News

mindfork 0.23.1 — a backup says what it holds

The backup made before a data upgrade no longer warns the version before the upgrade against its own data: every backup now records the format of the files it packs, and a warning that is due names the files that are newer.

mindfork 0.23.0 — pictures

The assistant can draw: with OpenAI, Gemini, xAI or one of OpenRouter's picture models, whatever model you chat with — and change a picture the chat already has. /image generate draws one for you, and a picture asked for from Telegram comes back to the phone. Spoken replies no longer fall silent in the middle.

mindfork 0.22.0 — a platform for a companion

A profile can now be a companion: it knows when you last spoke, has your notes from the first message, comes back on its own within quiet hours, keeps every version of its self-model, forks and travels, reads character cards, speaks in a voice of its own, and answers from your phone through Telegram — text, pictures and voice messages.

mindfork 0.21.0 — DeepSeek as a provider of its own

DeepSeek gets a mode of its own: the window and image support read from its list, its thinking switch from the first request, and /continue through its prefix completion. Measuring it found the external route failing /continue there. The token counter moves to the conversation's frame.

mindfork 0.20.0 — a conversation reused from turn to turn

The cache share 0.19.0 added showed most turns reusing nothing. The self-model's observations are now picked by what the chat opened with, so from the second turn the counter reads cache 94–99%. Every observation picked fits the prompt, an old 1200-character limit is raised, and DeepSeek's API works.

mindfork 0.19.0 — how much of a conversation was reused

The token counter now shows how much of the prompt the provider read from its cache: cache 99% on a warm turn, cache 0% when nothing was reused. MINDFORK_MODEL starts a local engine on its own, and the zip library is updated past three advisories.

mindfork 0.18.3 — a plain code block can be quiet

A code block with no highlighting can now be drawn without its reversed background, the way a highlighted block looks. The animated demo is drawn that way, so its diagram no longer flashes up as a bright block of source first. The default is unchanged.

mindfork 0.18.2 — what a tool tells the model is what happened

Nine fixes from running the live tests on Gemma 4, Qwen and gpt-oss. A tool's answer now says what actually happened — a scene already staged, an image not shown, a note already written, a file saved where nothing keeps it, a call that arrived broken — so a turn ends in an answer instead of code, silence, a made-up link or a loop.

mindfork 0.18.1 — where everything is

F1 → About now names the folders mindfork uses: the program's, the data's with the setting that chose it, and the logs'. Full paths, ready to paste into a file manager or a shell — and on a rented GPU box, one look says whether the data is on the volume.

mindfork 0.18.0 — a day on a real Mac

mindfork has run on macOS since 0.15.0, tested on CI runners with no screen. For this release it ran for a day on a Mac mini (M4) with macOS 26 and 15: installs, Metal, Ollama and LM Studio, three terminals key by key. What the day found is fixed, and mindfork keys shows what your terminal sends.

mindfork 0.17.0 — your Ollama or LM Studio, found and offered

Start mindfork with no model set up and it looks for a local Ollama or LM Studio, lists what they serve, and connects the one you pick with Enter — its embedding model too. LM Studio joins Ollama as a server mindfork knows: its window is read, and its silent cut of a long conversation, which drops the whole middle, is told.

mindfork 0.16.1 — when Ollama cuts the prompt, you are told

Ollama answers a prompt longer than its window with a reply, not an error: it drops the oldest messages, and when the last one alone does not fit beside the instructions, it cuts the prompt to half the window from the start. mindfork now tells such a cut from the size Ollama reports, says once that the reply was made without part of the conversation, and folds the conversation into a summary when there is anything to fold.

mindfork 0.16.0 — Ollama, with nothing to keep in step

mindfork now reads the context window of an Ollama server from Ollama itself, so a long conversation is folded into a summary before Ollama would cut the prompt — which it does without a word, the instructions first. A window too small for the conversation is said once, with the setting that raises it, and the first reply no longer ends in advice meant for llama-server.

mindfork 0.15.0 — on a Mac, in Scoop, and an animated demo

mindfork comes to macOS on Apple Silicon as a preview — through Homebrew or the same install line as on Linux, with stored keys and Metal for local models — and to Windows through Scoop. Every release now draws an animated demo from its own interface. Option and Alt move by words in every terminal, Ctrl+J breaks a line where nothing else can, and on Windows text typed with AltGr is no longer lost.

mindfork 0.14.1 — a rented GPU box that starts, or says why not

On a rented B200 the one-line install ended with neither server working and a status that blamed the model. A local server that dies now shows the error it logged and names the llama.cpp build that works, a port another program holds is refused, a bare Linux image gets the library llama.cpp needs, and the GPU kernels a server compiles are kept with the data — about a minute back on every start of a B200.

mindfork 0.14.0 — a window of any size, and one reasoning effort for every model

A window too small for the screen in front says so instead of drawing a frame with parts missing, and the chat gives up its chrome one piece at a time and stays a chat down to three rows. The reasoning effort is one setting for every model: a value a model does not have becomes the nearest it has, with a new top tier, max. And the models a recent setting had stopped — Claude Haiku 4.5, three of Grok's, titles on GPT 6.1 and Gemini 3.8 Flash — answer again.

mindfork 0.13.0 — OpenRouter as a mode of its own

One OpenRouter key for the assistant, impersonation, memory, speech and YouTube video: a mode beside the four clouds and your own server, with a model list that shows each model's context window and price, and a key checked before your first message. A message read aloud no longer stops half way, and an edit of the embedding settings no longer switches the embedder's checks off until the restart.

mindfork 0.12.0 — the app paints its own window, and a theme can be yours

Three ways of drawing the interface: the one it has always had, one where the app paints its own background with every colour held to a contrast floor, and one with no colours at all — NO_COLOR included. A theme is a file with only the background in it; the colours you leave out are fitted to stay readable on it. And one misspelt value in settings.json no longer costs every setting, the stored API keys included.

mindfork 0.11.2 — a bigger model on a smaller card, and a long page read whole

The managed llama.cpp server takes the arguments the settings have no field for, so a 26B mixture-of-experts model fits in 8.4 GB of video memory instead of 17.3. And a page the assistant fetches arrives whole up to a million characters and can be searched in the same reply — the last chapter of an 81-page specification, found with one search.

mindfork 0.11.1 — the install script works on a rented pod

0.11.0's one-line install failed on RunPod before it had unpacked a single file. The first real run found it; this release carries the fix, the test that reproduces the pod's shape, and the result that the rest of the line — llama.cpp with CUDA, the sandbox, a 31B model verified at a 131k context — ran on that pod as designed.

mindfork 0.11.0 — from a bare Linux machine to a local model, in one line

A new command installs the Python sandbox and llama.cpp and writes the model, its projector, the embedding model and the context into the settings — checked before anything is downloaded, and started once to prove it loads. An install script puts the app on the machine first. Written for a rented GPU box that is new at every stop; just as good on a desktop. And llama setup installs CUDA builds on Linux, which it had been refusing.

mindfork 0.10.2 — which copy of your data is the newest

One command says when each copy of your data was last written and what it holds — the data on this computer or a backup archive, without restoring anything — and a second says what each copy has that the other lacks. And opening a chat no longer flashes the one you were leaving.

mindfork is public

The repository is open, the first public release is out, and this is what the project is — an AI chat that lives in your terminal, with memory that persists and tools that stay behind switches you set.

mindfork 0.9.9 — files in, charts out, work in the background

The assistant's Python code reads the chat's files and hands back what it saves. Subagents and staged dialogues run in the background while you keep talking, with a screen to watch them. The engine downloads itself — and the web tools wait for your say-so.

mindfork 0.9.8 — give it real work

A code project attached to a chat — read, edit, build and test, every change reviewable and revertible. Subagents with real tools and live transcripts. And web search that can stop being rate-limited.

mindfork 0.9.7 — wherever your terminal is

Every action has a typed command, copying works over SSH, a conversation exports to a file — and the assistant can find what you said in another chat.

mindfork 0.9.6 — the model can see

Images in a message from a file, the clipboard or a URL; tool screenshots reach the model; the web tools stop at your front door; a cloud blip no longer costs the turn.

Hello from mindfork 0.9.5

The project gets a home on the web — and a status report: an interactive demo, a fourth cloud provider, history compression.