01 / watch it work

A lead agent. A crew of workers. One set of weights.

Hand Judas a task. The lead splits it, the workers run in parallel on the same loaded model, and every write lands in the snapshot journal.

$ judas run "add a --json flag to export and update the tests"
model   one local GGUF, loaded once, 8 slots
lead      plan: 3 parallel jobs, 1 review
lead      spawn worker-1 (flag) worker-2 (tests) worker-3 (docs)
worker-1  read_file  src/export.cpp
worker-2  read_file  tests/export_test.cpp
worker-1  edit_file  src/export.cpp         snapshot saved
worker-3  edit_file  docs/export.md         snapshot saved
worker-2  shell      sudo apt install fixtures
           blocked in code: sudo is never allowed, in any mode
worker-2  shell      ctest   (bubblewrap sandbox)
reviewer  read the diff, 3 files changed, no objections
done       review the session in the web app, or revert it step by step

Illustrative session. Agent names and tool output are examples of the shape of a run, not a recording.

04 / everything in the box

One binary. The whole harness.

06 / install

Four steps to a working crew.

Linux, a C++23 compiler, CMake, Ninja, SQLite, OpenSSL, toml++ and bubblewrap. Judas links a llama.cpp you build for your GPU.

1. 1

   Build llama.cpp for your GPU

   Vulkan ROCm CUDA
   copy

   git clone https://github.com/ggml-org/llama.cpp ~/src/llama.cpp
   cmake -S ~/src/llama.cpp -B ~/src/llama.cpp/build -G Ninja \
         -DCMAKE_BUILD_TYPE=Release -DGGML_VULKAN=ON
   cmake --build ~/src/llama.cpp/build --target llama llama-common

   copy

   git clone https://github.com/ggml-org/llama.cpp ~/src/llama.cpp
   cmake -S ~/src/llama.cpp -B ~/src/llama.cpp/build -G Ninja \
         -DCMAKE_BUILD_TYPE=Release -DGGML_HIP=ON -DGPU_TARGETS=gfx1030
   cmake --build ~/src/llama.cpp/build --target llama llama-common

   copy

   git clone https://github.com/ggml-org/llama.cpp ~/src/llama.cpp
   cmake -S ~/src/llama.cpp -B ~/src/llama.cpp/build -G Ninja \
         -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON
   cmake --build ~/src/llama.cpp/build --target llama llama-common

2. 2

   Build and install Judas

   copy

   git clone https://github.com/00jnr00/judas-public judas && cd judas
   ./install.sh

3. 3

   Connect a model, in one step

   new

   judas models finds GGUF models already on the machine: your folders, LM Studio, the Hugging Face cache and Ollama. It shows size, quant, context and whether each one fits the GPU.

   copy

   judas models                     # what you already have
   judas models search WORDS        # look on Hugging Face
   judas models pull <hf-repo> --add --use

   sourcemodelquantcontextsizefits GPU

   folderyour-model-a.ggufQ4_K_MN kN GBfits

   LM Studioyour-model-b.ggufQ8_0N kN GBtoo big

   HF cacheyour-model-c.ggufQ4_K_MN kN GBfits

   Ollamayour-model-dQ4_K_MN kN GBfits

   Sample layout only. Real rows show your own files and numbers.

4. 4

   Start it

   copy

   systemctl --user start judas
   judas                            # terminal UI, or open http://127.0.0.1:8079

   The installer never enables or starts the service on its own: loading a model takes the whole GPU, so you choose when.

07 / hardware

GPU support, honestly.

If you run it on hardware not listed here, an issue with your GPU, backend, model and tok/s helps everyone.