CAGE — Composable Agent Graph Engine: Self-hosted AI coding agent you wire up as a graph

I built AIWrapper for one job: leave it running when I go to bed, and have something finished in the morning. Not a tool that stops on question 40 to ask whether it may write a file, and not one that tells me my session is out. It's self-hosted — a C++20 inference server plus a VSCode client — so the models are mine, the GPUs are mine, and a run ends when the work is done rather than when a quota says so. Nothing leaves the machine unless a node is explicitly pointed at a remote endpoint. Three things make the unattended run actually work: • The Logic Graph. The agent loop isn't hard-coded, it's an executable graph you edit visually — prompt nodes, format validation, multi-case branching, dispatch/fan-out, loops with re-entry. Each node picks its own model, quantization, GPU set and sampling, so a small fast model plans and a large one writes, inside one run. You design the loop up front instead of steering it all night. • Interruptions are the exception, not the interface. File writes and patches land in a per-session staging area and never touch your workspace, so there is nothing to approve mid-run — you read the whole diff once, in the morning, and flush it. There is an AskUser node, but it runs a model first to decide whether the question is worth waking you for; most tasks produce none. • It isn't only code. Image, text-to-speech, audio, music and image-to-3D mesh each run as their own isolated worker, so the same graph that writes the code can also produce the art, the voice lines, the soundtrack and the meshes. Every asset a project needs, from one run — that's the target, not an add-on. Underneath: topology-aware multi-GPU placement (a no-alloc dry run measures the real requirement, then it reads the NVLink topology and picks split mode, GPU layers, tensor split and MoE CPU offload — you just pick the cards), retrieval over your project, MCP for every tool, accounts and per-user sessions. Honest state, since this is 0.1, the first evaluation build: • Local inference is what I run daily and what has been tested. Remote OpenAI-compatible endpoints are implemented but barely exercised — point it at a hosted API and expect rough edges. Same for the remote image/TTS/audio providers. • Of the nine worker environments, two need a hand: music wants FFmpeg headers (the Docker image carries them), and the 3D one needs upstream's interactive setup. • Server-side session history is recorded but not yet replayed into a resumed conversation. The control plane sees several machines, but there is no scheduling across them. • Requirements: Linux x86-64, NVIDIA driver + CUDA 12 (compute capability 7.0+, so V100 / RTX 20xx and newer; CPU-only works, slowly), VSCode 1.85+. Docker or native install, both documented. It will keep getting better, but toward that one goal above and on my own roadmap — so read the gaps as known rather than as a request queue. That said, I do read the thread, and anything genuinely useful in it has a good chance of turning up in a later build; I just won't be working the comments much. Binaries are free to evaluate (proprietary license; the source and full commercial rights are available if someone wants to talk).

AGE 0.1.1 is out (formerly AIWrapper). GPU support is now a separate per-card download, so the server itself is ~19 MB. New AMD/ROCm packs cover RDNA2 through RDNA4 and CDNA — they load cleanly but have never run on real AMD hardware, so that's the part I'd most like tested. Breaking update: refresh the server and the extension together. https://github.com/real0000/AIWr...