↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Spec-driven development/

MCP Probe Kit

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~900 words)
agent-curated 600 fully-ai-generated 703 llm=glm-5.3-flash 610 mcp 31 open-source 194 research-note 570 spec-driven-development 22 workflow 14
Table of Contents

MCP Probe Kit is a solo-maintained, MIT-licensed MCP server that runs a spec-driven delivery loop (requirements, spec, implementation, test, review) as 29 model-visible tools, with a convergence gate that blocks closure until that evidence chain is complete.

Its bet is protocol-native enforcement: where this category’s other members ship slash commands, skills, or CLIs that the harness interprets, MCP Probe Kit makes the workflow itself a set of MCP tools the agent calls, and the converge tool refuses to close a feature until the spec, test, and review evidence exists.

What it is
#

An npm package (npx mcp-probe-kit@latest, first published 2025-10-25, at v4.1.1 after 82 versions) that any MCP client can mount: Cursor, Claude Desktop, Cline, Continue, and others, speaking both the legacy MCP protocol and the 2026-07-28 revision, on Node.js 20 or newer. The tool surface splits into plan persistence and convergence (plan_heartbeat, resume_plan, converge), workflow orchestration (start_feature, start_bugfix, start_onboard, start_ui, start_product, start_ralph), and project and specification tools (init_project, init_project_context, check_spec, estimate), beside code, test, git, and structured-interview utilities, with a 39-tool compatibility surface behind an opt-in flag. Five native MCP Apps render the workflow: a Memory Center, Feature, Bug, and Product Workbenches, and the Convergence Gate itself. State is local: delegated plans persist progress through plan_heartbeat and resume across sessions with resume_plan, chat handoff archives losslessly under docs/history-session/ without a vector database, and the optional memory layer needs a self-hosted Qdrant plus an embedding endpoint.

Status
#

Active and shipping, thinly adopted: 41 stars, 5 forks, created 2025-10-25, pushed 2026-11-01, as of 2026-11-01 (GitHub API), MIT, solo steward (mybolide). The npm package has 82 published versions in its first year, v4.1.1 (2026-10-13) the newest stable, and 859 downloads last month (the 2026-09-30 to 2026-10-29 window, re-read 2026-11-01). The v4 line (2026-08 through 10) shipped native MCP Apps, resumable plans, evidence convergence, parent-child specs, and a managed GitNexus sidecar.

Star History Chart

The independent footprint is empty, and that is this note’s critical angle: a Hacker News search returns nothing, and the only third-party badge is the MseeP listing the README itself carries, so the claims rest on the steward’s own README, homepage, and demo animations.

Strengths
#

  • The convergence gate is protocol-native: converge refuses closure when requirements, spec, implementation, test, or review evidence is incomplete, the category’s convergence idea enforced as a tool call instead of a convention.
  • Plans survive sessions: plan_heartbeat persists actual progress and resume_plan restores the next executable step, so delegated agent work does not restart from zero.
  • One mount covers any MCP client with no per-harness packaging, and both the legacy and 2026-07-28 protocol revisions are supported.
  • The surface is explicit about its own breadth: the compact 29-tool default hides the rest, and the full 39-tool compatibility set is opt-in rather than forced.

Cautions
#

  • Zero independent coverage and a solo steward: 41 stars and 859 npm downloads a month after a year is thin validation for a 29-tool surface.
  • The breadth invites drift: workflow orchestration, memory, GitNexus sidecar management, UI utilities, and commit generation in one server, with 82 npm versions in a year signaling churn.
  • The memory layer’s defaults need self-hosted infrastructure (Qdrant, an embedding endpoint), so the out-of-box experience is the no-memory subset.
  • Enforcement is still the agent choosing to call converge: a harness or prompt that skips tool calls bypasses the gate the same way a yolo-configured agent skips any convention.

Pricing
#

Does not apply: MIT, free, no paid tier found as of 2026-11-01.

Compared to
#

  • GitHub Spec Kit: the movement root scaffolds markdown conventions the harness drives through slash commands; MCP Probe Kit exposes the loop as protocol tools any MCP client can call.
  • Spec Kitty: the other gated-pipeline member; Spec Kitty keeps review state in your repository and runs work packages in worktrees, MCP Probe Kit keeps evidence state in tool results and runs wherever the client does.
  • Archcore: the other convergence-first member; Archcore reviews typed documents in git and names the side that is wrong, MCP Probe Kit gates closure on recorded evidence without a document layer.

Bottom line
#

Recommended for engineers already running an MCP client who want the spec loop delivered as tools, with delegated plans that resume across sessions and a convergence gate, and who accept an unvalidated, fast-churning surface. Not for anyone who needs independent validation, a stable tool API, or a workflow whose state lives in the repository rather than in a server process.

Changes
#

  • 2026-11-01 - Created from the Engineering4AI awesome-list entrant scan (the MCP Servers entry merged 2026-11-01), resolving mybolide/mcp-probe-kit as the protocol-native workflow column with the zero-coverage record as the critical angle.

See also
#

References
#