Skip to content

MCP_PERFORMANCE_ROADMAP.md

Latest commit

 

History

History
60 lines (43 loc) · 2.97 KB

MCP_PERFORMANCE_ROADMAP.md

File metadata and controls

60 lines (43 loc) · 2.97 KB

MCP Performance Roadmap

This is the execution plan for making @geometra/mcp clearly outperform Playwright-style MCP flows on real agent workloads, not just synthetic demos.

Goals

  • Win on agent-facing payload size in every discovery-heavy flow.
  • Reach parity or better on live ATS form completion.
  • Reduce end-to-end wall-clock time, including browser startup.
  • Keep Geometra as the primary interface even when lower-level fallback is needed.

Success Metrics

  • Discovery-heavy workflows: more than 70% less model-facing payload than Playwright MCP.
  • Full multi-step workflows: more than 50% less model-facing payload than Playwright MCP.
  • End-to-end runtime including connect: more than 25% faster than Playwright MCP on warm paths.
  • Live application forms: invalidCount parity or better on Ashby, Greenhouse, Lever, and Workday.
  • Tool turns: fewer MCP calls than Playwright MCP for the same successful workflow.

Phase 1: Reliability and Semantic Actions

  • Replace coordinate-only fallbacks with stable semantic targets wherever possible.
  • Keep reveal, click, and wait flows centered on node ids and semantic filters.
  • Close live-site form gaps on ATS workflows before adding more surface area.
  • Gate regressions with real MCP-vs-MCP benchmarks, not Playwright stand-ins.

Status:

  • geometra_click and batch click now accept semantic targets and auto-reveal before clicking.

Phase 2: Lower Chatter Per Task

  • Expand one-call server-side batches for common workflows like connect + page_model, reveal + click + wait, and fill + validate.
  • Keep default outputs final-only and compact unless the caller explicitly asks for verbose debugging.
  • Trim snapshot-like fallback summaries so the model only receives the fields and actions it can use next.

Phase 3: Runtime Performance

  • Add warm proxy pools instead of a single reusable session slot.
  • Reuse headed/headless browser contexts aggressively across related workflows.
  • Prewarm Chromium for benchmark and agent sessions that are likely to chain multiple tasks.
  • Measure warm and cold paths separately so startup cost does not hide interaction efficiency.

Status:

  • MCP now keeps a small warm proxy pool instead of a single reusable slot, so compatible headed and headless sessions do not evict each other immediately.

Phase 4: Transparent Fallback

  • Keep the Geometra MCP surface stable even when semantic actions need a lower-level browser escape hatch.
  • Detect recoverable action failures and fall back server-side instead of forcing the agent to switch tools.
  • Report when fallback was used so the team can prioritize native fixes.

Benchmark Policy

  • Prefer real MCP-vs-MCP comparisons: @geometra/mcp versus Playwright MCP.
  • Avoid hardcoded local paths in benchmarks; use CLI args, env vars, or repo-relative discovery.
  • Track both payload size and wall-clock time on:
    • discovery-heavy page understanding
    • single-page applications
    • long-running multi-step application workflows
    • warm session reuse