build-mcp

An MCP server that runs a build tool and returns structured compiler errors to an LLM agent in opencode.

The value is in the parser: instead of dumping thousands of lines of noisy build output into the agent's context, it returns a clean list of {file, line, severity, message} diagnostics.

Files

• build_mcp.py — the MCP server. Exposes build and clean tools.

• fake_junkbuild.py — a noisy fake build tool for testing without a real toolchain.

• opencode.jsonc — example opencode config registering the server.

• requirements.txt — the mcp package.

Related MCP server: Build Runner MCP Server

Setup

git clone https://github.com/bpbenpatterson/build-mcp.git
cd build-mcp
python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt

Smoke test (no opencode needed)

python build_mcp.py --selftest

You should see JSON with "error_count": 3 and structured error objects.

Wire into opencode

Copy the mcp block from opencode.jsonc into your opencode config (~/.config/opencode/opencode.jsonc or a project opencode.json), replacing /absolute/path/to/build-mcp with the path you cloned to. Point BUILD_CMD at the real build tool, and command at the venv's python.

Then restart opencode and prompt:

use the build-tools tool to build the project, then fix the errors

The agent will call build, receive the structured diagnostics, and act on them.

Using the real build tool

Set BUILD_CMD to your real command. The server appends --target <name> (or --clean --target <name>). If your tool uses different flags, edit run_build() in build_mcp.py.

Adapt ERROR_RE in build_mcp.py to match your tool's actual error format. That regex is the part that carries your domain knowledge, and it is what makes the tool useful.

Safety

• build has a timeout (BUILD_TIMEOUT_S, default 300s).

• No destructive targets are exposed. If you add deploy or similar, put an approval gate in front of it.