Groundline

What Groundline MCP server does

Groundline MCP server provides an MCP interface for analyzing engine hot-fire and other bench-test data. It is intended for workflows where report values must be linked to reproducible calculations rather than generated directly from raw samples by a language model. The project is listed in the MCP Registry as io.github.fhs1220888/groundline.

Analysis can cover run phases, sensor condition, limit violations, valve timing, oscillations, simulation comparison, and pulse-shaped measurements. The repository also includes a rule-based agent that can run a fixed checklist without an LLM, plus optional LLM agents for planning and writing findings.

How it works

A client first opens a run and can inspect the available analysis operations. It then asks Groundline to run deterministic checks on the prepared data. Each tool invocation is recorded in an evidence ledger with its parameters, the input file SHA-256, a hash of the tool source, the result, and any figure produced.

Findings must cite ledger entries. Verification checks that reported numbers occur in the cited evidence, allowing documented rounding and conversions between seconds and milliseconds or fractions and percentages. It also checks units and semantic roles, so values labeled as peaks, durations, means, or impulses must come from matching evidence fields. groundline reproduce report.json reruns the ledger against the raw data and compares the results.

The MCP server leaves planning to the connected client. Claude Desktop, Cursor, or another MCP-capable agent can choose the sequence of calls, while the verification rules remain in Groundline.

Setup and configuration

Install the project with its minimal Python dependencies using pip install -e .; the optional all extra is also available. A registry-installed client can launch the server with uvx groundline mcp. From a local checkout, use groundline mcp or groundline-mcp over stdio.

Input data can be a CSV with time in the first column, or NI TDMS data. The command-line workflow supports groundline ingest for converting raw DAQ exports using a JSON mapping, followed by analysis of the generated run file. Reference simulation data and limits can be supplied for comparison and redline checks.

Optional LLM operation supports OpenAI, OpenAI-compatible endpoints, and Anthropic. Configuration may be supplied in .env, environment variables, command-line arguments, or a local endpoint such as Ollama. The rule agent does not require a model.

Tools and capabilities

The MCP surface contains five tools:

• open_run opens the test data for analysis.
• list_analysis_tools lists the available deterministic operations.
• run_analysis executes selected analysis tools.
• verify checks findings against their cited evidence.
• write_html_report produces an HTML report.

The underlying checks include data descriptions, phase segmentation, sensor-health checks, persistent redline detection, valve-response latency, sliding-window FFT oscillation detection, simulation-reference comparison, pulse metrics, and thrust-to-chamber-pressure ratio checks. The latter flags changes greater than 50% during the burn.

Limitations and notes

Groundline is marked as a v0.1 prototype. Public examples include synthetic data and several real test examples, but the real examples do not include redlines, valve commands, or simulation predictions; those checks are validated on synthetic data only. The ingestion draft must be reviewed before use because channel names, windows, scaling, offsets, grid rate, and gap handling may require correction.

The LLM does not receive raw samples; it receives summaries of tool results. This supports on-premise use, but an external model endpoint can still be involved when OpenAI or Anthropic configuration is selected. Groundline MCP server does not replace review of the source data or the generated mapping.