Model Context Protocol / tool surfaces

MCPs should make capability visible.

These projects turn software into explicit tools for agents. The transport, authority boundary, storage model, and failure state are part of the product—not plumbing to hide.

Available now

Remote Streamable HTTP server

TurtlePen MCP

The canonical 73-tool TurtlePen engine over public Streamable HTTP. Each MCP session retains one active document and a confined artifact workspace across requests.

73 tools · v0.3.2 · D1 session index · R2 state + artifacts

Source
Available now

Local stdio server

Local Files MCP

A GUI-installed MCP server that gives ChatGPT constrained local file access without pretending a remote service can see the user’s machine.

Python · MIT · explicit local boundary

Source
Prototype lineage

Source-intelligence research

RegistryRouter MCP

Earlier Brainn.dev research into routing framework questions through official source registries, separating discovery from execution, and keeping approval explicit.

Package and policy lineage in the prior Brainn.dev repository

Source

Hosted TurtlePen boundary

One engine. Durable sessions.

Brainn.dev vendors the same source used by the local stdio server and calls its live createTools(session) registry. The web adapter owns transport and persistence only; it does not reimplement measurement, validation, or SVG rendering.

Live service contract

01

Streamable HTTP

POST JSON-RPC with SSE responses and MCP protocol 2025-06-18.

02

Session continuity

Reuse Mcp-Session-Id; one isolated active document persists across calls.

03

Honest limits

Rate, payload, history, and storage caps refuse by name; nothing is truncated.

04

Confined artifacts

Session files and inline SVG renders stay in expiring, versioned R2 storage.

Connect to the canonical TurtlePen server.

Add https://brainn.dev/api/mcp/turtlepen to a Streamable HTTP MCP client. The Sites discovery route /api/mcp reaches the same engine.

Open workspace