bossless

AI you own. No API key. No cloud. No boss.

Status: proof of concept. This is the "AI Without Bosses" thesis as a runnable artifact, not a hosted product — there is no public instance and none is planned. Clone it, build it, own your own.

A complete AI assistant in one static binary: a chat web UI + an OpenAI-compatible API + a 1B-parameter LLM inference engine — written entirely in machin (MFL, the machine-first language), running on your CPU. scp it to a $4 VPS or a 2015 laptop, point it at open weights, and you have an assistant that answers to nobody but you: no account, no token meter, no telemetry, no terms of service, no request ever leaving the machine.

Inspired by AI Without Bosses (Prof. R. Trebor Scholz, The New School) — a course on democratically controlled, community-owned alternatives to Big Tech AI. This repo is that idea as a runnable artifact: claiming control of AI means being able to hold the whole thing in your hand.

What you get

• Chat UI at / — streaming tokens, multi-turn, dark, zero JS dependencies
• OpenAI-compatible API at /v1/chat/completions (stream + non-stream + tool calls) and /v1/models — point any existing OpenAI-client tool at it
• The engine: machin-colibri's pure-MFL int8 inference (vendored) — Q8_0 quantized llama2.c-format checkpoints, GQA, KV prefix cache, worker-pool parallel matmul. 20+ tok/s on a laptop CPU (i5-11320H, 6 threads); output greedy-argmax-identical to the C reference.
• One machin build produces the binary; --static gives you a musl binary that runs FROM scratch in a container.

Honesty, up front

The model is 1B parameters (Llama-3.2-1B-Instruct by default). It is a modest assistant: good at conversation, summaries, simple questions, simple tool calls — not a frontier model, and it will not pretend to be. The claim here is not parity with Big Tech; it is ownership: this is the floor of what you can fully own today, and the floor keeps rising.

Run

./build.sh                                # needs machin + a C compiler
./bossless models/llama32-1b-q80.bin 8091 # chat at http://localhost:8091/

Weights: any llama2.c version-2 (Q8_0, magic ak42) or ak44 checkpoint works. Convert Llama-3.2-1B / TinyLlama with llama2.c's export.py (GQA patch) — see the colibri README for the exact recipe. COLIBRI_THREADS=N sets the worker count (default 6).

Agent-first test hook: GET /?q=hello auto-sends a message — drive the full UI loop from curl + a headless browser, no clicking required.

Coop mode — AI without an admin

BOSSLESS_COOP=coop.db ./bossless models/llama32-1b-q80.bin 8091

One instance, many members, no root user:

• Join by invite — first boot prints a founding invite code; after that, any member can mint invites (/join?code=…). Members get a signed session cookie for the chat UI and a personal Authorization: Bearer token for the /v1 API.
• The ledger — every member's token usage (today / all-time) is visible to every member at /coop. Radical transparency instead of a billing dashboard only the owner can see.
• Policy by vote — instance policy (max tokens per reply, daily token budget per member) changes only by proposal + strict majority of all members: one member, one vote; a lone early voter can't decide for the coop. When a budget passes, the API enforces it (429 with "propose a higher budget and let the members vote").

Big Tech has admins; bossless has members.

Roadmap

• M2 — bossless agent: the roam-style tool loop backed by the local model.
• Static release binaries + weight download helper.
• Coop: invite links with names, proposal comments, more policies (retention, model choice).

Credits

Engine: machin-colibri (pure-MFL clean-room inference, inspired by karpathy's llama2.c checkpoint format). Framing: AI Without Bosses, Trebor Scholz. Part of awesome-machin.