VibeDrift MCP server checks codebase conventions before agent writes code

VibeDrift originally operated as a scanner that measured codebase drift after code was already written. The new MCP server mode moves that check earlier in the process — the agent can now pause mid-task, query the dominant pattern in a section of the repo, check whether a similar function already exists, and confirm a planned change matches the project's conventions, all before writing anything. The setup requires a single config block using `npx @vibedrift/cli mcp` and works with any MCP-compatible agent without going through a marketplace.

To validate the tool's impact, the team ran a controlled experiment: the same model wrote the same new file under two conditions — alone, and after receiving VibeDrift's convention signal.

To validate the tool's impact, the team ran a controlled experiment: the same model wrote the same new file under two conditions — alone, and after receiving VibeDrift's convention signal. A separate drift check graded both outputs to avoid self-scoring. The headline result was a delta of 0.84 (95% CI: [0.57, 1.11]) in the condition where the convention was non-default and hidden from the agent's context window. A concrete example illustrates the mechanism: on a repo whose house style is `.then()` chains, the agent alone defaulted to `async/await`, while the agent with VibeDrift's signal correctly produced `.then()` chain syntax. Equally important, the effect was a clean zero in all conditions where the convention was already visible in context or was the model's own default — no inflated wins. The post notes the experiment used a small number of tasks per condition and a deliberately clean convention stand-in, so the exact numbers are directional, but the direction is described as solid and the null results as tight.

Key facts

1. 01VibeDrift now runs as an MCP server, moving convention checks to before the agent writes code, not after it ships.
2. 02It integrates with Claude Code, Cursor, or any MCP-compatible agent via a single `npx @vibedrift/cli mcp` config block.
3. 03Controlled experiment: the same model wrote the same file twice — once alone, once with VibeDrift's convention signal — graded by a separate drift check.
4. 04When the convention was non-default and hidden from context, VibeDrift cut drift by a delta of 0.84 (95% CI: [0.57, 1.11]).
5. 05When the convention was already visible in the agent's context or matched the model's default, the measured effect was exactly zero.
6. 06The post acknowledges the experiment used a small number of tasks per condition and a clean stand-in convention, calling the exact numbers directional.
7. 07The tool is described as most valuable on large, established codebases whose conventions exceed the agent's context window.

Topics