Preflight

What preflight-dev/preflight MCP server does

The preflight-dev/preflight MCP server provides 24 MCP tools for improving prompt quality and session control in Claude Code. Its central entry point, preflight_check, classifies an incoming prompt and selects follow-up checks. Short routine commands can pass through, while vague, multi-step, or cross-service requests can trigger clarification and planning workflows.

The tool set covers five areas: prompt discipline, session-history intelligence, analysis and reporting, verification, and workspace hygiene. It can create execution plans, resolve vague follow-ups to file targets, enrich sub-agent tasks, order work by dependency and risk, and save checkpoints with commits and resumption notes.

How it works

Prompt triage uses signals such as skip keywords, multi-step language, cross-service terms, prompt length, file references, and vague verbs or pronouns. A correction pattern can raise the required review level when a prompt resembles a previously logged mistake. The force_level option on preflight_check can explicitly select skip, light, or full handling.

The server reads Claude Code session data stored as JSONL under ~/.claude/projects/. It extracts prompts, assistant responses, tool calls, sub-agent activity, corrections, context compactions, errors, and commits. Session events are indexed in per-project LanceDB databases for semantic search. Related project data can also be used for cross-service contract and history lookups.

Setup and configuration

The documented quick-start command runs the server through npm:

npx -y preflight-dev-serve

Claude Code can register that command with claude mcp add preflight -- npx -y preflight-dev-serve. For project-specific analysis, the README shows setting CLAUDE_PROJECT_DIR to the target workspace. A cloned setup can run the TypeScript entry point with tsx, while a global npm installation provides the preflight-dev-serve executable.

The project supports a .preflight/ configuration directory with YAML configuration for triage rules and thresholds. Its local data includes timeline indexes, extracted contracts, and project metadata under ~/.preflight/projects/.

Tools and capabilities

The preflight-dev/preflight MCP server includes tools for:

• Checking, scoring, clarifying, scoping, enriching, and sequencing prompts or tasks.
• Tracking corrections and matching new prompts against learned patterns.
• Monitoring session health, scanning active sessions, viewing timelines, and searching history.
• Indexing project history and searching API contracts, types, interfaces, routes, and schemas.
• Creating scorecards with session, trend, or cross-project views, including radar SVG, PDF, or Markdown output.
• Estimating token usage, dollar cost, correction waste, and potential preflight savings.
• Running type checks, tests, and builds through verify_completion.
• Auditing workspace documentation and summarizing changes since a checkpoint.

Limitations and notes

The README specifically documents use with Claude Code; it does not establish compatibility with other MCP clients. The server relies on Claude Code session files for its history-oriented features, so results depend on those files being available. Semantic search uses local vector embeddings by default according to the documented LanceDB schema, with OpenAI-sized vectors also described as an option. Cost figures are estimates rather than billing records, and completion verification only runs the checks configured or available for the workspace.