Forge itself is closed
The CUDA embedding engine, the serving stack and the models are private. What is open is every path into them, which is the part that decides whether you can leave.
OPEN SOURCE // MIT, AND HONEST ABOUT THE EDGES
The clients are yours.
The engine is not.
Everything you need to call Voxell from your own stack is MIT licensed and on the public registries: a Go client, an MCP server, LangChain and LlamaIndex integrations, a Vercel AI SDK provider. The embedding engine and the models are not open, and that line is drawn on purpose rather than blurred.
Clients and integrations
Call Forge from the tool you already use. Each of these is a thin, readable client you can fork.
forge-gov0.1.0 · MIT
Go client, native gRPC
Speaks protobuf over mutual TLS rather than HTTP and JSON. Vectors travel as packed binary at four bytes per float, and the client authenticates with a certificate instead of putting a bearer token on the wire.
Read the source →forge-mcpv0.1.5 · MIT
MCP server for coding agents
Gives Claude Code, Codex, Cursor and any stdio MCP client a retrieval tool backed by your own corpus. The most used thing on this page: 481 installs in the last thirty days.
npx -y @voxell/forge-mcp →langchain-voxell0.1.1 · MIT
LangChain embeddings
Sets Forge's asymmetric input type for you, so queries embed as queries and documents embed as documents. Mixing those two is a quiet way to lose recall.
pip install langchain-voxell →llama-index-embeddings-forge0.1.0 · MIT
LlamaIndex embeddings
The same asymmetric handling, wired into LlamaIndex. Matryoshka dimensions are real here: a truncated vector is re-normalized, so a shorter dimension is a usable prefix rather than a broken one.
Read the source →forge-ai-providerMIT
Vercel AI SDK provider
Drop-in for the AI SDK's embedding interface. The OpenAI-compatible path preserves your existing width, so a swap does not force you to rebuild an index you already have.
Read the source →n8n-nodes-forgeSOURCE ONLY
n8n embedding node
An AiEmbedding sub-node for n8n workflows. Complete and readable, but not published to npm yet, so today it is source you clone rather than a package you install. The README says otherwise and is wrong.
Read the source →Tools that came out of the work
Built because something broke here first, then made general enough to hand over.
featherweightv0.1.0 · MIT
Cross-device state without a backend
A static site has no server to remember anything, and the usual answers are a framework runtime or a whole backend to store one blob. This keeps the state in the browser, instant and offline, and syncs it between devices through an edge key-value store that has no idea what the data means. Around 1 KB, zero dependencies, no build step, and it works with any framework or none.
npm i @voxell/featherweight →cssguardv0.1.0 · MIT
The other half of CSS tooling
Purge tools find CSS with no HTML. This finds the inverse: an HTML class with no CSS definition, the failure where a purge step removes a style that JavaScript adds at runtime and the page silently stops working. It found the bug on this site.
go install …/cssguard@latest →warp_pos_spacySOURCE ONLY
A pressure gauge for template collapse
Generated text often looks varied while repeating one sentence shape underneath. This hashes small part-of-speech windows and reports whether a corpus is drifting into a monoculture. spaCy keeps the tagging and parsing; the hot loop moves to Rust, where the original scoring job went from about 3.5 minutes of Python loops to about 6.6 ms. Roughly 1,500x on that step. Source only, not yet packaged or licensed.
Read the source →mash-verificationSOURCE ONLY
Check the sorting benchmarks yourself
The harness, the integrity proofs and the statistics behind MASH Sort's numbers. The sort implementation itself is not included, so this verifies the claims without opening the algorithm. Not yet licensed.
Read the source →Lux, the open half
The client and the edge are MIT. The high-performance broker is the commercial product.
lux_sdk0.1.1 · MIT
Local-first state sync client
7,220 bytes raw, 2,600 gzipped, zero dependencies. Reads hit local storage immediately and sync over a WebSocket, so an offline tab keeps working and reconciles when it comes back.
npm i @voxell/lux →lux_osMIT
The edge, on your own account
Cloudflare Functions and the key-value layer behind them, deployable to your account so state sync needs no backend of yours and no vendor between you and it.
Read the source →What is not open
Worth saying plainly, because it is the first thing worth checking.
Weights are not published
The research model has a public model card and a demo endpoint, but the weights behind the commercial tiers are not downloadable. If that is a hard requirement, say so early and we will tell you honestly whether this fits.
MASH is verifiable, not reproducible
You can re-run the benchmarks and check the integrity proofs. You cannot read the algorithm. That is a real limit on what independent verification means here, and it is stated rather than left to be discovered.
Fork it, or just call it.
The Turbo tier is free forever and needs no card, so the fastest way to evaluate any of this is to run a real call before you read a line of the source.