MCP · Workflow orchestration
CLIv4 stableThe problem with AI coding agents is rarely that they cannot write code — it is that they skip the thinking. mcp-probe-kit turns "understand first, plan, then prove it is done" into a path the agent cannot route around: requirements become executable steps, every step leaves evidence, and nothing closes early.
npm i mcp-probe-kitcannot close: review evidence missing
start_feature does not jump into writing code. It reads project context, refreshes the index, runs impact analysis, then breaks the requirement into tasks with acceptance criteria before handing them to the agent.
plan_heartbeat persists completed, skipped and unresolved steps along with their evidence. After a dropped conversation, a restart, or a different agent, resume_plan recomputes what runs next.
converge checks steps, unresolved items, and five kinds of evidence: requirements, spec, implementation, tests and review. If any is missing it refuses closure — "mostly done" does not pass.
start_bugfix runs the SRC-8 eight-step root-cause protocol (inspired by Toyota TBP): reproduction contract, attribution layers, contributing factors. No code changes until the worksheet is closed.
History Session is on by default and archives each turn — real user wording, decisions, open items — as local Markdown. A new chat picks up where the last one stopped, no vector database required.
Native MCP Apps including Feature Workbench, Memory Center and Convergence Gate render plan progress, memory entries and convergence state right inside the client.
Cursor, Claude Desktop, OpenCode and Cline are all supported. Paste one mcpServers entry pointing at npx -y mcp-probe-kit@latest.
Any tool call writes SKILL.md into the workspace and merges the mcp-probe block into AGENTS.md, so the agent knows which step to use and when.
Have the agent split requirements with start_feature, track progress with plan_heartbeat, and close with converge. 29 tools by default, 35 with memory configured.
29 tools by default (24 base + 5 History Session), grouped by responsibility:
These animations are rendered from the same MCP App source shipped in the npm package — not mockups.
No. mcp-probe-kit does not edit code or take over the filesystem. It orchestrates and validates: which workflow applies, how far execution got, whether the evidence is complete. Reading files, editing code and running commands stay with the client agent.
No. The default 29 tools work out of the box and History Session needs no vector database. Only add Qdrant plus an embedding model if you want cross-project reusable experience — that raises the surface from 29 to 35 tools.
Any standard MCP client: Cursor, Claude Desktop, Cline, OpenCode and others. Both legacy MCP and the 2026-07-28 protocol are supported. Node.js 20 or newer is required.
No. The server runs as a local npx process, plan state lives under .mcp-probe-kit/ in your project, and history lives under docs/history-session/. The only external dependency is the optional memory service, which you host yourself.
Local-first — code and execution stay in your own environment.