Purpose
Forge MCP is the MCP server for Voxell's hosted text embedding API (Forge), used to generate vector embeddings in semantic search and RAG scenarios. It exposes Forge to any MCP client and provides only two tools:
- embed: converts text into vectors
- list_models: lists available models and their dimensions
Main capabilities
- Three quality tiers: turbo (1024 dimensions, fast, default tier), pro (2560 dimensions), ultra (4096 dimensions, highest quality), allowing you to choose on the quality-versus-cost curve.
- Matryoshka (MRL): set dim to truncate (renormalize), yielding vectors about 4 times smaller and cheaper.
- Low latency (Go + CUDA engine), zero trust (authentication by key, mTLS can be provided).
- Benchmark performance: Voxell's Ingot-8B-R3 ranks first in the English entries of the public MTEB leaderboard (English v2), with an average score of 75.98 across 41 tasks.
- Stateless service: Voxell does not store the text you send or the vectors returned, and only records usage metadata (token count) for billing.
- Compatible with the OpenAI embeddings API: any OpenAI client can be pointed at Forge without changing code, while preserving existing vector dimensions.
How to use
Prerequisites: Node.js ≥ 18 (tested on 20); a Forge API key is required, which can be created at https://dash.voxell.ai turbo is free forever and requires no credit card.
Most MCP clients run on demand via npx; add the standard mcpServers configuration block to the client's MCP configuration (Cursor, Claude Desktop, Cline, Windsurf, and VS Code all use this structure); Claude Code can be installed with one command. After installation, replace your-key-here in the configuration with a real key from dash.voxell.ai.
The embed tool returns vectors, as well as the model, dimensions, and token count; the default is turbo.
Forge also supports the OpenAI embeddings API: existing OpenAI model names map to the corresponding Forge tiers with matching dimensions (text-embedding-3-small / ada-002 → 1536 dimensions, text-embedding-3-large → 3072 dimensions), so existing vector stores can be connected as-is; you can also use the Forge tier names turbo | pro | ultra directly, and it supports dimensions (Matryoshka, renormalized) and encoding_format: "base64".
Use cases
- Add semantic search: embed documents with input_type: "document", embed queries with input_type: "query", then rank by cosine similarity.
- Build RAG: embed a knowledge base, store the vectors, and retrieve the closest chunks to support an LLM.
- Find similar or duplicate text: embed two pieces of text and compare the vectors.
- Clustering or classification: after batch embedding, cluster the vectors or train a classifier on them.
- Shrink vector storage: set dim to truncate (Matryoshka), trading a little accuracy for smaller, cheaper vectors.
- Call directly in your editor: let an AI agent (Cursor, Claude, etc.) embed snippets, batches, or files through the embed tool, without writing a separate script.
Limitations
This server only does embedding—it does not provide storage, search, or RAG, which belong to different products. It only returns vectors. Voxell also has a retrieval pipeline built on Forge, with public evaluations: four public corpora (a total of 7,817 documents, 980,885 passages), 800 questions (200 per corpus), measurement date 2026-10-05, the top ten results for 95% of the questions include the correct document; these numbers describe the performance of that pipeline on these corpora, not a comparison with other vendors, and not the performance of the embed tool alone.
Other
License: MIT © Voxell, Inc.
