MCP App Store
by KeenEthics
Averth icon

Averth MCP Server for ChatGPT

By Averth

Overview

Averth is a cost-analysis plugin for AI agents: it measures the fully loaded cost of every agent run. Submit a sanitized batch of runs and get the true cost per accepted outcome, the share of spend burned by the most expensive runs, and an offline replay of the cost policies that would have prevented the waste. Analysis is retrospective only. Averth never enforces policies, never touches live systems, and needs no account, no integration, and no raw prompts or customer data.

How to connect

  1. ChatGPT: Open the Averth listing in the ChatGPT app directory and choose Connect. ChatGPT listing Averth exposes 4 tools here.

Tools

analyze_run_batch

ChatGPT
Analyze a sanitized batch of agent run records and compute fully-loaded cost per accepted outcome, the expensive-tail share of spend, and a breakdown across the five cost layers (context compounding, external tool spend, token yield, heavy tail, model mix plus human review). ledger_json is a JSON array of run records; each record may include run_id, model_cost, tool_cost, retry_cost, human_cost, human_minutes, retries, input_tokens, output_tokens, waste_tokens, total_cost, success, outcome_inferred, outcome_unknown, reopened, business_value, model_name. An averth export_ledger object (with an 'attempts' array) is also accepted. A run with no recorded outcome is outcome-unknown: it counts as neither a success nor a failure, and cost per accepted outcome is never computed from it. Redaction warning: ledgers export caller free text verbatim, including case_id, tool names, and retry reasons, and that free text can contain customer data; redact sensitive values before sharing. The ledger is processed in memory only and is never stored.

explain_cost_layers

ChatGPT
Explain the five cost layers used to compute fully-loaded agent economics: context compounding, external tool spend, token yield, heavy-tail outliers, and multi-model attribution with loaded human labor. Returns a static explainer; takes no input.

simulate_policy

ChatGPT
Replay a hypothetical cost policy against a submitted run ledger, offline. Reports which completed runs crossed the proposed thresholds and how much historical spend is associated with those runs, split between historical spend on failed runs that would have been stopped and collateral (successful runs that would have been stopped). This is never projected savings: a live policy acts mid-run, not on completed runs. policy_json is a JSON object with optional keys: max_cost_per_attempt (stop runs costing more than this many dollars) and yield_floor (stop runs whose token yield falls below this 0-1 fraction). At least one must be set. This is a historical replay only: it enforces nothing, changes nothing, and the ledger is processed in memory and never stored.

verify_evidence

ChatGPT
Judge human-written acceptance checks against submitted recorded runs of coding-agent work, using Averth V0 (experimental, merged to master). Read-only, in memory, no execution, no repo access, no merge guard, no repair. evidence_json is a JSON object with 'task_intent', 'task_intent_source', and a 'checks' array; each check has id, text, check_type (recorded_run, artifact_presence, external, or inferred), and for recorded_run a required_command, for artifact_presence an artifact_path. Runs are supplied in evidence_json as 'runs' (a recorded-run ledger object with a 'runs' array, or a plain array of run records with command, start_timestamp, end_timestamp, and exit_code), or as raw text in ingest_text (any supported shape: JSON, JSONL, pretty JSON, plain log, OpenTelemetry-style span records, meter-style events; ingested, never judged by itself). trace_jsonl is summarised only, never a verdict. repo_snapshot_json is an optional object of artifact path to caller-supplied SHA-256 digest; it is NOT a server scan, so artifact rows are Tier 1, not Tier 2. final_edit_marker is an optional ISO timestamp; runs before it are superseded, retained, never cited. repo_ref is an optional caller-supplied label for the repo, not a fetched repo. infer_checks adds INFERRED rows from observed commands; they are always NOT OBSERVED, never human acceptance, never VERIFIED. Returns a JSON ledger plus a Markdown report. Statuses are VERIFIED, FAILED, NOT OBSERVED, EXTERNAL-UNVERIFIABLE, or NOT APPLICABLE. NOT OBSERVED means the records do not show it, never that it did not happen, and never a failure. A VERIFIED row proves only what its cited observation shows. V0 was tested only on a local fixture and public traces, not on a real developer workflow; no validated repair, no measured savings, no paying customer. The ledger is processed in memory only and is never stored.

Capabilities

Analyze Agent Run CostsReplay Cost Policies

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/averth"><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/averth)