FlowMCP
# š FlowMCP (formerly FlowMCP)
**Cross-model AI memory persistence.** *Your AI remembers everything, everywhere.*
FlowMCP is an MCP (Model Context Protocol) server that gives every AI tool you use ā Claude, ChatGPT, Cursor, and more ā access to the same persistent memory. Explain your project once, and every AI model knows the context.
---
## š ļø Choose Your setup
### š Option 1: Open-Source / Local Mode (Zero Limits)
Perfect for local development. Everything runs on your computer with offline access and **no usage capping**.
#### 1. Start Postgres & FlowMCP
```bash
docker compose up -d
```
*This handles database tables seeding automatically.*
#### 2. Install & Build
```bash
npm install
npm run build
```
#### 3. Setup Environment
```bash
cp .env.example .env
```
The defaults support local Docker correctly. No changes required.
---
### āļø Option 2: Hosted Node Mode (Supabase + Stripe)
For scaling to multiple users, team group spaces, and payment unlocks.
1. Create a **Supabase** project and run `sql/supabase-full-migration.sql` inside your SQL Query Editor.
2. Update your `.env` with:
```bash
DATABASE_URL="your-supabase-connection-string"
TRANSPORT=http
PORT=3001
```
3. (Optional) Add `STRIPE_SECRET_KEY` and `STRIPE_WEBHOOK_SECRET` for upgrades handles.
---
## š Connecting to AI Tools
### š¦ Claude Desktop
Edit `claude_desktop_config.json`:
```json
{
"mcpServers": {
"flowmcp": {
"command": "node",
"args": ["/absolute/path/to/flowmcp-mcp-server/dist/index.js"]
}
}
}
```
### š» Cursor
Create `.cursor/mcp.json` (or add via Cursor Settings Settings ā Features ā MCP):
* **Name:** `FlowMCP`
* **Type:** `command`
* **Command:** `node /absolute/path/to/flowmcp-mcp-server/dist/index.js`
---
## š§ Main Commands
| Command | Mode | Target Audience |
| :--- | :--- | :--- |
| `npm start` | **Stdio Mode** | Standard offline triggers (Claude / Cursor) |
| `TRANSPORT=http npm start` | **HTTP Mode** | Remote Web App triggers (ChatGPT integrations) |
| `npm run dev` | **Watch Mode** | Live rebuilding updates |
---
## šļø Architecture
```text
AI Tool Interface (Claude / ChatGPT / Cursor)
ā
ā MCP Protocol (stdio / HTTP)
ā¼
āāāāāāāāāāāāāāāāāāāāāāāāāāāā
ā FlowMCP Server ā
ā ⢠memory_store ā
ā ⢠memory_recall ā
ā ⢠Auto-assign scopes ā
āāāāāāāāāāāā¬āāāāāāāāāāāāāāāā
ā
ā¼
āāāāāāāāāāāāāāāāāāāāāāāāāāāā
ā PostgreSQL DB ā
ā ⢠Full-text Search ā
ā ⢠Relevance Scoring ā
āāāāāāāāāāāāāāāāāāāāāāāāāāāā
```
## š¤ Collaborative Seating
FlowMCP supports **Group Spaces**.
* **Personal Space**: Automatically triggers on setup for solo memories.
* **Group Spaces**: Allows shared triggers if deployed in Hosted mode with the Next.js Dashboard companion connected.
---
*Created with š Flow Labs frameworks.*
TDQS
Scored across 7 tools
Each tool has a distinct primary purpose: listing spaces, storing/searching/deleting memories, sharing memories, and managing projects. The overlap between flowmcp_store and flowmcp_project is mitigated by explicit guidance to prefer project for structured context. flowmcp_recall and flowmcp_load are clearly separated by memory vs. project.
All tools share the flowmcp_ prefix, but naming conventions vary: list_spaces follows verb_noun, while store, recall, forget, push, project, and load are bare verbs or nouns. 'project' is a noun used as a tool name, and 'load' is generic, making the pattern somewhat inconsistent.
Seven tools is a reasonable count for a memory management server. However, flowmcp_project bundles five distinct actions (snapshot, update_section, list, save_skill, load_skill), effectively expanding the surface area. Still, the overall count is well within the ideal range.
The memory lifecycle (create, read, delete, share) is covered, with update handled via store-with-recall. Project management covers snapshot/update/list/skills but lacks a delete project/section action. There is also no direct 'list all memories' tool, only search via recall. Minor gaps but core workflows are functional.