MCP server for semantic code analysis, contract auditing, evidence drift detection, and performance evidence correlation using the Lifeblood semantic graph engine.
Lifeblood MCP server defines 7 tools with detailed schemas and descriptions. Naming follows verb_noun convention (lifeblood_*). Most tools have comprehensive parameter documentation with enums, ranges, and constraints. However, descriptions vary in quality, some are verbose (lifeblood_performance_evidence: 400+ chars) while others lack clarity on when to use them. Output schemas are not explicitly documented in the source. Error handling guidance is minimal. No tool annotations (readOnlyHint/destructiveHint) visible. Composition is sound: tools are single-responsibility and chainable via IDs. Security: no credentials exposed as parameters. Overall: solid technical foundation but lacks LLM-optimized descriptions and output schema documentation.
Analyzes a C# workspace to extract semantic graph, operation facts, and source evidence. Supports incremental analysis with coalescing of concurrent requests.
Executes an ordered read-only query plan of multiple tool calls serially within a single snapshot lease. Hard-capped at 32 calls. Rejects non-observation/exclusive/nested-batch tools before any call runs.
Returns the MCP server capabilities and available tools based on current session state.
Audits a contract manifest against the loaded semantic graph. Validates call routes, state access patterns, source text policies, and invariant evidence. Returns findings with optional summarization.
Evaluates repository evidence drift by comparing a baseline evidence manifest against live semantic graph metrics. Returns verdict on whether evidence is current, stale, or flagged based on configurable tolerance policies.
Output schemas not documented. Tools return complex objects (ContractAuditReport, PerformanceEvidence, etc.) but LLMs cannot see the response structure. This forces agents to guess what fields are available and risks failed downstream tool calls.
Descriptions exceed 200 chars for lifeblood_performance_evidence (400+ chars) and lifeblood_batch (300+ chars). Verbose descriptions waste tokens and bury key decision points. LLMs struggle to extract 'when to use this' from walls of text.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) visible in source. Agents cannot distinguish safe read-only tools from state-modifying ones. lifeblood_snapshots (pin/unpin/evict) and lifeblood_analyze (readOnly param) should be annotated.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Processes runtime performance evidence: import validates and summarizes a capture, correlate joins a capture to the leased semantic graph, compare validates two capture identities and returns normalized deltas.
Manages workspace snapshots: list current and retained graph-only history, pin/unpin/evict historical publications, and optionally check drift status.
Error handling lacks recovery guidance. ContractAuditToolHandler throws ArgumentException for missing manifest but does not suggest which parameter to supply or how to fix it. Agents receive no actionable next steps.
Parameter descriptions for lifeblood_performance_evidence use technical jargon ('crossDevice', 'markerAliases', 'maxMeasurements') without explaining when an LLM should set them. Descriptions read as API docs, not agent guidance.