An asynchronous developer queue and multi-agent orchestrator built specifically for local LLMs. Eliminates GPU idle stalls by automatically checkpointing Git state and context-switching tasks when human input is requested.
When running local models, hardware is strictly bounded by VRAM. Conventional coding agents freeze when asking questions, wasting hours of compute.
Local LLM inference requires dedicated VRAM. Running multiple agents concurrently causes VRAM thrashing and Out-Of-Memory (OOM) crashes.
NightCodeForge decouples tasks from the GPU worker through native Git branching, instant WIP state commits, and priority rescheduling.
llm-urgent priority.
Designed from the ground up as a production-grade developer platform combining determinism with local LLM capabilities.
A dedicated Hangfire inference queue enforces WorkerCount = 1 across tasks, guaranteeing zero VRAM thrashing or OOM crashes on consumer and workstation GPUs.
Direct execution via CliWrap on isolated feature/task-{id} branches. Zero reliance on brittle wrappers, with deterministic branch resets and squash merges.
Stage 1 reviews and prompts humans proactively before coding begins; Stage 2 handles implementation, test execution and repair; Stage 3 inspects live git diffs and fixes regressions.
Review → Code → AuditRun in-process native C# agents via Crucible, run external agents via Pi JSON-RPC, or connect open coding agents via the Agent Client Protocol (ACP).
Crucible • Pi • ACPRoute 95% of heavy backlog jobs to a dedicated high-VRAM LAN server while keeping your workstation GPU 100% responsive for real-time interactive ideation chat.
Local Host + LAN Server
The ncf CLI tool prepares repositories in seconds: merges gitignore, applies editorconfig, installs agent skills, and runs a 5-pillar project readiness audit.
ncf init & ncf audit
Explore the components powering NightCodeForge's autonomous workflow execution.
Pre-inspects the codebase, clarifies ambiguities, and asks the human developer clarifying questions upfront before touching any code.
Receives the approved technical specification and developer answers, modifies code, runs compiler/build loops, executes tests, and iterates to green.
Generates a live git diff against the base branch, evaluates against rigorous quality checklists, and applies corrective patches.
Running local models doesn't mean your workstation has to grind to a halt. NightCodeForge integrates first-class with OllamaRouter to dynamically partition workloads:
NightCodeForge is currently in active pre-release development. Here is exactly where the project stands today.
NightCodeForge is advancing rapidly through core architectural phases. Pre-compiled binaries, packaged desktop installers, and public NuGet packages are currently in development and will be released once initial integration phases conclude. In the meantime, full source code is public and can be tested from GitHub.
Single-worker Hangfire queue, task DAG dependencies, and EF Core persistence.
CliWrap branch sandboxing, automatic WIP checkpointing, and safe diff handling.
Blazor Server dashboard, live terminal stream, question inbox, and review room.
Crucible C# harness, ACP protocol support, and the ncf preparation tool.
Dynamic tool timeouts, containerized ICU support, command transparency, and test expansion.
Packaged standalone server binaries, NuGet global tools, and official download bundles.
# Clone the repository
git clone https://github.com/jetelain/NightCodeForge.git
cd NightCodeForge
# Build the solution (.NET 10)
dotnet build
# Run tests
dotnet test --filter "Category!=Acceptance&Category!=E2E"