# ChimeraForge [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/Sahil170595/Chimeraforge  
**GitHub Stars:** 2  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/chimeraforge

## Description
Local-first LLM deployment planner: GPU/VRAM sizing, cost and latency, with provenance

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "chimeraforge": {
    "command": "uvx",
    "args": ["chimeraforge"]
  }
}
```

## Documentation

## What the ChimeraForge MCP server does

The ChimeraForge MCP server connects an MCP-compatible assistant to ChimeraForge, a local-first planner for choosing how to deploy language models. It evaluates combinations of model, quantization, serving backend, GPU count, and tensor or pipeline parallelism against constraints such as VRAM, quality, latency, cost, energy, and request rate.

The planner can work from a registry size class, a Hugging Face repository, an Ollama tag, or manual model details for an unreleased model. Its output identifies how each figure was obtained. Values can be measured, extrapolated, derived, estimated, or unknown; unsupported quality results remain unknown rather than being filled with fabricated scores.

## How it works

The server runs the `chimeraforge mcp` command over stdio. An assistant sends planning requests to the local process, which uses ChimeraForge's bundled tables and planning logic. The supplied corpus includes throughput, quantization, quality, safety, service-time, and third-party audit data, but its coverage is limited. Throughput rows are FP16 measurements from an RTX 4080 Laptop 12GB, while other GPUs may receive bandwidth-based extrapolations. Quantized throughput uses multipliers applied to FP16 rows rather than direct measurements of each quantization.

The ChimeraForge MCP server does not phone home and can operate air-gapped. It can also use model identifiers that require resolution through Hugging Face or Ollama, depending on the request and available network access. The command-line planner supports workload profiles derived from request logs or live vLLM and SGLang metrics; the MCP interface is described as exposing five tools, including the named tools below.

## Setup and configuration

Install the MCP extra with:

```bash
pip install "chimeraforge[mcp]"
```

Alternatively, configure a client to launch the package through `uvx`:

```json
{
  "mcpServers": {
    "chimeraforge": {
      "command": "uvx",
      "args": ["--from", "chimeraforge[mcp]", "chimeraforge", "mcp"]
    }
  }
}
```

The README provides configurations for Claude Desktop, Cursor, and Claude Code. If the package is already installed in the environment used by the client, the command can be shortened to `chimeraforge mcp`.

## Tools and capabilities

The documented MCP tools include:

- `chimeraforge_plan` for the full deployment search.
- `chimeraforge_suggest` for ranking options that fit a specified GPU.
- `chimeraforge_compare_api` for comparing self-hosting with hosted API cost and break-even volume.
- `chimeraforge_resolve_model` for grounding a model identifier in its parameters and architecture.

The planner can account for request rate, KV-cache settings, quality and safety gates, and measured benchmark data when those inputs are available. It can identify the gate that rejected all candidates in a zero-result plan instead of returning a generic failure.

## Limitations and notes

The bundled planning tables are much smaller than the total research corpus. The largest directly measured model is 3.21B, and predictions above that size use extrapolation. No SGLang throughput rows are present, so that backend follows the FP16 and power-law path. Every plan should therefore be read with its provenance labels and coverage warnings rather than treated as a benchmark of the target hardware.

_Full upstream README: https://allmcps.com/mcp/chimeraforge/readme_