Flint: A Local BYO Agent and Bubble MCP Outside AI Credits

October 2026 changed several things in Bubble at once. Agent 2 reached general availability. Bubble MCP (https://mcp.bubble.io/mcp) went into public beta. Pricing moved too: Starter Web is now $39/month, and existing apps get double credit allowances until October 31. Same topic: Agent 2 GA, MCP Beta,.

For a non-technical founder, the job hasn't changed. You need a working minimum viable product (MVP) out fast, so you can get early customer feedback or investor interest without burning through capital. I still see Bubble as a reliable engine for visual prototyping. But now there's a fork: use the embedded Agent 2 inside the editor, or connect an external agent through the Model Context Protocol (MCP).

What Flint Is: A Local Sidecar, Not a Runtime

First, the naming trap. Recursive distributes an open-weight coding model called Flint under Apache 2.0 via Ollama. That's not what I'm talking about. I mean flint-agent: an npm package under the MIT license, version 1.14.0 released in early October 2026. It's a terminal agent that runs on your own machine, built as a BYO (Bring Your Own) harness:

• It runs against the API providers you choose or a local Ollama instance.
• It connects to external MCP servers through MCP_SERVERS or ~/.flint/mcp.json.
• It ships a headless HTTP API for running local workflows.

Flint is not a Bubble runtime, and it doesn't replace the visual editor. It's a sidecar. It inspects schemas and drives changes over standard protocols.

Agent 2 vs. MCP: Choosing the Right Path for Your MVP

If you're non-technical and want feedback now, I'd start with Agent 2. No terminal, no config, it works right in the browser. The cost is that it burns the Bubble AI credits included in your subscription tier. Same topic: MCP With No Credits and.

If you already use terminal tools and vibe coding workflows, an MCP-connected sidecar gives you a few real advantages:

1. Separate billing. Bubble MCP doesn't draw down Bubble AI credits. You pay your model provider directly or run open models locally.
2. The editor stays in sync. Anything done through https://mcp.bubble.io/mcp lands in your Bubble development environment, so the visual editor remains the source of truth.
3. Local permissions. The agent runs inside your terminal environment, and Bubble MCP gets read/write access to your development database (dev DB). I keep sensitive keys separate.
4. Beta limits. Per the Bubble Manual, MCP doesn't do production deploys or branch merges. You can query server logs through query_logs.

Recommendations for Non-Technical Founders

If you need to validate an MVP this week with no technical overhead, use Bubble's native canvas and Agent 2. If you hit credit limits, or you're already working with local AI terminal tools like Flint, connect to Bubble MCP. That gives you a controlled, cheaper sidecar for testing product features quickly. Same topic: Bubble MCP on Free With.