AgenticSettle Verify
OfficialAgenticSettle Verify is a free MCP server for objectively scoring AI-generated outputs and managing verification records, appeals, feedback, templates, and alerts—without moving money.
Score AI output 0–100 with PASS/PARTIAL/FAIL or tier verdicts using
verify_output, including optional SLA checks.Retrieve an existing verdict by report ID with
check_verdict.List and filter past verification records with pagination via
list_verifications.Get aggregate agent performance statistics with
get_insights.Submit feedback about the server or a verdict using
submit_feedback.Dispute a verdict with a deterministic tamper check via
submit_appeal, and check its status withcheck_appeal.List bundled domain rubric templates with
list_criteria_templates.List this account's own Manager AI oversight alerts with
get_manager_alerts.
AgenticSettle Verify MCP Server
Free VOP (Verified Output Protocol) verification for AI agent outputs — lets Claude and any MCP-compatible AI agent objectively score task outputs 0–100 and get a PASS/PARTIAL/FAIL verdict.
This server exposes the free, verification-only slice of a larger settlement
protocol. VOP scores an agent's output; in the full protocol the verdict is
then anchored on-chain as an evidence_hash, and (on AgenticSettle's paid
platform, not in this server) escrowed payment is released once that verdict
passes. That full loop — verify → anchor → settle — has been exercised
end-to-end as a test on Base Sepolia (an OP Stack testnet, test tokens only,
not production settlement) using the x402 payment protocol. The hosted
verification service computes an evidence_hash for each verdict but does not
currently write it on-chain; on-chain anchoring has only been exercised on a
testnet. This repository is
only the verification step, free and standalone: it sends no blockchain
transactions and moves no money.
What this server is: A quality-verification tool. Submit a task description and an agent's output; get back an objective 0–100 score, a verdict, and a tier. What this server is not: A payment processor, escrow service, or anything that moves money, tokens, or any financial asset. This server has no such capability — every tool here is read-only or write-only-to-a-verification-record, and none of them transfer value between parties. (AgenticSettle's full platform does support quality-gated escrow settlement for paying customers, but those tools are intentionally not part of this MCP server — see Why this is a separate, smaller server.)
Installation
Requirements: Python 3.10+, an AgenticSettle API key (free, self-service, no credit card).
Get a key instantly — no waiting, no email verification step (an email address is still required):
curl -X POST https://app.agenticsettle.io/v2/signup \
-H "Content-Type: application/json" \
-d '{"email":"you@example.com"}'This returns a free-tier key immediately (up to 10 verify_output calls/day per key, and a shared daily cap
across all free keys may apply first; no per-key daily limit on the other 8 tools). It's shown once in the response — save it.
pip install git+https://github.com/agenticsettleio/agenticsettle-mcp.gitAdd to Claude Code
export AGENTIC_SETTLE_API_KEY="your-api-key-here"
claude mcp add agenticsettle-verify -- python -m agenticsettle_verify_mcpAdd to Claude Desktop
Add this block to your claude_desktop_config.json:
{
"mcpServers": {
"agenticsettle-verify": {
"command": "python",
"args": ["-m", "agenticsettle_verify_mcp"],
"env": {
"AGENTIC_SETTLE_BASE_URL": "https://app.agenticsettle.io",
"AGENTIC_SETTLE_API_KEY": "your-api-key-here"
}
}
}
}Or install the packaged .mcpb extension (single-click, see Releases — built automatically per-platform on each tagged release) — Claude Desktop will prompt you for your API key on install.
Run manually (stdio transport)
python -m agenticsettle_verify_mcpRelated MCP server: agentic-platform
Tools
All 9 tools are free — no paid account, no escrow, no payment of any kind (a free API key is required).
The local server exposes all 9 tools. The hosted endpoint (https://app.agenticsettle.io/mcp) exposes 6 of them: verify_output, check_verdict, get_insights, list_verifications, list_criteria_templates, submit_feedback. submit_appeal, check_appeal and get_manager_alerts are local-server only.
Tool | What it does |
| Score AI agent output 0–100, get a PASS/PARTIAL/FAIL verdict |
| Retrieve an existing verification verdict by report ID |
| Get an agent's aggregate VOP performance statistics |
| List past verification records with filters and pagination |
| Report an issue or suggestion about this server or a verdict |
| Dispute a verdict and get a deterministic tamper check |
| Check the status and outcome of a submitted appeal |
| List bundled domain rubric templates |
| List this account's own oversight incidents (never another account's) |
Example
result = await verify_output(
task_description="Write a 500-word blog intro about caching.",
result_content="<agent's actual output>",
sla={"min_words": 500, "required_sections": ["introduction"]},
)
# {"report_id": "VER-...", "verdict": "PASS", "score": 87, "tier": "Standard", ...}Runnable examples for LangGraph, CrewAI and the hosted MCP endpoint are in examples/.
A small agent-agnostic demo that releases work only when its visible tests pass and this verification agrees (the settlement receipt is simulated, no money moves) is in agenticsettle-test-gated-release.
Getting reliable auto-invocation
Some MCP hosts (including Claude Desktop) don't automatically call verify_output
on every "is this ready?" / "check this for me" request, especially right after
install — Claude may just answer from its own judgment instead of invoking the
tool. If you notice this, add the following to Claude's personal
instructions/memory (Claude.ai Settings → Profile → "What personal preferences
should Claude consider in responses?", or the equivalent in Claude Desktop):
When asked for a final check on AI-generated output (a report, essay, code,
analysis, or any deliverable the user will submit, ship, or rely on) — phrases
like "is this ready to submit", "verify this", "check this is correct", "make
sure there are no errors" — do not answer from your own judgment alone. Call
the AgenticSettle Verify tool (verify_output) first and base your answer on
its result. This applies equally to content you (Claude) just generated in
this conversation, not only to content the user wrote.
Call discipline:
- Call verify_output exactly once per user request for unchanged content. Do
not retry hoping for a better score — grounded_factcheck uses live web
search, so minor score variance between runs on identical content is
expected and is not a reason to retry.
- If the result is PARTIAL/FAIL, report it to the user first rather than
automatically revising and re-verifying in a loop. Explain what needs
fixing; only produce and re-verify a new version once the user asks for a
revision.
- Only re-verify when the content has substantively changed — cosmetic or
formatting-only edits do not need a fresh verification.This is optional — the tool works without it — but it makes Claude invoke
verify_output proactively rather than only when explicitly told to use the
AgenticSettle tool.
Why this is a separate, smaller server
AgenticSettle's full platform also supports quality-gated escrow settlement —
locking tokens between a buyer and an agent, releasing them automatically
once VOP confirms delivery. Those tools (submit_task, complete_task,
settle_payment, cancel_task, register_token, get_token_balance,
create_criteria, sign_criteria, get_criteria, list_tasks, get_task)
move value between two parties, so they are intentionally not included
in this MCP server. This server exists to let anyone try objective AI
output verification — free, no account commitment — as a way to build
intuition about output quality before ever touching a paid or escrow
workflow. If you need the escrow/settlement tools, use the AgenticSettle
API/SDK directly (see agenticsettle.io).
Configuration
Two environment variables are required to get started; the rest are optional tuning knobs:
export AGENTIC_SETTLE_BASE_URL="https://app.agenticsettle.io" # default if omitted
export AGENTIC_SETTLE_API_KEY="your-api-key-here" # requiredVariable | Default | Purpose |
|
| Backend URL |
| (required) |
|
|
| Per-HTTP-request timeout in seconds (grounded verification can take up to ~53s) |
|
| Max attempts on 429/502/503/504 or network errors |
|
| Comma-separated backoff seconds between retries |
| (unset) | If set, |
If AGENTIC_SETTLE_API_KEY is not set, every tool call raises an error immediately:
AGENTIC_SETTLE_API_KEY not configured. Get a free key instantly: POST https://app.agenticsettle.io/v2/signup with {"email": "you@example.com"}, then set the environment variable.Security
AGENTIC_SETTLE_API_KEYis read from environment variables — never hardcoded.The key is sent only in the
x-api-keyrequest header over HTTPS; it is never included in server responses or logs.If the key is missing, every tool call raises immediately without making any network call — this is one of the two exceptions to the dict-error contract below.
The server runs over stdio transport — no network port is opened; communication is exclusively through the MCP host process (Claude, Claude Desktop, etc.).
This server transfers no money, tokens, or financial assets of any kind — every tool either reads a verification/incident record or writes a verification/feedback/appeal record.
Rate limits
Tier | Endpoint | Limit |
No API key (public) | none — all tools here require a key | — |
Free API key |
| Up to 10 verify calls/day per key (a shared daily cap across free keys may apply first) |
Free API key | all other tools | No daily quota — auth only |
When the limit is exceeded the backend returns HTTP 429. This server retries automatically with backoff, so brief bursts are absorbed transparently.
Fault tolerance
The server retries automatically on HTTP 429/502/503/504 and network errors with backoff (5s → 15s, max 3 attempts).
Error reference
Most errors are returned, not raised: on invalid input or a backend 4xx response, a tool returns {"error": str, "status_code": int} instead of throwing — check for an "error" key in the result rather than wrapping calls in try/except. Only two situations raise an actual exception:
A missing
AGENTIC_SETTLE_API_KEY— raises immediately, since no tool can function without it.The backend unreachable after all retries — raises
RuntimeError(a connectivity failure, not a 4xx/5xx response, so there's nostatus_codeto attach).
Situation | What you get |
Missing/invalid API key | Raises |
Invalid | Returns |
| Returns |
Backend temporarily down (429/502/503/504) | Auto-retry, then success or a structured error dict |
Backend down after retries exhausted | Raises |
Empty/short | Not an error — returns a normal report with |
Privacy Policy
See PRIVACY.md for the full policy — what data this server sends to the AgenticSettle backend, how it's used, stored, and how to request access or deletion.
FAQ
Q: Is this a financial service or payment processor? A: No. This server has no capability to move money, tokens, or any financial asset — every tool either scores/retrieves a verification or records feedback.
Q: Do I need a paid plan? A: No — every tool in this server is free. AgenticSettle's separate paid tier (escrow-gated settlement) is not exposed here at all.
Q: Which AI agents are supported? A: Any MCP-compatible client can connect. We have tested Claude Code and Claude Desktop; other clients should work but have not been tested by us. VOP is model-neutral; it evaluates the output, not the model that produced it.
Q: What happens to the content I submit? A: See PRIVACY.md — briefly, an external AI service (Anthropic) uses it to compute your verification score. The record we keep for retrieval stores the score, verdict, hashes, and short AI-extracted excerpts, not your full text.
License
Available Tools
9 toolscheck_appealARead-only
Retrieve an appeal's status and tamper-check result. Free tier.
Args: appeal_id: The appeal_id returned by submit_appeal.
Returns: dict with keys: appeal_id, verification_id, appellant, reason, original_score, original_verdict, original_evidence_hash, recomputed_score, recomputed_verdict, recomputed_evidence_hash, hash_match (bool), status ("OPEN"|"RESOLVED"), outcome (None|"UPHELD"|"OVERTURNED" — set once resolved), resolution_score, resolution_note, created_at, resolved_at
Inputs: appeal_id (required). Returns: object
check_verdictARead-only
Retrieve an existing VOP verification verdict by report ID.
Use the report_id returned by verify_output to fetch full verdict details.
Args: report_id: The verification ID returned by verify_output (the report_id field). audience: Response detail level. "agent" (default) — machine-readable verdict for agent consumption. "customer" — adds human-readable explanations for presenting results to end users. Do not use "admin" — it is reserved for platform operators only.
Returns: dict with keys: report_id, verdict, score, tier, fail_codes, issued_at (ISO 8601), settlement (always None), agent_performance
Inputs: report_id (required), audience. Returns: object
get_insightsARead-only
Retrieve VOP performance statistics for an agent. Free tier.
Aggregates verdict counts and average score across this agent's verify_output history. Use to evaluate an agent's track record before relying on its output.
Args: agent_id: The agent identifier to look up.
Returns: dict with keys: agent_id, total_jobs (int), pass_count (int), partial_count (int), fail_count (int), pass_rate_pct (float, 0-100 — NOT a 0.0-1.0 fraction), avg_vop_score (float, 0-100).
Inputs: agent_id (required). Returns: object
get_manager_alertsARead-only
List Manager AI oversight alerts (incidents) for this account. Free tier.
Manager AI continuously watches for customer-behavior anomalies (repeat rework, repeat disputes) and verification-quality anomalies in the background, and raises an incident when something looks off. This tool is always scoped to the account behind this server's own API key — there is no customer_id/agent_id parameter, so it is impossible to query another tenant's alerts.
Args: domain: Filter by domain (optional). Values seen in practice: "customer_behavior", "vop_quality", "verification_audit". severity: Filter by severity (optional). Values seen in practice: "info", "low", "medium", "high", "critical". status: Filter by status (optional). Values: "open", "resolved". limit: Maximum number of incidents to return (1-500, default 50).
Returns: dict with keys: total (int), incidents (list of dicts with keys: incident_id, domain, severity, signal, evidence (dict), status, created_at)
Inputs: limit, domain, status, severity. Returns: object
list_criteria_templatesARead-only
List bundled domain criteria templates showing what a well-specified verification rubric looks like. Free tier.
Args: domain: Filter by domain name (optional, e.g. "data_analysis"). Matches domain equality or a template_id prefix match.
Returns: dict with key: templates (list of dicts, each with template_id, domain, items (list of {id, name, weight}), pass_threshold, partial_range ([low, high]))
Inputs: domain. Returns: object
list_verificationsARead-only
List VOP verification records with optional filters and pagination. Free tier.
Returns verifications in reverse chronological order.
Args: agent_id: Filter by agent ID (optional). verdict: Filter by verdict (optional). Values: "PASS", "FAIL", "PARTIAL" tier: Filter by tier (optional). Values: "Platinum", "Standard", "Partial", "Risk" date_from: ISO 8601 date lower bound, inclusive (optional). Example: "2026-06-01" date_to: ISO 8601 date upper bound, inclusive (optional). Example: "2026-06-30" limit: Maximum number of records to return (1–500, default 50). offset: Pagination offset — number of records to skip (default 0).
Returns: dict with keys: total (int), limit (int), offset (int), items (list of verification dicts with keys: verification_id, agent_id, verdict, score, tier, fail_codes, created_at)
Inputs: tier, limit, offset, date_to, +3 more. Returns: object
submit_appealA
Dispute a VOP verdict and get a deterministic tamper check. Free tier.
Re-runs the SAME engine version on the same inputs and compares the recomputed evidence hash to the one stored at issue time. A hash match proves the stored score was not altered after issue (the dispute is about scoring quality, not tampering); a mismatch signals engine version drift or a stored-score discrepancy that warrants manual review.
You must supply the EXACT original task_description (as request_content) and result_content used in the disputed verify_output call — they are hash-checked against the stored verification, so reconstructed or paraphrased text will be rejected. There is a time window to appeal after a verdict is issued (window length is server-configured); appealing after it closes returns an error.
Args: verification_id: The report_id from the disputed verify_output/check_verdict call. request_content: The exact original task_description/instructions (must hash-match what was stored). result_content: The exact original output text that was verified (must hash-match what was stored). appellant: Who is appealing (optional free text, e.g. "agent", "customer", or an identifier). reason: Why the verdict is being disputed (optional, recommended).
Returns: dict with keys: appeal_id, verification_id, status ("OPEN"), appeal_window (dict: window_hours, issued_at, deadline, enforced), tamper_check (dict: hash_match (bool) — false does NOT necessarily mean tampering, see interpretation; original_evidence_hash, recomputed_evidence_hash, interpretation (str)), original (dict: score, verdict), recomputed (dict: score, verdict, confidence, review_recommended, review_reasons)
Inputs: result_content (required), request_content (required), verification_id (required), reason, +1 more. Returns: object
submit_feedbackA
Submit feedback about this MCP server or a VOP verification result. Free tier.
Use after any tool call to report issues, request features, or share what worked well. Feedback is reviewed weekly and drives product improvements.
Args: rating: Satisfaction score 1–5 (1 = very poor, 5 = excellent). category: Feedback type. One of: "wrong_verdict" — VOP verdict seems incorrect for the output "feature_request" — request a new capability or parameter "bug" — tool raised an error or behaved unexpectedly "praise" — something worked especially well "other" — anything else comment: Description of the issue or suggestion (max 2,000 characters). tool_name: The MCP tool name this feedback is about (optional). Example: "verify_output", "check_verdict" report_id: The report_id from a specific verification (optional). Helps correlate feedback with the exact VOP result.
Returns: dict with keys: received (bool), feedback_id (str), message (str)
Inputs: rating (required), comment (required), category (required), report_id, +1 more. Returns: object
verify_outputA
Verify AI-generated output — including the calling model's own prior responses, not just another agent's — using VOP (Verified Output Protocol). Free tier, no account or escrow setup required.
Primary use case: verifying AI-generated content (a report, essay, code, analysis, or any output a user will submit, ship, or otherwise rely on as final) before it is presented as finished, including content generated earlier in the same conversation. An LLM's own stated confidence in its output is not evidence of factual correctness; this tool checks factual correctness independently via grounded web search rather than the model's self-assessment.
A verdict describes the exact content submitted, not any later revision of it. If the content is subsequently changed, the earlier PASS/FAIL result no longer describes what now exists — verifying a revision requires submitting the revised content as its own call.
Score variance: because the grounded critic uses live web search, re-running this tool on identical, unchanged content can produce a slightly different score between runs. This reflects live-search noise rather than a defect that repeated calls would resolve — the score does not converge toward a "truer" value with more calls on unchanged input, and each call counts against the daily quota.
Returns a 4-bucket verdict (Platinum/Standard/Partial/Risk), a score (0-100), and fail codes for a single piece of agent-generated output.
Args: task_description: The original task instructions or requirements to verify against. result_content: The agent-generated output text to be evaluated. Maximum 200,000 characters. agent_id: Agent identifier (optional). Used to track performance history — pass the same value across calls to build up get_insights data. sla: Inline evaluation criteria dict (optional). Supported keys: "required_sections": list[str] — headings that must appear in the output "min_words": int — minimum word count (hard-fail if violated) "min_chars": int — minimum character count "max_chars": int — maximum character count (0 = no limit) "min_numbers": int — minimum numeric data points required "min_citations": int — minimum citation/reference count required Example: {"min_words": 500, "min_citations": 3, "required_sections": ["introduction", "conclusion"]}
Returns: dict with keys: report_id, verdict ("PASS"|"PARTIAL"|"FAIL"), score (0-100), tier ("Platinum"|"Standard"|"Partial"|"Risk"), fail_codes (list), settlement (always None — this server has no settlement capability), agent_performance
Inputs: result_content (required), task_description (required), sla, agent_id. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 9 tools
Each tool has a clearly distinct purpose: verify_output submits content for verification, check_verdict retrieves a specific result, get_insights tracks agent performance, list_verifications searches records, submit_feedback and submit_appeal handle user-side actions. No two tools have overlapping functionality.
All tools follow a consistent verb_noun pattern (verify_output, check_verdict, get_insights, list_verifications, submit_feedback, list_criteria_templates, get_manager_alerts). The naming is predictable and clearly conveys the action and resource involved.
With 9 tools, the set is well-scoped for the domain of AI output verification. Each tool serves a necessary function in the verification, retrieval, feedback, and appeal lifecycle without redundancy or bloat.
The tool surface covers the full verification lifecycle: submitting output, retrieving results, listing history, checking appeals, tracking agent performance, providing feedback, and browsing templates. No obvious gaps exist for the stated purpose of verifying AI-generated content.
Maintenance
Related MCP Connectors
Verify an agent's output before you pay: JSON Schema + rules, signed receipts. Free during launch.
Independent signed math checks for AI agents: PASS or FAIL plus a link anyone can open. Free beta.
Free hosted notary: make any AI agent's actions provable in one call. Verify signed agent logs.
One-stop claim verification for AI agents: verdicts with confidence, evidence, signed receipts.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceCertification authority for AI agents. Register, take adversarial exams, earn cryptographically signed credentials (Ed25519). Get paid to examine other agents. 20,000 free credits on registration — no payment needed to start.MIT
- AlicenseAqualityFmaintenanceScore your agent's governance (0-100), lint MCP tool definitions, and estimate costs across all major models. Free diagnostic tools with no API key needed. Expert skill files on governance, economics, and system architecture available with free tier.81MIT
- AlicenseAqualityDmaintenanceIndependent trust scores, claim audits, and side-by-side comparisons for AI agents, queryable over MCP. Every verdict is backed by hands-on testing and signed evidence from Hlido (hlido.eu). Hosted endpoint available at https://hlido.eu/mcp — no auth required.14MIT
- AlicenseNot gradedqualityBmaintenanceEnables deterministic verification of AI agent decisions and actions, providing PASS/FAIL/ABSTAIN verdicts with replayable proofs and an optional signed receipt ledger.4 npmApache 2.0