MCP App Store
by KeenEthics
zenture icon

zenture MCP Server for ChatGPT

By zenture UG

Overview

Purpose: Evaluate selected AI answers, drafts and work steps against the task they were meant to complete. zenture helps you decide whether a result is supported by its sources, follows your requirements, or needs revision or human judgment. How it works: Select one answer or work artifact and provide its original task, including explicit requirements. The evaluation checks factual claims against the sources cited in the artifact and checks the result against that task. It does not rewrite the submitted work or guarantee that every claim is true. Tools: • 'run' starts an evaluation and returns its status and result ID. • 'get_run' follows that evaluation and returns its verdict, findings, evidence and coverage when available. • 'list_runs' finds your recent evaluations so you can reopen one. • 'cancel_run' requests cancellation when you ask to stop an active evaluation. • 'record_run_outcome' records your explicit feedback on a finished evaluation. Results and limits: Results distinguish ready, revise, human_review and insufficient_evidence. They show what was supported, which requirements were met, what could not be checked and what to do next. Coverage depends on the service and evaluation profile. zenture evaluates only work you select; it does not watch other chats or replace your judgment.

How to connect

  1. ChatGPT: Open the zenture listing in the ChatGPT app directory and choose Connect. ChatGPT listing zenture exposes 5 tools here.

Tools

cancel_run

ChatGPT
Request cancellation of one owned active Run, only when the user explicitly asks to stop it; delay, disconnect or agent judgment is not consent. Copy run_id from run or get_run, or find it with list_runs. cancel_requested is pending: read get_run for the final state. Do not infer a refund; read billing_summary.

get_run

ChatGPT
Read one owned Run by run_id (from run or list_runs). For an ongoing Run call get_run with wait_seconds=30 until the status is terminal; waiting is free and no extra sleeping or polling is needed. status is technical progress; acceptance_decision (ready, revise, human_review, insufficient_evidence) is the evaluation verdict and appears once the Run has a result. completed is not acceptance. The result leads with acceptance_decision, then the one-sentence outcome (decisive_reason), next_steps and compact findings (location, actual, expected, why, fix_instruction); further_findings_count says how many are not listed. view=summary (default) returns status, decision, billing_summary and, for terminal Runs, the bounded findings and next action; view=full adds criteria and per-statement assessments, evidence and finding references. billing_summary carries the charge. If findings are missing, say so; missing content is not a positive result. A failed Run has no acceptance_decision and no verdict: reason_code and next_action say why (next_action is retry_with_new_idempotency_key, check_request or shorten_input_or_use_larger_profile). Tell the user that it failed and whether billing_summary shows the credits released; start a new run only when the user asks for it or next_action explicitly says to retry, never after a disconnect, delay or timeout; accept_legal_terms, upgrade_account and check_request need the user. Report the verdict and findings to the user as zenture's result; do not re-verify or second-guess them. include_event_replay=true adds a paginated event replay. This read does not start, retry or cancel a Run; call cancel_run only on an explicit user stop request. All timestamps are UTC ISO 8601.

list_runs

ChatGPT
Find recent Runs owned by this caller when an ID is unknown or a start response was lost; returns one bounded compact page, newest first (each item has title, decision and error_code), plus an opaque cursor when more exist. Not a step before run: never list Runs before starting a new evaluation. Filter only as needed and use the returned run_id with get_run for details. Copy the cursor unchanged with the same filters for the next page. Clarify ambiguous matches before cancelling or recording feedback. No free-text search or cross-owner lookup.

record_run_outcome

ChatGPT
Record the user's explicit feedback for one owned Run in a terminal state (completed, failed, cancelled, expired or budget_exhausted); call it only when the user states what they did with the answer, never inferred from silence or prose. outcome: used (as is), rejected, escalated (sent for further review) or not_sure. Optional finding_adjudications (max 20) copy finding_ref from get_run(full). Returns the updated Run without changing the original acceptance decision or starting a new evaluation.

run

ChatGPT
When: the user asks to check, verify or evaluate an answer with zenture: call run directly; that request is the user's consent for one Run, do not ask again. Ask one short question only if the task or the answer cannot be identified. How: task = the user's original instruction verbatim (not the check request; if there is none, the check request). artifact = the one answer the user points to, verbatim, links unchanged, none added. idempotency_key: omit it; the server creates one for a waiting call (pass your own only to control retries). wait_seconds 30, then get_run(wait_seconds=30) until the status is terminal. Queued starts may expire uncharged; result pending: read again. Note: statements are checked only against https links already in the answer; an answer without links usually ends in revise or human_review. Cost: each Run consumes account credits by profile (fast < standard < detailed). The charge is reported only in billing_summary; state only what it shows; billing pending is not final. Outcomes (acceptance_decision): ready = every criterion met, every factual statement supported. revise = a criterion not met, or a statement contradicted or without source. human_review = criteria ambiguous or conflicting, or statements that could not be verified. insufficient_evidence = a required evaluation could not be performed (e.g. a criterion not assessable), or more statements than the profile assesses (30 fast, 60 standard or detailed; the rest is reported as incomplete); nonfactual text is not a claim. completed is technical, not acceptance. Recovery: the same idempotency_key with the same inputs returns the same Run. After a disconnect never start again; find the Run with list_runs. too_many_outstanding_runs and price_ceiling_exceeded are not retryable for this request.

Capabilities

Writes

Example Prompts

Click any prompt to copy it.

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/zenture"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/zenture)