AISBFWhat CoderAI is
CoderAI is a multimodal and multi-backend local model orchestrator with an OpenAI-compatible API server. It runs models on local GPUs and supports NVIDIA CUDA, AMD Vulkan, Intel Vulkan, macOS Metal paths, plus CPU fallback where appropriate.
OpenAI-compatible API
Expose /v1/models, /v1/chat/completions, completions, embeddings, media generation, and generated file routes from your own machine.
Browser generation UI
Use the Web Studio and admin UI for chat, image, video, audio, pipelines, archives, models, settings, users, and tokens.
Broker worker mode
CoderAI can connect outward to AISBF over WSS, register capabilities and hardware, then receive routed work without opening inbound firewall ports.
Documentation map
CoderAI does not require AISBF
CoderAI can run standalone with its own browser UI and OpenAI-compatible API. AISBF is an optional routing, metering, clustering, and provider-management layer when you want to sell or share CoderAI capacity through a broker.
Install CoderAI standalone →
Use the CoderAI API directly →
AISBF and CoderAI tutorials →
Optional AISBF broker tutorial →
Install from the packages
docker pull ghcr.io/nextime/coderai:latest — the full CoderAI as one public image; nineteen single-capability images beside it; CoderAI-Setup.exe for Windows. Nothing to build.
Install and run
Build scripts, platform support, default ports, admin login, config files, package outputs, and driver expectations.
API reference
Base URL, authentication, OpenAI-style text routes, image/video/audio endpoints, archives, profiles, and pipelines.
Inference engines
Per-model engine selection: Transformers, llama.cpp GGUF, ds4, colibri, kimi-k3-in-c, ktransformers/SGLang and vLLM — how each one is chosen, built and supervised.
Remote GPUs (RunPod)
Serve a model on a GPU rented by the second: pod pools with price and VRAM caps, serverless endpoints, spend budgets, idle teardown, the stale-pod reaper and local-to-cloud spillover.
Remote executionCluster
Several machines as one CoderAI: other installs as engines of one front, found over mDNS with one shared token; a GGUF over machines through llama.cpp RPC, vLLM on Ray, requests split over every card that has the model, a video pipeline's parts on different machines, LoRA training data-parallel across nodes; prefix-cache routing, Prometheus metrics, node logs, signed images.
Speech & speakers
Per-model STT backends, word timestamps, translation, diarization, and enrolled speaker identification and verification.
Document OCR
PaddleOCR, docTR and Surya behind /v1/ocr, with schema-driven structured extraction and stamp/signature detection.
Embeddings & rerank
Text, image, multi-vector BGE-M3, GME-Qwen2-VL, GeoCLIP geolocation, visual place recognition, and cross-encoder reranking.
Multimodal capabilities
Text, image, video, audio, 2D/3D, profiles, pipelines, model capability badges, common workflows, and bundled tools such as the video editor, VideoGen Studio, and Township Fighters demo.
Demo/example tools
CoderAI ships three demo apps under tools/: video_editor.py for timeline editing, videogen.py for profile-driven short movies, and gen_township_fighters.py for an MMA-style generated video workflow.
AISBF broker integration
Provider identity, user/global scopes, WebSocket endpoints, registration envelopes, heartbeats, and routing gotchas.
Operations
Frontend/engine split, multi-engine routing, VRAM/RAM offload, queues, thermal protection, tokens, archives, and troubleshooting.
Docker / OCI image
Build or run the provided OCI image with coderai-docker, persistent volumes, GPU modes, external storage, logs, AISBF broker config, and bundled tool routes /editor/, /videogen/, and /township/.
Reverse proxy
nginx root/subdomain and sub-path configurations, large upload settings, SSE streaming, and forwarded headers.