imagin-studio-api-docs-mcp 0.1.14

Local MCP server for IMAGIN.studio documentation search

IMAGIN.studio API Docs MCP Server - Setup Guide

Works on Windows, macOS, and Linux. No cloning, no venv — just install uv and paste a config snippet.

The server provides one tool — search_docs — that gives your AI coding assistant semantic search over the IMAGIN.studio documentation.

Dependency note — fastembed-imagin

This package depends on fastembed-imagin, a temporary fork of fastembed by Qdrant.

Why: upstream fastembed 0.7.4 pins pillow<12.0, which blocks Pillow 12.x security fixes (CVE-2026-25990). The fix is merged on main but not yet released.

What changed: only the pillow version constraint — relaxed from <12.0 to <13.0 for Python 3.10+. No logic changes. Source: IMAGIN-studio/fastembed-imagin.

pip-audit: because fastembed-imagin is not tracked by any security advisory database, pip-audit will report no known vulnerabilities for it (it passes clean). It will not flag it as an unknown or suspicious package.

Revert plan: once upstream releases fastembed 0.7.5+ with the pillow fix, this package will switch back to fastembed and fastembed-imagin will be retired.

---

Quick start

If your AI agent supports MCP, you can try pasting this into the chat:

Read the setup instructions on this page first, then install this MCP server: https://pypi.org/project/imagin-studio-api-docs-mcp/

Agent: when the user sends you this link, visit the URL and read the full setup instructions on that page. It contains step-by-step configuration for every major AI coding assistant.

Most agents (Claude Code, Cursor, etc.) will walk you through the setup. For full manual instructions, see below.

First-time install note: The very first time the MCP server starts, it downloads dependencies and builds the index. Most MCP clients will time out waiting for this and report a connection failure. This is expected — just type /mcp (or use your agent's MCP reconnect command) to reconnect, and it will work. Subsequent starts are near-instant.

---

1. Install uv

uv is a fast Python package manager. Install it once:

macOS / Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Or via Homebrew:

brew install uv

Windows (PowerShell):

irm https://astral.sh/uv/install.ps1 | iex

Or via a package manager:

winget install --id=astral-sh.uv -e
# or
scoop install main/uv

Verify:

uv --version

To update uv later: uv self update

---

2. Verify it works

Run the server manually to confirm everything installs and starts:

uvx imagin-studio-api-docs-mcp

On first run this will:

1. Download Python (if needed) and the imagin-studio-api-docs-mcp package
2. Clone the IMAGIN.studio docs from GitHub
3. Build a local vector index (~20-30 seconds)
4. Start listening on stdio (you'll see no output — that's normal)

Press Ctrl+C to stop. The index is cached for next time.

---

3. Alternative: pip install

If you prefer a traditional Python install over uvx, you can install the package directly with pip:

pip install imagin-studio-api-docs-mcp

This installs a console script called imagin-studio-api-docs-mcp that you can run directly:

imagin-studio-api-docs-mcp

Tip: Use a virtual environment to avoid polluting your global Python:

python3 -m venv .venv
source .venv/bin/activate   # macOS / Linux
.venv\Scripts\activate      # Windows
pip install imagin-studio-api-docs-mcp

When using pip install, replace "uvx" with "imagin-studio-api-docs-mcp" and remove "args" in the agent config snippets below. For example:

"imagin-docs": {
  "command": "imagin-studio-api-docs-mcp",
  "args": []
}

If your agent can't find the command, use the full path to the installed script:

# Find it with:
which imagin-studio-api-docs-mcp          # macOS / Linux
where imagin-studio-api-docs-mcp          # Windows

Then use that path as "command" in your agent config.

uvx vs pip: uvx is the recommended method because it handles Python version management, dependency isolation, and auto-updates. With pip you manage the environment yourself, but it works just as well once installed.

---

4. Configure your agent

Pick your agent below and paste the config. uvx handles Python, dependencies, and isolation automatically. If you installed via pip instead, replace "uvx" with "imagin-studio-api-docs-mcp" and set "args": [] (see section 3).

Server config reference

Most agents work with just "uvx" as the command. If your agent can't find uvx, replace "uvx" with the full path (see troubleshooting).

The one exception is Claude Desktop App, which requires the full path upfront.

"command": "uvx",
"args":    ["imagin-studio-api-docs-mcp"]

Alternative: npx (Node.js)

If you have Node.js 18+ installed, you can use npx instead of uvx:

npx -y imagin-studio-api-docs-mcp

This downloads the npm wrapper and starts the MCP server. On first run it will also download the Python package via uvx behind the scenes.

Prerequisite: you still need uv installed (see step 1). The npx package is a thin wrapper that finds and launches uvx for you — it does not bundle Python.

Why use npx? GUI apps like Claude Desktop and Cursor often can't find uvx because they don't inherit your shell PATH. The npx wrapper probes common install locations (~/.local/bin, Homebrew, cargo) automatically, so you don't need to configure full paths.

In agent config snippets, replace "uvx" with "npx" and use ["-y", "imagin-studio-api-docs-mcp"] as args:

"imagin-docs": {
  "command": "npx",
  "args": ["-y", "imagin-studio-api-docs-mcp"]
}

npx vs uvx: uvx is still the recommended method — it's faster and has no Node.js dependency. Use npx when your agent has trouble finding uvx in PATH (common with GUI-launched apps).

---

4A. Claude Code (CLI)

Claude Code runs in your terminal, so it inherits your shell PATH. You can use just "uvx" here (no full path needed).

1. Open a terminal
2. Navigate to your project directory (or any directory you want to use Claude Code in)
3. Create a file called .mcp.json in that directory with this content:

{
  "mcpServers": {
    "imagin-docs": {
      "type": "stdio",
      "command": "uvx",
      "args": ["imagin-studio-api-docs-mcp"]
    }
  }
}

4. From that same directory, run: claude
5. Claude Code will detect the .mcp.json and start the MCP server. You should see imagin-docs listed when it connects.
6. Test it by asking: "Search the IMAGIN docs for CDN cache invalidation"

To debug connection issues, run: claude --mcp-debug

---

4B. Claude Desktop App

If you already have MCP servers installed via Desktop Extensions (Settings > Extensions), those are managed separately by Claude and will keep working. The steps below add a stdio server via the config file, which is the standard way to add custom/local MCP servers.

1. Open Claude Desktop App

2. Open Settings:

   • macOS: Claude (menu bar) > Settings... — or Cmd + ,
   • Windows: File > Settings... — or Ctrl + ,

3. Click "Developer" in the left sidebar

4. Click "Edit Config" — this opens the config file in your default editor:

   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
   • Windows: %APPDATA%\Claude\claude_desktop_config.json
   • Linux: ~/.config/Claude/claude_desktop_config.json

5. Add the "mcpServers" key to the file. Your file probably already has some content (like "preferences"). Add "mcpServers" alongside it:

   Before (example):

   {
     "preferences": {
       "coworkScheduledTasksEnabled": false,
       "sidebarMode": "chat"
     }
   }
   

   After (replace <UVX_PATH> with your full path from which uvx):

   {
     "preferences": {
       "coworkScheduledTasksEnabled": false,
       "sidebarMode": "chat"
     },
     "mcpServers": {
       "imagin-docs": {
         "command": "<UVX_PATH>",
         "args": ["imagin-studio-api-docs-mcp"]
       }
     }
   }
   

   Example full paths:

   • macOS: /Users/you/.local/bin/uvx
   • Linux: /home/you/.local/bin/uvx
   • Windows: C:/Users/you/.local/bin/uvx.exe

   Important: Claude Desktop App does not inherit your shell PATH, so just "uvx" will fail with "No such file or directory". You must use the full absolute path.

   Key things to get right:

   • Use the full path to uvx (not just "uvx")
   • Add a comma after the closing } of "preferences" (before "mcpServers")
   • Keep your existing keys intact — only add the "mcpServers" block
   • If "mcpServers" already exists, just add the "imagin-docs" entry inside it

6. Save the file

7. Fully quit Claude Desktop App (not just close the window):

   • macOS: Claude (menu bar) > Quit Claude — or Cmd + Q
   • Windows: right-click the Claude icon in the system tray > Quit

8. Reopen Claude Desktop App

9. Start a new chat. You should see a slider/hammer icon at the bottom of the input box. Click it — search_docs should be listed.

If the icon doesn't appear, check the logs for errors:

• macOS: ~/Library/Logs/Claude/mcp*.log
• Windows: %APPDATA%\Claude\logs\mcp*.log

---

4C. Cursor

1. Open Cursor

2. Open Settings:

   • macOS: Cursor (menu bar) > Settings... > Cursor Settings
   • Windows: File > Preferences > Cursor Settings
   • Or press Cmd + Shift + J (macOS) / Ctrl + Shift + J (Windows)

3. Click "Tools & MCP" in the left sidebar

4. Click "+ New MCP Server" > "Add a Custom MCP Server"

5. Cursor will open a JSON file. Paste this inside the "mcpServers" object:

   "imagin-docs": {
     "command": "uvx",
     "args": ["imagin-studio-api-docs-mcp"]
   }
   

6. Save the file

7. Back in the MCP settings page, you should see imagin-docs listed. The dot next to it should turn green (connected). If it shows red, toggle the server off and on again using the switch. If it stays red, try replacing "uvx" with the full path (see troubleshooting).

Alternative — Config file (manual):

• Global: ~/.cursor/mcp.json
• Per-project: .cursor/mcp.json (in your project root)

1. Create or open the file
2. Paste:

{
  "mcpServers": {
    "imagin-docs": {
      "command": "uvx",
      "args": ["imagin-studio-api-docs-mcp"]
    }
  }
}

3. Save the file
4. Restart Cursor (Cmd + Q / Alt + F4, then reopen)
5. Verify in Settings > Tools & MCP — imagin-docs should show a green dot

---

4D. Windsurf (Codeium)

1. Open Windsurf
2. Go to Windsurf Settings > Cascade > MCP Servers
3. Click "Open MCP Marketplace", then click the gear icon next to "Installed MCPs" — this opens mcp_config.json:
   • macOS: ~/.codeium/windsurf/mcp_config.json
   • Windows: %USERPROFILE%\.codeium\windsurf\mcp_config.json
   • Linux: ~/.codeium/windsurf/mcp_config.json
4. Add the "imagin-docs" server inside the "mcpServers" object:

{
  "mcpServers": {
    "imagin-docs": {
      "command": "uvx",
      "args": ["imagin-studio-api-docs-mcp"]
    }
  }
}

If the file already has other servers, just add the "imagin-docs" entry inside the existing "mcpServers" object. If Windsurf can't find uvx, replace "uvx" with the full path (see troubleshooting).

5. Save the file
6. Back in MCP Servers, imagin-docs should appear in the installed list with a green status indicator. The search_docs tool should be listed under it.

---

4E. VS Code + Copilot (Agent Mode)

Requires VS Code 1.99+ and the GitHub Copilot extension.

Option 1 — Per-project config (recommended):

1. Open your project in VS Code
2. Create a .vscode folder in the project root if it doesn't exist
3. Create a file called .vscode/mcp.json with this content:

{
  "servers": {
    "imagin-docs": {
      "type": "stdio",
      "command": "uvx",
      "args": ["imagin-studio-api-docs-mcp"]
    }
  }
}

4. Open Copilot Chat: Cmd + Shift + I (macOS) / Ctrl + Shift + I (Windows)
5. Switch to Agent mode by clicking the mode dropdown at the top of the chat panel and selecting "Agent"
6. Click the tools icon (wrench) in the chat input — search_docs should be listed under imagin-docs

Option 2 — Global config (available in all projects):

1. Open VS Code Settings: Cmd + , (macOS) / Ctrl + , (Windows)
2. Click the {} icon (top-right) to open settings as JSON
3. Add this block to your settings JSON:

"mcp": {
  "servers": {
    "imagin-docs": {
      "type": "stdio",
      "command": "uvx",
      "args": ["imagin-studio-api-docs-mcp"]
    }
  }
}

4. Save and reload VS Code: Cmd + Shift + P > "Developer: Reload Window"
5. Verify in Copilot Chat (Agent mode) — the tool should appear

If VS Code can't find uvx, replace "uvx" with the full path (see troubleshooting).

Note: VS Code uses "servers" (not "mcpServers") as the top-level key in .vscode/mcp.json. This is different from most other agents.

---

4F. Cline (VS Code Extension)

Option 1 — Via the Cline UI (recommended):

1. Open VS Code with the Cline extension installed

2. Open the Cline sidebar (click the Cline icon in the activity bar)

3. Click the MCP servers icon (server/plug icon at the top of the Cline panel)

4. Click "Configure MCP Servers" — this opens the settings JSON file

5. Add the "imagin-docs" server inside the "mcpServers" object:

   "imagin-docs": {
     "command": "uvx",
     "args": ["imagin-studio-api-docs-mcp"]
   }
   

6. Save the file

7. Back in the MCP servers panel, imagin-docs should appear. Click the refresh icon if it doesn't connect immediately.

8. The green dot means connected. Click on the server name to see the search_docs tool listed.

Option 2 — Edit the config file directly:

The file location depends on your OS:

• macOS: ~/Library/Application Support/Code/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json
• Windows: %APPDATA%\Code\User\globalStorage\saoudrizwan.claude-dev\settings\cline_mcp_settings.json
• Linux: ~/.config/Code/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json

1. Open the file in any text editor
2. Add or replace with:

{
  "mcpServers": {
    "imagin-docs": {
      "command": "uvx",
      "args": ["imagin-studio-api-docs-mcp"]
    }
  }
}

3. Save the file
4. In VS Code, open the Cline sidebar and click the refresh icon on the MCP servers panel

If Cline can't find uvx, replace "uvx" with the full path (see troubleshooting).

---

4G. Zed

Zed uses a different config format than the other agents.

1. Open Zed

2. Open Settings:

   • macOS: Zed (menu bar) > Settings... — or Cmd + ,
   • Linux: File > Settings — or Ctrl + ,

   This opens ~/.config/zed/settings.json

3. Find or create a "context_servers" key at the top level of the JSON. Add the "imagin-docs" entry inside it:

   "context_servers": {
     "imagin-docs": {
       "command": {
         "path": "uvx",
         "args": ["imagin-studio-api-docs-mcp"],
         "env": {}
       },
       "settings": {}
     }
   }
   

   If Zed can't find uvx, replace "uvx" with the full path from which uvx (see troubleshooting).

   If you already have a "context_servers" block with other servers, just add the "imagin-docs": { ... } entry inside it.

4. Save the file (Zed auto-reloads settings — no restart needed)

5. Open the AI assistant panel: Cmd + Shift + A (macOS) / Ctrl + Shift + A

6. The search_docs tool should now be available to the assistant

Note: Zed's config format is different from all other agents:

• Uses "context_servers" instead of "mcpServers"
• Uses "path" instead of "command"
• Has a nested "command" object wrapping "path" and "args"

---

5. First run

On first launch the server will:

1. Clone the IMAGIN.studio public docs from GitHub (~5 sec)
2. Parse and chunk the markdown files
3. Download the embedding model (~50 MB, once)
4. Build a local vector index (~20-30 sec)

The first launch will likely fail. Because uvx downloads Python, the package, and the embedding model on the first run, the MCP client usually times out before the server is ready. This is normal. Type /mcp in your agent (or use its MCP reconnect/restart command) to reconnect. The second connection will succeed because everything is already cached.

During indexing, search_docs returns: "Index is not yet built... Please retry shortly."

After that, searches are instant. The index is cached and only rebuilds when the upstream documentation changes.

---

6. Data storage

All data lives under your home directory:

	Path
macOS / Linux	~/.imagin-docs-mcp/
Windows	C:\Users\<YOU>\.imagin-docs-mcp\

docs/           Cloned documentation repository
lancedb/        Vector search index
models/         Cached embedding model
metadata.json   Index state (commit hash, timestamps)

To fully reset, delete that folder and restart your agent:

# macOS / Linux
rm -rf ~/.imagin-docs-mcp

# Windows CMD
rmdir /s /q %USERPROFILE%\.imagin-docs-mcp

# Windows PowerShell
Remove-Item -Recurse -Force $env:USERPROFILE\.imagin-docs-mcp

---

7. Telemetry (opt-in)

By default, the MCP server sends no data anywhere. All processing is fully local.

If you'd like to help improve the server, you can opt in to anonymous crash reporting:

export IMAGIN_DOCS_TELEMETRY=1

When enabled, only unhandled exceptions (stack traces) are sent to Google Cloud Error Reporting. This helps us detect and fix startup failures, indexing errors, and compatibility issues.

What is collected:

• Exception type and stack trace
• Package version and OS platform
• No search queries, no document content, no personally identifiable information

What is NOT collected:

• Your search queries or results
• Document content from the index
• IP addresses, usernames, or machine identifiers

To disable, unset the variable or set it to 0:

unset IMAGIN_DOCS_TELEMETRY

---

8. Troubleshooting

"uvx" not found / "No such file or directory" / "Failed to spawn"

• Replace "uvx" in your config with the full absolute path:
  • macOS / Linux: run which uvx (typically ~/.local/bin/uvx)
  • Windows: run where uvx (typically C:\Users\<you>\.local\bin\uvx.exe)
• Claude Desktop App always requires the full path.
• If uvx isn't installed at all, re-run the uv installer from step 1.
• Windows: restart your terminal after installing so PATH updates.
• Verify with: uv --version

"Index is not yet built..."

• Normal on first start. Wait ~30 seconds and retry.
• Ensure you have internet (the server clones docs from GitHub).
• Some agents may retry too quickly before the index finishes building. If you see repeated "not yet built" messages, wait for the initial indexing to complete (~20-30 sec) before sending another query.

MCP server fails on first install

• This is expected. The first run downloads Python, dependencies, and the embedding model, which takes longer than most MCP clients will wait.
• Type /mcp in your agent (or use its reconnect/restart MCP command) to reconnect. The second attempt will work because everything is cached.
• In Claude Code: type /mcp, then reconnect the imagin-docs server.
• In Cursor: toggle the server off and on in Settings > Tools & MCP.
• In other agents: restart the agent or IDE.

Agent doesn't see the MCP server

• Double-check the config file path for your agent (section 4).
• Restart the agent / IDE after saving.
• Claude Code CLI: run claude --mcp-debug for protocol logs.

Slow first startup

• The embedding model (~50 MB) downloads once on first run.
• Subsequent starts reuse the cached model and are near-instant.

Windows long path errors

• Enable long paths: reg add HKLM\SYSTEM\CurrentControlSet\Control\FileSystem /v LongPathsEnabled /t REG_DWORD /d 1 /f
• Restart your terminal.

PowerShell blocks the uv install script

• Run: Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned
• Then re-run the install command.
• Alternatively, run with: powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"