to Codex, OpenCode, and other MCP-compatible clients.

OPEN FROM CORE TO INTERFACE

Inspect it. Fork it.
Make it think your way.

No opaque orchestration layer. No model lock-in. The runtime, workflows, plugins, and interface are all yours to change.

View source ↗ Read the quick start →
quick start
# clone the open-source runtime
$ git clone https://github.com/Ilyaberdar/local-cognitive-AI-system.git
$ cd local-cognitive-AI-system
$ npm install
$ npm run dev

✓ dashboard running at 127.0.0.1:3000

LOCAL COGNITIVE

Your agents.
Your models. Your machine.

Build the system you actually want to run.

Start building on GitHub ↗