Most AI tools are private — you and a model, alone in a tab. The Clan makes AI team infrastructure: models are contacts, applications are participants, prompts are shared artefacts, and the conversation is the workspace. Write, illustrate, analyse, build — together.
A group chat where AI models are full members. Tick the boxes for who should hear you — two colleagues, a language model, a vision model, a code executor — and send. Replies land in the same thread for everyone to see.
Humans and AIs draw from the same message log. No copy-paste between tools, no lost context, no "which tab was that in?"
Draft a story with one model, illustrate a scene with a vision model, sanity-check the code with another — all in one thread, all visible to the group.
Add a model to your contacts, give it an alias and a priority, route to it like a person. Different people can talk to different models in the same room.
Drop a sketch, a photo, a screenshot. Vision models see it; the group sees it; it's archived with the thread. No "attach this file separately."
With a shared key, encrypted exchanges show a green sender name, plaintext shows red. Trust you can see, not assume.
Highlight any message or model reply → capture it as a reusable prompt in one click. Your team's best prompts live in the sidebar, not in scrollback.
The problem isn't that AI models are unavailable —
it's that their work is scattered across private tabs,
invisible to the team, and impossible to audit.
The Clan puts the models in the room.
The system is not a generic AI tool. It is a shared workspace where humans and AI models collaborate as peers within a single conversation thread — with inspectable context, reusable prompts, and visible trust signals. It is as comfortable with a story draft as with a code review.
The router keeps a single message log that everyone draws from. A human asks a question, an AI answers, a second human quotes that answer, a script runs and posts its output — all in one stream, with sender, recipients, and reply-chains preserved.
Nobody copy-pastes between tools. The AI sees the human discussion; the humans see the AI's reasoning. The group's context is collective.
Draft a story beat with one model, ask a vision model to paint the scene, have a third model critique the pacing, then send the whole thing to a colleague for a human read. The conversation is not text-only and it is not code-only. It is whatever the group is working on.
Drop a sketch, a photo, a screenshot, a generated illustration. Vision models can see it directly in the thread; the humans see it too; it is hashed, stored, and archived with the message. No "attach this file separately, then paste the link into chat."
Generated images come back into the same thread as attachments, so the group's visual history is as intact as its written history.
Any script, Python tool, or HTML app on the server can be added to the sidebar with an alias and launched by anyone with a checkbox. The Code Executor receives a block of code and returns the output — rendered as an image, a live HTML page, an audio clip, or text — inside the chat.
The barrier between "chatting about an idea" and "running the idea" is one button press.
You add a model to your contacts once. It appears as ol.llama3,
hf.flux-vision, or under a friendly alias you set. You can
give it a priority number, toggle between raw IDs and human-friendly
names, and route a message to it exactly the way you'd route it to a
person.
Different people in the group can be talking to different models in the same room without stepping on each other.
You can claim an alias and protect it with a secret. It becomes a durable identity others can address, not a throwaway session name. The group has stable members.
When a shared AES-256 key is set, sensitive operations — AI calls, file writes, process launches, private history — must arrive encrypted. Incoming sender names render green when the frame was genuinely encrypted and red when it was plaintext.
That's not decoration; it's a visible trust signal at a glance. Everyone can see whether a given exchange happened over a protected channel. Human-to-human chat stays open by default; the moment a model or the executor is involved, the gate applies.
The "VIEW" panel lets each person filter the inbox by sender. You see the whole group's traffic, but you can narrow to the voices you care about. Individual attention, collective record.
The group can snapshot the entire message log as a named context and start fresh — or load a prior context back in. Long-running collaborations — a novel, a research thread, a codebase — become resumable.
This is the part prompt engineers, writers, and researchers feel immediately. The system treats prompts as first-class, shared, and inspectable — not as throwaway text in a scrollback buffer. A prompt for a story opening and a prompt for a code review live side by side.
The system narrows the audience considerably — it is a coordination layer for mixed teams, not a general-purpose chatbot. Here's who leans in.
People who already hop between ChatGPT, Claude, and local Ollama models depending on the task. They feel the friction of copy-pasting context between tabs and want a single thread where a model's answer is visible to the whole team.
People drafting fiction, tabletop campaigns, or narrative games who want a text model for prose, a vision model for scene art, and a human editor in the same room. The thread becomes the story bible.
Designers and illustrators who want to iterate on images with a vision model while a colleague comments in real time. Drop a sketch, ask for a variation, keep every version archived with the conversation.
The highlight-to-capture-prompt feature and the exact payload debug view are catnip for this group. They want to see what the model actually received, iterate on prompts collectively, and build a shared library.
The crowd around EleutherAI, Nous Research, and local model communities. They care about inspectability, self-hosting, and not shipping every prompt to a SaaS vendor. A local router is exactly their shape.
Fixed prompts, context save/load, and a shared message log make this a strong teaching environment. A facilitator demonstrates a prompt, the group sees the model's response, everyone captures the useful bits.
Legal, research, and internal-ops teams that cannot paste client data into a web UI. The encryption gate on sensitive operations and the ability to run everything on localhost is a real differentiator.
Paper clubs and research communities are natural early adopters for the "collective context" and "inspectable payload" features. Shared reasoning, auditable prompts, resumable threads.
Don't sell it as "a better chat app."
Sell it as the missing coordination layer
for teams that already use AI — to write, to draw, to build.
If you're assembling a human-AI group, this is the room you put it in: one log, many minds, every exchange inspectable, every prompt reusable, every tool one checkbox away. Bring a story, a sketch, a dataset, or a codebase — the room adapts.