Loading ToolBench
MCP server for automated loop engineering in AI coding workflows
Loop Engineering MCP has 31 tools with generally clear naming (verb_noun pattern) and comprehensive descriptions (avg ~150 chars). However, critical gaps exist: output schemas are entirely undocumented across all tools, LLMs cannot predict return structures; parameter descriptions vary widely in quality; error handling guidance is absent; and several tools lack actionable recovery paths. The server implements sophisticated domain logic (maker/checker separation, budget caps, autonomy scoring) but fails to expose this structure to agents via schema documentation. Most tools fall into the 50-70 range due to missing output schemas and incomplete error guidance.
Automatically merge a PR if merge_policy is 'auto_low_risk' and risk score passes threshold.
Generate a multi-loop autonomy report over the last N days.
Run the 4-condition loop suitability test BEFORE building a loop: does the task repeat, is verification automated, does the agent have real tools, and can the budget absorb iterations?
View the current status of a loop's feature checklist.
Submit an attempt for a loop run (the loop body). Runs the maker self-check then the independent checker gate. On success opens a PR; on a fresh failure asks you to iterate with the same run_id; after max_attempts or a repeated failure it escalates to a human.
Configure a hidden out-of-sample verification command (never shown to the maker agent). Used for orchestrator-only validation.
No output schemas documented for any tool. LLMs cannot predict return structure, field names, or types. This forces agents to guess what fields exist and breaks tool chaining (e.g., run_loop_now returns run_id, but agents have no schema proof).
Error handling is absent. No tool description explains what happens on failure, what errors are retryable, or how to recover. E.g., run_loop_now could fail if the loop doesn't exist, but agents have no guidance.
Destructive operations (auto_merge, delete_loop) lack confirmation or dry-run support. Agents can irreversibly merge PRs or delete loops without a safety gate.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | <=2025-11-25 | v2 |
Create a new automated loop with skill template, state tracking, an independent checker gate, stop rules, and budget caps
Delete a loop and all its associated state.
Evaluate merge risk for a loop PR using deterministic scoring (CI status, diff size, path risk, review comments, flake history, test delta).
Get the full configuration for a specific loop.
Get the skill instructions for a loop.
Get the current state of a loop (active run, metrics, escalations).
Ingest and classify CI failures from a branch using pattern matching (flake, env, dependency, bug).
Initialize a feature checklist for a loop (Anthropic long-running agent pattern).
List all configured loops with their status and metadata.
Mark a feature in the checklist as passed or reset it.
Submit a PR review verdict (approve, request_changes, or escalate) from a reviewer subagent.
Run the hidden verification command for a loop (orchestrator-only, never exposed to maker).
Begin a loop run — returns skill instructions for YOU (the host agent) to execute. No API keys needed. After making changes, call complete_loop_run.
Score a loop using the Loop Engineering Scoring (LES) system. Returns a report on loop health, safety, and autonomy metrics.
Score a specific loop run using LES metrics.
Set or update the independent checker gate (goal_check_command) for a loop. This is the command that objectively decides whether the goal is met, separate from the maker's self-check.
Spawn a read-only checker subagent brief for the active run. The checker is independent from the maker and defaults to REJECT.
Spawn a read-only PR review brief for an independent code reviewer subagent.
Start or resume a loop (activates cron scheduler)
Stop a running loop (doesn't delete it)
Submit the checker's verdict (PASS, REJECT, or ESCALATE) for the active run.
Suggest a loop configuration based on a classified CI failure.
Update configuration fields for an existing loop.
Run HTTP marker checks against a deployed environment to verify deployment success.
View hidden verification metrics for a loop (orchestrator dashboard only).
Parameter descriptions are inconsistent in detail. Some (e.g., create_loop.isolation) are thorough; others (e.g., autonomy_report.days) lack context on valid ranges or defaults. No tool documents parameter constraints (min/max for integers, regex for strings).
Tool descriptions do not explain when to use one tool vs. a similar one. E.g., spawn_checker vs. submit_checker_verdict, or score_loop vs. score_run, lack guidance on the workflow sequence.