MCP Integration

What Your Agent Can Do

Seven tools, one persistent knowledge layer. Your agent reads and writes knowledge using the standard Model Context Protocol.

get_instructions

Returns organization guide, best practices, and accessible harbors. Call this first before other tools.

search

Semantic search across all pearls. Returns titles, summaries, and relevance scores. Parameter: query

save_pearl

Create a knowledge item with title, content, summary, tags, and type classification. Parameters: content, title, tags

get_pearl

Retrieve the full content of any pearl by its ID. Parameter: pearl_id

list_harbors

List all harbors (workspaces) accessible to the API key. Each harbor is an isolated knowledge boundary.

list_buoys

List topic clusters within a harbor with pearl counts. Parameters: harbor_id, limit

bulk_save

Save up to 100 pearls in a single call. Ideal for batch imports, session summaries, and research dumps.

Get Started

Three Steps to Persistent Memory

Get an API Key

Download the MCP Server

Grab the Python server script and install dependencies:

# Install dependencies
pip install mcp httpx python-dotenv

# Download the server script
curl -o flowtether_mcp.py https://flowtether.ai/downloads/flowtether_mcp.py

Configure Your AI Tool

Add FlowTether to your Claude Code config or any MCP-compatible client:

{
  "mcpServers": {
    "flowtether": {
      "command": "python",
      "args": ["/path/to/flowtether_mcp.py"],
      "env": {
        "FLOWTETHER_API_KEY": "ft_live_your_key_here",
        "FLOWTETHER_HARBOR_ID": "your-harbor-uuid"
      }
    }
  }
}

API Reference

Authentication & Endpoints

Authentication

# All requests require an API key and harbor context
Authorization: Bearer ft_live_your_key_here
X-Harbor-ID: your-harbor-uuid
Content-Type: application/json

Response Format

# Success
{ "success": true, "data": { ... } }

# Error
{ "success": false, "error": "ERR_CODE", "message": "Human readable" }

MCP Endpoints

Method	Path	Purpose
GET	`/api/mcp/harbors`	List accessible harbors
POST	`/api/mcp/create_pearl`	Create a single pearl
POST	`/api/mcp/bulk_ingest`	Batch import (max 100 pearls)
GET	`/api/mcp/buoys`	List topic clusters
GET	`/api/mcp/pearl/{id}`	Get full pearl content
GET	`/api/mcp/duplicates`	Check for duplicate content
GET	`/api/search/`	Semantic + full-text search

Pearl Types

Classify knowledge when saving via MCP:

insight — Realizations and discoveries
decision — Choices made (include the rationale)
question — Unresolved items to investigate
idea — Concepts to explore
reference — External documentation or links
note — General knowledge capture

Full documentation: flowtether.ai/llms-full.txt

Why It Matters

What Agents Build with Persistent Memory

Cross-Session Context

Your agent remembers decisions, preferences, and project context from previous conversations. No more re-explaining.

Research Accumulation

Save findings as you go. Each research session builds on the last. Search semantically to find connections across months of work.

Decision Logging

Record architectural decisions with rationale. When someone asks "why did we choose X?", the agent finds the answer instantly.

Knowledge Organization

Auto-classify knowledge into topic clusters (Buoys). Duplicate detection prevents redundant saves. Tags enable precise retrieval.

Team Knowledge Base

Multiple agents can read and write to the same harbor. Shared context without shared chat history.

Import Everything

Bring in existing knowledge from ChatGPT, Claude, Evernote, Confluence, Notion, Obsidian, PDFs, and 12+ other sources.

For the Human Deciding

Your Data, Your Rules

EU Hosted

Hetzner data center in Germany. European data protection laws (GDPR) apply.

No AI Training

Your content is never used to train models. Ever. Sealed content is excluded from all AI reads.

Full Data Ownership

Export everything at any time. Delete your account and all data is permanently removed.

Harbor Isolation

Row-level security at the database layer. API keys scoped to specific harbors. No cross-workspace leakage.

Secret Scanning

Credentials found in imported content are automatically quarantined. Your API keys stay safe.

Open Protocol

Built on the Model Context Protocol standard. No vendor lock-in. Works with any MCP-compatible client.

Your AI Agent Needs a Knowledge Base