AI chat participant for Etherpad. When users @mention the AI in pad chat, it reads the document, understands who wrote what, and responds. It can also edit the pad directly when asked.

pnpm run plugins i ep_ai_chat

This pulls in ep_ai_core automatically — it is a required runtime dependency.

Configure the LLM provider in settings.json under the ep_ai_core key. See the ep_ai_core README for full configuration details.

Type @ai followed by your message in the pad's chat box:

@ai summarize this document
@ai who wrote the introduction?
@ai fix the spelling errors in paragraph 3

The AI will respond in chat. If you ask it to make changes and the pad has full access mode, it will edit the document directly.

Chat-specific settings go under ep_ai_core.chat in settings.json. The apiBaseUrl, apiKey, model, and provider fields belong directly under ep_ai_core and tell the plugin which LLM to talk to. See "LLM provider examples" below for the most common providers.

{
  "ep_ai_core": {
    "apiBaseUrl": "https://api.anthropic.com/v1",
    "apiKey": "sk-ant-...",
    "model": "claude-sonnet-4-20250514",
    "chat": {
      "trigger": "@ai",
      "authorName": "AI Assistant",
      "authorColor": "#7c4dff",
      "systemPrompt": "You are a helpful writing assistant.",
      "maxContextChars": 50000,
      "chatHistoryLength": 20,
      "conversationBufferSize": 10
    }
  }
}

The plugin speaks two protocols: Anthropic's /messages and the OpenAI /chat/completions shape. Anything that exposes one of those works. The provider is auto-detected from the URL; set "provider": "anthropic" or "provider": "openai" to force it.

{
  "ep_ai_core": {
    "apiBaseUrl": "https://api.anthropic.com/v1",
    "apiKey": "sk-ant-...",
    "model": "claude-sonnet-4-20250514",
    "provider": "anthropic"
  }
}

{
  "ep_ai_core": {
    "apiBaseUrl": "https://api.openai.com/v1",
    "apiKey": "sk-...",
    "model": "gpt-4o",
    "provider": "openai"
  }
}

Gemini exposes an OpenAI-compatible endpoint, so the openai provider works straight out of the box.

{
  "ep_ai_core": {
    "apiBaseUrl": "https://generativelanguage.googleapis.com/v1beta/openai",
    "apiKey": "AIza...",
    "model": "gemini-2.5-flash",
    "provider": "openai"
  }
}

The GitHub Models inference API is OpenAI-compatible. The apiKey is a GitHub personal access token with the models:read scope.

{
  "ep_ai_core": {
    "apiBaseUrl": "https://models.github.ai/inference",
    "apiKey": "ghp_...",
    "model": "openai/gpt-4o",
    "provider": "openai"
  }
}

{
  "ep_ai_core": {
    "apiBaseUrl": "https://api.x.ai/v1",
    "apiKey": "xai-...",
    "model": "grok-2-latest",
    "provider": "openai"
  }
}

Ollama exposes an OpenAI-compatible endpoint at /v1, so it slots in like any other provider. No API key is required — Ollama ignores the Authorization header — but apiKey must still be set to a non-empty string (any value works).

Pull a model first, then point Etherpad at the local server:

ollama pull gemma3
# or: llama3.1, llama3.2, qwen2.5, mistral, deepseek-r1, ...

{
  "ep_ai_core": {
    "apiBaseUrl": "http://localhost:11434/v1",
    "apiKey": "ollama",
    "model": "gemma3:latest",
    "provider": "openai"
  }
}

The model string is whatever ollama list shows (including the tag, e.g. gemma3:latest or llama3.1:8b). Anything Ollama can serve will work for plain chat replies.

Note on edits with smaller models. When a user asks the AI to edit the pad, the plugin asks the model to return a fenced ```json block of the form {"action":"edit","findText":"...","replaceText":"..."}. Larger frontier models follow this reliably; smaller local models (e.g. 3–4B parameter quantised builds) sometimes wander off-format, in which case the response is sent to chat as-is instead of being applied as an edit. Reach for an 8B+ instruction-tuned model (llama3.1:8b, qwen2.5:14b, etc.) for the most reliable in-pad editing; smaller models are fine for Q&A, summarisation, and discussion.

If Etherpad runs in a container, replace localhost with the host reachable from inside the container — typically http://host.docker.internal:11434/v1 on Docker Desktop, or the host's LAN IP elsewhere. By default Ollama binds to 127.0.0.1; export OLLAMA_HOST=0.0.0.0 to bind all interfaces if Etherpad runs on a different machine.

Tip: any other provider that ships an OpenAI-compatible endpoint (Mistral, Groq, OpenRouter, LM Studio, vLLM, an internal gateway, etc.) works the same way — point apiBaseUrl at it, set "provider": "openai", pick the right model string for that provider, and you're done.

When ep_comments_page is also installed, the AI defaults to creating a suggestion comment instead of editing the pad directly. The comment shows up in the comments sidebar with Accept and Revert controls — the document author reviews each change before it lands.

If ep_comments_page is not installed, the AI falls back to direct editing exactly as before.

You can override the configured mode for a single chat message:

@ai apply: fix the typo in line 3
@ai suggest: rewrite the introduction

@ai apply: always edits the pad directly. @ai suggest: always creates a suggestion comment (or, if ep_comments_page is missing, replies in chat explaining the missing dep and applies directly).

Add to ep_ai_core.chat in settings.json:

{
  "ep_ai_core": {
    "chat": {
      "suggestionMode": "auto",
      "suggestionModePads": { "important-pad": "suggest" }
    }
  }
}

The resolution cascade (highest priority wins): per-request override → per-pad → global → built-in default.

When a user asks the AI to change the document, the AI responds with a structured edit (find text, replace with new text). The plugin:

1. Locates the exact text in the pad
2. Applies the change as a native Etherpad changeset
3. Attributes the edit to the AI author (with the configured color)
4. Broadcasts the update to all connected clients
5. Announces the AI in the user list so its color appears in the author palette

If the pad's access mode is readOnly, the AI will answer questions but decline edit requests. If access is none, the AI will not respond at all.

The AI maintains a short conversation buffer per pad. Follow-up messages can reference earlier exchanges without repeating context. The buffer is kept in memory and resets when the server restarts.

Apache-2.0