Skip to main content
Glama
smartcomputer-ai

imaginator

by smartcomputer-ai
OverviewSchemaSourceRelated ServersDiscussionsAdmin
TypeScript
Remote

Imaginator

Run image generation experiments across different models and providers.

Work in a grid: prompts are rows, models are columns. Think, "Excel for image prompts". Any changes to a row or column immediately regenerates the relevant cells in parallel.

Designed to be used by agents via an MCP backend for rapid experimentation.

Imaginator grid

Quick start

corepack enable            # pnpm via corepack
pnpm install
cp .env.example .env       # then add your provider API keys
pnpm dev                   # server on :4747, web on :5173

Open http://localhost:5173.

Providers are enabled by the keys in .env (OPENAI_API_KEY, FAL_KEY, BFL_API_KEY, GOOGLE_API_KEY, REPLICATE_API_TOKEN). To try it without spending anything, leave the keys empty or set IMAGINATOR_MOCK=1: the mock/* models render prompts onto colored images and exercise the whole workflow.

Related MCP server: imagen-mcp

Connect an agent (MCP)

The server hosts an MCP endpoint at http://127.0.0.1:4747/mcp (Streamable HTTP, both the 2026-07-28 and the 2025-era protocol). Claude Code:

claude mcp add --transport http imaginator http://127.0.0.1:4747/mcp

Clients that only launch local stdio servers (Claude Desktop, most config files) use the bridge, which forwards to the running server:

{ "mcpServers": { "imaginator": { "command": "pnpm", "args": ["--dir", "/path/to/imaginator", "mcp"] } } }

The agent gets tools to create collections, add rows and columns, wait for generations, and look at the results as images. Try the compare-models prompt: give it a few prompts and it sets up the grid, waits, and writes a comparison. Details, including what different MCP clients can and cannot do with images, are in packages/server/README.md and docs/MCP-CLIENTS.md.

Authenticated mode

By default the server is an open localhost tool. To expose it, set in .env:

AUTH_ENABLED=1
AUTH_PASSWORD=...        # web UI login
AUTH_API_KEY=...         # MCP clients and scripts: Authorization: Bearer <key>
IMAGINATOR_HOST=0.0.0.0  # optional: listen on all interfaces

The web app then shows a login page and keeps a session cookie (30 days, AUTH_SESSION_DAYS to change). MCP and /api calls need the bearer key:

claude mcp add --transport http imaginator http://host:4747/mcp --header "Authorization: Bearer $AUTH_API_KEY"
curl -H "Authorization: Bearer $AUTH_API_KEY" http://host:4747/api/collections.list

The stdio bridge (pnpm mcp) picks AUTH_API_KEY up from the environment or .env on its own. Only /api/health, the login routes and the static web app stay open. Put TLS in front (a reverse proxy) when the host is not on a trusted network; the cookie is marked Secure when the request arrives over HTTPS or with X-Forwarded-Proto: https.

How it works

  • Nothing is imperative. You never press "generate". Adding a row or a column to a live collection is the request; the server fills every cell that is missing and leaves alone every cell whose content did not change.

  • Edits are cheap and precise. Change one prompt and only that row regenerates. Change a column's model settings and only that column does. Revert an edit and the earlier result comes back without a new run, because results are keyed by what was asked, not by when.

  • Rows own the prompt, columns own the model. Rows carry the prompt, optional input images (reference, init, mask) and common settings such as aspect ratio or seed. Columns carry the model and its provider-specific settings. That split keeps every cell in a column comparable.

  • Pause to plan, resume to run. A paused collection can be edited freely; resuming runs one wave of generations for everything that is missing.

  • History is kept. Every generation stays as a version of its cell. "Regenerate" asks the same request for another sample; "retry" re-runs a failed one. Failed cells never retry on their own, so a broken prompt cannot burn money.

  • Assets are shared. Uploaded images and generated outputs live in one library and can be used as inputs in any row of any collection.

Development

pnpm typecheck        # all packages
pnpm test             # all packages
pnpm dev:server       # server only (tsx watch)
pnpm dev:web          # web only
pnpm mcp              # stdio bridge to the running server's MCP endpoint (/mcp)

Package

What

packages/core

Domain types, zod schemas, IDs and addresses, request resolution and hashing, Provider interface, command IO schemas. Browser-safe, no I/O.

packages/server

SQLite persistence, asset store, services and event bus, reconciler, runner, provider adapters, HTTP + SSE, MCP.

packages/web

Vite + React grid UI.

Every operation is also a plain POST /api/<command> with a JSON body; GET /api lists them all with their schemas.

License

Apache 2.0, see LICENSE.

This server cannot be deployed

One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

Maintenance

ActivityMaintained
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers