ModelFoundry

ModelFoundry is Axiom-Labs' OpenAI-compatible optional model gateway, router, dashboard, and comparison layer for Hermes Agent, OpenCode, and adjacent local AI workflows. It combines benchmark-informed model quality with live provider health, then routes requests through a single local endpoint without forcing you to replace your primary paid model stack.

Why it exists

• One local endpoint for many providers
• OpenAI-compatible API for existing tools and scripts
• Automatic routing based on quality + latency + uptime
• Live dashboard and comparison layer for informed model decisions
• Optional provider lane for Hermes Agent and OpenCode instead of a hard stack replacement

Product position

ModelFoundry is best treated as:

• an optional provider lane
• a router
• a dashboard
• a comparison layer

It is not meant to replace your main paid stack by default.

Quick start

Install with npm

npm install -g model-foundry
model-foundry

Then open:

• Dashboard: http://localhost:7352/
• API base URL: http://127.0.0.1:7352/v1
• API key: any string
• Default routed model: auto-fastest

Run with Docker

mkdir model-foundry
cd model-foundry
curl -fsSL -o Dockerfile https://raw.githubusercontent.com/Codename-11/model-foundry/master/Dockerfile
curl -fsSL -o docker-compose.yml https://raw.githubusercontent.com/Codename-11/model-foundry/master/docker-compose.yml

docker compose up -d --build

Core CLI

model-foundry [--port <number>] [--log] [--ban <model1,model2>]
model-foundry onboard [--port <number>]
model-foundry install --autostart
model-foundry start --autostart
model-foundry uninstall --autostart
model-foundry status --autostart
model-foundry update
model-foundry refresh-scores
model-foundry autoupdate [--enable|--disable|--status] [--interval <hours>]
model-foundry config export
model-foundry config import <token>

Legacy binary alias:

modelrelay

Integrations

Run onboarding for guided setup:

model-foundry onboard

The quick-start path now recommends OpenRouter as the default first provider, with Groq as the fast single-provider option and OpenAI-Compatible as the bring-your-own-upstream path.

Manual docs:

• Hermes Agent integration
• OpenCode integration
• Hermes Proxy integration
• Frontier provider catalog

Docs

• Documentation index
• Architecture overview
• Routing, scoring, and evaluation
• Configuration and operations
• CLI reference
• Roadmap
• Code Arena score reference

OpenAI-compatible endpoints

POST /v1/chat/completions

Use model: "auto-fastest" to let the router choose the current best backend. You can also target grouped model IDs such as minimax-m2.5, kimi-k2.5, or glm4.7 and let ModelFoundry choose the best provider for that model family.

GET /v1/models

Returns the router-exposed model list, including grouped slugs and auto-fastest.

Config basics

• Canonical config path: ~/.model-foundry.json
• Legacy compatibility mirror: ~/.modelrelay.json
• Request logging is off by default; enable with --log
• Config transfer is supported with model-foundry config export / import

Supported API key env vars include:

• NVIDIA_API_KEY
• GROQ_API_KEY
• CEREBRAS_API_KEY
• OPENROUTER_API_KEY
• OPENCODE_API_KEY
• OPENAI_COMPATIBLE_API_KEY
• ANTHROPIC_API_KEY
• OPENAI_API_KEY
• CODESTRAL_API_KEY
• SCALEWAY_API_KEY
• QWEN_CODE_API_KEY or DASHSCOPE_API_KEY
• KILOCODE_API_KEY
• GOOGLE_API_KEY

The frontend can also use the shared frontier catalog endpoint:

• GET /api/provider-meta

Development

Install deps:

npm install

Run the backend and Vite frontend together:

npm run dev

That gives you:

• Vite frontend: http://localhost:5173/
• Router/backend API: http://127.0.0.1:7352/

Build the production frontend bundle:

npm run build

Run tests:

npm test

Production/server mode serves the built dist/ app automatically when it exists.

Project

• GitHub: https://github.com/Codename-11/model-foundry
• Organization: Axiom-Labs