MCP Server — Programmatic Tool Interface for Agentic Workflows
Why
While the Agent Skill handles 90% of developer interactions via CLI invocation, agentic workflows (CI/CD pipelines, autonomous coding agents, multi-step validation loops) need programmatic tool access with structured inputs and outputs. MCP (Model Context Protocol) — adopted by every major AI platform with 97M+ monthly SDK downloads — provides the standard interface. An MCP server exposing 3-5 high-level tools with compressed, summarized results gives SpecFact programmatic accessibility without the context bloat of naive MCP implementations.
What Changes
- NEW: MCP server in
src/specfact_cli/mcp/ exposing high-level tools:
validate_spec(spec_path, options) — run validation (spec-only or full-chain), return compressed summarycheck_compatibility(base_spec, revision_spec) — breaking change detection between spec versionsget_requirement_coverage(requirement_id) — traceability status for a specific requirementsuggest_fixes(validation_report_path) — actionable fix recommendations for validation failures
- NEW: Server-side summarization — tools never return raw specs or full diffs; only compressed, actionable summaries suitable for LLM context windows
- NEW: Dual-response pattern: concise text summary for LLM reasoning + ResourceLink URI for full reports that agents can fetch on demand
- NEW:
specfact mcp serve — start the MCP server (stdio transport for IDE integration, HTTP/SSE for CI/CD)
- NEW:
specfact mcp install — generate MCP server configuration for Claude Desktop, Cursor, VS Code, etc.
- NEW: Tool count disciplined to 3-5 tools per MCP best practices (Phil Schmid: 5-15 maximum)
- DESIGN DECISION: MCP server wraps CLI commands internally — the CLI remains the source of truth, MCP is a thin adapter layer
Capabilities
New Capabilities
mcp-server: MCP server exposing 3-5 high-level specification intelligence tools with server-side summarization, dual-response pattern, stdio and HTTP/SSE transports. Wraps CLI commands as thin adapter layer.
Modified Capabilities
(none)
OpenSpec Change Proposal: ai-integration-02-mcp-server
MCP Server — Programmatic Tool Interface for Agentic Workflows
Why
While the Agent Skill handles 90% of developer interactions via CLI invocation, agentic workflows (CI/CD pipelines, autonomous coding agents, multi-step validation loops) need programmatic tool access with structured inputs and outputs. MCP (Model Context Protocol) — adopted by every major AI platform with 97M+ monthly SDK downloads — provides the standard interface. An MCP server exposing 3-5 high-level tools with compressed, summarized results gives SpecFact programmatic accessibility without the context bloat of naive MCP implementations.
What Changes
src/specfact_cli/mcp/exposing high-level tools:validate_spec(spec_path, options)— run validation (spec-only or full-chain), return compressed summarycheck_compatibility(base_spec, revision_spec)— breaking change detection between spec versionsget_requirement_coverage(requirement_id)— traceability status for a specific requirementsuggest_fixes(validation_report_path)— actionable fix recommendations for validation failuresspecfact mcp serve— start the MCP server (stdio transport for IDE integration, HTTP/SSE for CI/CD)specfact mcp install— generate MCP server configuration for Claude Desktop, Cursor, VS Code, etc.Capabilities
New Capabilities
mcp-server: MCP server exposing 3-5 high-level specification intelligence tools with server-side summarization, dual-response pattern, stdio and HTTP/SSE transports. Wraps CLI commands as thin adapter layer.Modified Capabilities
(none)
OpenSpec Change Proposal:
ai-integration-02-mcp-server