uncensored-fleet

Deploy a local multi-model LLM fleet in one command — llama.cpp slots, an agent harness, and engram memory.

Your hardware. Your models. No API keys, no rate limits, nothing leaves the box.

pip install cognis-uncensored-fleet
bash scripts/build-llamacpp.sh     # build the engine (CUDA/Metal/Vulkan auto)
fleet pull all                     # download the model fleet
fleet up uncensored                # start the commander slot
fleet agent "summarize ./notes and propose next steps"

🔎 Example output

Real, reproducible output from the tool — runs offline:

$ uncensored-fleet-emit --version
uncensored-fleet 0.3.0

$ uncensored-fleet-emit --help
usage: fleet [-h] [--version] {models,pull,up,down,status,setup,run,agent} ...

uncensored-fleet — local multi-model LLM fleet + harness (engram-integrated)

positional arguments:
  {models,pull,up,down,status,setup,run,agent}
    models              list fleet slots
    pull                download a slot's model (or 'all')
    up                  start a slot (or 'all')
    down                stop a slot (or all)
    status              show fleet status
    setup               install llama.cpp + pull all models
    run                 one-shot prompt to a slot
    agent               run the agent harness on a task

options:
  -h, --help            show this help message and exit
  --version             show program's version number and exit

$ uncensored-fleet-emit status
reasoning    :8771  down  planning, analysis, chain-of-thought
  math         :8772  down  math, SymPy-verifiable problems
  coding       :8773  down  code generation + edits
  vision       :8775  down  image understanding / OCR
  uncensored   :8774  UP    lead/commander, unrestricted local reasoning

Blocks above are real uncensored-fleet output — reproduce them from a clone.

Usage — step by step

1. Install the fleet command, then build llama.cpp + pull models (setup prints the platform script to run first):

   pip install -e .
   fleet setup

2. List the slots to see each model, its port, VRAM and role:

   fleet models

3. Pull and start a slot (use all for the whole fleet), then check it's serving:

   fleet pull all
   fleet up uncensored
   fleet status

4. Use the fleet — send a one-shot prompt to a slot, or run the agent harness on a task (--slot defaults to uncensored):

   fleet run uncensored "summarize this repo"
   fleet agent "scaffold a fizzbuzz module with tests" --slot coding

5. Tear down when finished (omit the slot to stop all):

   fleet down

Contents

• Why · The fleet · Quick start · The harness · Engram memory · Explore the suite

Why

Cloud LLMs gate you on price, rate limits, and content policy. uncensored-fleet stands up a fleet of local models — reasoning, math, coding, vision, and an abliterated "commander" — each served by llama.cpp on your own GPU, and gives you a model-agnostic agent harness to drive them. It is the deployment layer for a private, unrestricted, self-improving local AI stack.

↑ back to top

The fleet

Slot	Role	Default model	Port
reasoning	planning / analysis	DeepSeek-R1-Distill-Qwen-7B	8771
math	SymPy-verifiable problems	Qwen2.5-Math-7B	8772
coding	code gen + edits	Qwen2.5-Coder-7B	8773
vision	image / OCR	Qwen2-VL-7B	8775
uncensored	commander, unrestricted	Josiefied-Qwen3-8B-abliterated	8774

VRAM-aware: conflicting slots auto-evict. Override any slot in fleet.yaml.

Quick start

fleet models                 # list slots
fleet pull all               # download GGUFs (huggingface_hub / hf-cli / direct)
fleet up all                 # start servers (or: fleet up coding)
fleet status                 # see what's live
fleet run coding "write a python LRU cache"
fleet agent "audit ./repo for secrets and write FINDINGS.md" --slot uncensored
fleet down                   # stop everything

The harness

A tiny, dependency-free agent loop (fleet/harness.py) talks to any slot over the llama.cpp OpenAI-compatible endpoint and supports safe tools (run_bash, read_file, write_file) via a simple TOOL:: / FINAL:: text protocol — model-agnostic, so it works with whatever you load.

Engram memory

The harness remembers. It uses the Cognis engram fork for portable, model-agnostic long-term memory when installed (pip install "cognis-uncensored-fleet[engram]"), and falls back to a local SQLite store otherwise. Every task's outcome is recalled on the next related run.

Explore the Cognis Neural Suite

uncensored-fleet is the local-AI backbone of the Cognis Neural Suite (170+ tools). Pair it with:

• 🧠 engram — model-agnostic agent memory
• 🛠️ skills — agent skill registry the harness can load
• 🤖 agentsmith · evalbench · modelroute — orchestrate, evaluate, route
• 📚 awesome-cognis · cognis-sources — the full index

Responsible use

Local, unrestricted models are powerful. Use them lawfully and ethically; you are responsible for what you generate and run.

How it fits

flowchart LR
  U[You / CI / Agent] --> R[uncensored-fleet]
  R --> O[Outputs & artifacts]
  R --> M[MCP / JSON]
  M --> AI[AI agents]
  R --> S[Cognis Neural Suite]

Loading

Explore the suite → 🗂️ all tools · ⭐ awesome-cognis · 🔗 cognis-sources

Interoperability

uncensored-fleet composes with the 300+ tool Cognis suite — JSON in/out and a shared OpenAI-compatible /v1 backbone. See INTEROP.md for the suite map, composition patterns, and reference stacks.

Integrations

Forward uncensored-fleet's findings to STIX/MISP/Sigma/Splunk/Elastic/Slack/webhooks via cognis-connect. See INTEGRATIONS.md.

License

Source-available under the Cognis Open Collaboration License (COCL) v1.0 — see LICENSE. Commercial use: licensing@cognis.digital.

---

_{Cognis Digital · part of the Cognis Neural Suite · Making Tomorrow Better Today}