MCP + Semantic Recall

Your AI memory, inside every coding agent you already use.

dropdat now ships an MCP server. Wire it into Claude Code, Cursor, Cline, or Claude Desktop — and every session can recall, read, and save capsules across providers without copy-pasting context.

Five tools. One library.

The MCP server exposes exactly the surface an agent needs — and nothing it doesn't.

dropdat_recall

Hybrid semantic + keyword search across every capsule the user has saved. Vector cosine over OpenAI embeddings, fused with BM25 ranking via RRF. The agent can ask in any phrasing — matches surface even when the original capsule used different words.

dropdat_read

Fetch one capsule's full contents (every message, every version) by id. Pair with dropdat_recall to drill into a hit.

dropdat_list

Browse recent capsules, optionally filtered by tag. Use when the user wants to skim, not search.

dropdat_capsule

Save the current conversation as a new capsule. Call when the user says 'remember this' or at the end of a meaningful session.

dropdat_autocapsule

Save the FULL verbatim Claude Code session by reading the on-disk .jsonl transcript — no model reconstruction, every message preserved.

Powered by hybrid search

Recall by meaning, not just by word.

Every capsule is embedded automatically on save. When an agent calls dropdat_recall, the query embeds, runs vector + BM25 in one SQL pass, and returns ranked hits — even if the wording is entirely different.

Hybrid retrieval

Vector cosine + tsvector BM25, fused with Reciprocal Rank Fusion. Catches semantic matches and exact keywords in one query.

pgvector + HNSW

Embeddings live in your own Postgres. HNSW index over 1536-dim vectors keeps recall fast even at hundreds of thousands of capsules.

Auto-embedded on write

New capsules and edits embed in the background. No worker queue to operate, no separate vector DB to babysit.

Graceful fallback

No OPENAI_API_KEY set? Search degrades to BM25 only — your library stays searchable, just less smart.

Install in under a minute.

One npm package, one config block, any MCP-capable client. Source on GitHub.

  1. 01

    Issue an API key

    Open the API Keys page in your dashboard and create a key. Tokens are shown once and look like dk_live_….
  2. 02

    Install from npm

    # zero-install — runs the latest version on demand
    npx -y @dropdat/mcp
    
    # or pin a global install
    npm install -g @dropdat/mcp
  3. 03

    Wire it into your client

    Claude Code (one-liner)
    claude mcp add dropdat \
      --env DROPDAT_API_KEY=dk_live_xxx \
      --env DROPDAT_API_BASE=https://dropdat.app \
      -- npx -y @dropdat/mcp
    Then /mcp inside Claude Code lists the new server and its five tools.
    Cursor, Cline, Claude Desktop, Continue
    {
      "mcpServers": {
        "dropdat": {
          "command": "npx",
          "args": ["-y", "@dropdat/mcp"],
          "env": {
            "DROPDAT_API_KEY": "dk_live_xxx",
            "DROPDAT_API_BASE": "https://dropdat.app"
          }
        }
      }
    }
    Drop into the client’s MCP config file (e.g. ~/.cursor/mcp.json or claude_desktop_config.json) and restart the client.
  4. 04

    Talk to it

    • “What did we decide about auth last week? Check dropdat.”
    • “Save this conversation as a dropdat capsule titled ‘billing migration notes’.”
    • “List my recent dropdat capsules tagged ‘refactor’.”

Per-client install guides

Pick your AI client for the exact config snippet, file path, and verification step.

Also works with

Standard MCP stdio transport. No fork, no proxy, no special build.

Claude Code

Anthropic's terminal coding agent

Cursor

Editor with native MCP support

Cline

VS Code autonomous coding agent

Claude Desktop

Mac & Windows chat app

Continue

Open-source VS Code / JetBrains extension

Anything MCP-capable

Stdio transport, standard spec

Stop pasting context. Start recalling it.

The MCP server is open and free with any dropdat account. Semantic recall ships on every tier.