Overview
Discover
Effects Examples Copyright protection New
Use it
Batch Manager Clipboard Manager Command line Your own programs AI agents (MCP) Team server
Reference
Getting started API reference Effects reference Math library
Download

Team server

One machine runs the engine, the whole team uses it. The OpticScript MCP server does not have to run on every desk. Start it once, on a fast machine on your network, and every workstation's agent (Claude Code, Claude Desktop, your own) gets the same engine, the same version and the same fonts. Images travel as an id, not as base64 through the model.

One OpticScript server for the whole team: workstations talk MCP over HTTP to mlcos-mcp; images and data go through the mlcartifact store by id

Two pieces on the server

  • mlcos-mcp — the OpticScript MCP server, this time over HTTP instead of stdio. It runs every script in a sandbox and wants a bearer token.
  • mlcartifact — a small store for files. A client puts an image in, gets an id back, and hands the model only that id. Results land there too.

Your workstations need nothing but an MCP client. For files that live on the workstation there is artifact-cli, a single binary that uploads and downloads.

Why ids and not base64

A model can pass an image to a tool as base64. For an icon that is fine. For a photo it is not: the model has to write every character into the tool call, and back comes the result the same way.

What passes through the model: 1,740,108 characters of base64 for a 1.3 MB PNG, against two artifact ids of 27 characters each

With the store in between, the model sees mlcartifact://1c48-78174ca3 going in and another id coming out. The pixels never pass through the context.

What it looks like

A real run from a Windows workstation (Claude Code in WSL). The task was to convert a local 1.3 MB PNG to a 640 px WebP; the prompt below is shortened, everything else is as it happened:

Claude Code uploads foto.png with artifact-cli, looks up the API, validates and runs the script on the team server, and downloads the 34,628-byte WebP — 21 seconds

Nine turns, 21 seconds, and the script ran on the first attempt: the model checked it with validate_script before calling run_script. run_script reports width and height of every output, so there was no second call just to measure the result.

The conversion itself is one line of OpticScript. Here with the eagle from the samples:

Converting on the server: PNG 1024×1024, 1,056,527 bytes, becomes WebP 640×640, 26,760 bytes — 97 percent smaller

Data from elsewhere

The store is not only for images. A database tool can drop its query result as CSV, and a script gets the id as a parameter and turns it into a chart:

//!PARAM: DATA:string=
//!OUTPUT: CHART
const rows = Engine.parseCSV("mlcartifact://" + DATA);
// … draw the chart …
chart.save(CHART);        // → mlcartifact://<new id> in the result

Every loader reads from the store — Engine.loadImage, parseCSV, parseJSON, readText, loadSVG. Engine.writeArtifact(name, imageOrText) stores a result and returns its id.

Setting it up

A team server needs an Enterprise licence — the 30-day evaluation unlocks it too. On the server, as the account the services run under:

mlcos-license activate <KEY> --name "First Last"
mlcos-license status        # what is unlocked, until when

Without a matching licence mlcos-mcp refuses to start on the network and names exactly this command. Then:

# the store — token required, even from the same machine
ARTIFACT_GRPC_TOKEN=<store-token> artifact-server \
    -grpc-addr 0.0.0.0:9590 -addr 127.0.0.1:9591 -require-token-localhost

# the engine
MLCOS_MCP_TOKEN=<secret> MLCOS_ARTIFACT_TOKEN=<store-token> \
  mlcos-mcp -addr 0.0.0.0:8765 \
    -artifacts http://127.0.0.1:9590 -artifact-user team

On each workstation:

claude mcp add --scope user --transport http opticscript \
    http://<server>:8765/mcp --header "Authorization: Bearer <secret>"

For artifact-cli, set ARTIFACT_GRPC_ADDR=<server>:9590, ARTIFACT_GRPC_TOKEN=<store-token> and ARTIFACT_USER_ID=team. Then artifact-cli create photo.png prints the id, and artifact-cli download <id> result.webp fetches a result.

On macOS both run fine as LaunchAgents, on Linux as systemd units.

Safe by default

Whoever reaches the server sends code. So:

  • A token is required as soon as the server listens beyond loopback; without one it refuses to start. -addr :8765 without a host means 127.0.0.1.
  • Scripts run in a sandbox: no file paths — neither from the caller nor inside the script. Inputs, outputs and artifacts only. Engine.env() returns nothing, Engine.fetch stays off, tool plugins only with -allow-tools.
  • Limits: 5 minutes per run, 64 MB per request.
  • The store wants its own token and rejects browser requests from foreign origins.

Limits

  • One area for everyone. All users of a server share the artifact area (-artifact-user). Separation per person needs sign-in (OAuth), which this version does not have.
  • claude.ai in the browser cannot use it yet. Its connectors reach servers from the internet and expect OAuth; a server on your LAN with a fixed token is neither.
  • Large files go through the store. Base64 works for icons, not for photos.

See also