$mindfork
An AI chat with a memory and a self-model
The assistant keeps notes about you, a model of itself that it revisits, and a knowledge base over your files. Run a local Gemma or Qwen through llama.cpp, Ollama or LM Studio — or bring OpenAI, Anthropic, Gemini, Grok, DeepSeek or an OpenRouter key. Real tools behind switches you set; one fast native binary for Windows, Linux and macOS.
$ mindfork demo — try it in a minute: seeded data, a scripted engine, no model download, nothing written outside a temp folder. Already running Ollama? Two lines.
What it does
Memory that persists
A self-model with goals and observations, notes with a link graph, and RAG over your documents — it remembers you between sessions, and can look back through your other conversations if you let it.
Local first
A managed llama-server under the hood — one command downloads the llama.cpp build for your GPU — or the Ollama or LM Studio you already run, found and offered on the first start, or any OpenAI-compatible endpoint you point it at. Your conversations never have to leave your machine.
Four clouds and a gateway
OpenAI, Anthropic, Gemini, Grok and DeepSeek as first-class backends, and OpenRouter as a mode of its own — one key for chat, embeddings, speech and YouTube video, in front of every vendor's models, picked from a list that shows each one's context window and price. Switch per profile; API keys are stored with machine-bound encryption.
Real tools
An agentic loop with web search, page fetching, file access and a sandboxed Python runtime that reads the chat's files and hands back what it saves, charts included — plus your own tools over MCP.
Hand it your project
Attach a code project to a chat: the assistant reads, searches and edits it, and runs the build and test commands you typed — never ones it composed. F4 shows every change as a diff, with per-file revert.
It can delegate
The assistant can hand a side-task to a subagent with the same tools — several at once, or in the background while it keeps talking to you. The transcript is a conversation of its own — nested in the chat list, searchable, and streaming live while it runs.
A real TUI
Markdown, syntax highlighting, Mermaid diagrams and LaTeX rendered as text, dark and light themes, English and Russian — all in the terminal.
Show it a picture
Attach a screenshot, a photo or a diagram to your message — from a file, straight off the clipboard, or by its web address. Works with a local vision model, with all four clouds and through OpenRouter, wherever the model takes images.
Find anything you said
Search inside the open conversation, or across every chat's messages — a hit opens the chat at the matching message, highlighted. Subagent transcripts are searched too.
Long chats keep working
When a conversation outgrows the model's context window, the early part is folded into a rolling summary — automatically, or on demand with /compact — and the assistant can still read back what was folded away.
It can read replies aloud
/tts speaks the last message, the last N, or the whole conversation, through OpenAI or Gemini speech — or through any of the speech models behind an OpenRouter key, in a voice picked from the model's own list. No new vendor to sign up for if you already use one of them.
Private by construction
Portable data that lives next to the binary, encrypted backups, machine-bound keys, no telemetry. The assistant's web tools stay on the public internet — your own machine and network are off limits unless you say otherwise.
Around the app
Latest news
- 2026-10-31 mindfork 0.23.1 — a backup says what it holds
- 2026-10-30 mindfork 0.23.0 — pictures
- 2026-10-24 mindfork 0.22.0 — a platform for a companion