Last updated

iblai-api-agent-chat

Mirrored from iblai/base · skills/agents/iblai-api-agent-chat/SKILL.md. This page is generated — edit it in the repository, not here.

Enable live chat with a deployed agent. Unlike the other iblai-* skills (which call the REST API directly), runtime chat (the agent's tools, RAG and memory) is served by the hosted iblai-api-agent-chat MCP server. This skill does the one-time wiring so your assistant can talk to an agent; the conversation itself then runs through that server's MCP tools.

When to use

  • You want to converse with an agent (not configure it). For configuring agents use the other /iblai-api-agent-* skills; to find an agent's unique_id use /iblai-api-search.

What you need

  • IBLAI_API_KEY — your Platform API Token (from .env; run /iblai-api-login if it is not set).
  • An agent's unique_id — the agent to chat with. Get one from /iblai-api-search (each result's unique_id), then confirm the choice with the user.

Setup

The agent's unique_id goes in the URL, and the server authenticates with one header, Authorization: Api-Token <key>:

https://asgi.data.iblai.app/mcp/agent-chat/<unique_id>

No agent header is needed. (The older X-Mentor-Unique-Id / X-Agent-Unique-Id header still works for existing configs; a URL id and a header naming different agents are refused.)

Claude Code

claude mcp add iblai-api-agent-chat \
  --transport http "https://asgi.data.iblai.app/mcp/agent-chat/$MENTOR" \
  --header "Authorization: Api-Token $IBLAI_API_KEY"

.mcp.json / Claude Desktop / Cursor

Add to the project's .mcp.json (or the client's MCP config). Substitute the real token and unique_id - do not commit secrets:

{
  "mcpServers": {
    "iblai-agent-chat": {
      "transport": "streamable-http",
      "url": "https://asgi.data.iblai.app/mcp/agent-chat/YOUR_AGENT_UNIQUE_ID",
      "headers": {
        "Authorization": "Api-Token YOUR_API_TOKEN"
      }
    }
  }
}

After adding it, the assistant gains the server's get_mentor_response(prompt, session_id?) tool. Each call is one chat turn: the agent replies as it does in the chat screen, with its own instructions, tools and trained documents. The tool returns the full reply text, ending with a session_id: <id> line (also in the result's structured content).

Notes

  • One agent per connection - the agent is fixed by the URL of each server entry. To chat with several agents, add multiple named entries (e.g. iblai-api-agent-chat-tutor, iblai-api-agent-chat-advisor), each with its own unique_id in the URL.
  • Follow-up questions - pass the session_id from the previous reply to continue the conversation; the agent sees the earlier turns. Omit it to start a new conversation. A session belongs to the token's owner and its agent; a session id of another user, organization or agent is refused.
  • Owner-mode tokens only - the turn is attributed to the token's owner and billed to the token's organization like a turn in the chat screen (credits, spend caps and moderation apply). A refused turn (out of credits, a non-owner-mode token, no access to the agent) comes back as a tool error with the reason.
  • Which agents - your organization's agents, plus main agents shared with every organization (visibility "viewable by anyone"). Those are chattable with any organization's token, and the turn is billed to the caller's organization.
  • The token lives in client config / .env - never commit it; .env and local MCP config should stay gitignored.
  • This is the only runtime surface among the iblai-api-* skills; everything else (configure, search, recommend, profile, analytics, admin) is a direct-REST skill.

Copyright © ibl.ai | support@iblai.zendesk.com