Agent Referee
NPM · AGENT-REFEREE-MCP
Neutral tie-breaks between AI agents and sealed review of an agent's output, with signed receipts.
Available components
Last scanned (2 hours ago)
Scores higher than 43% of npm servers
Category AI and Machine Learning
How this component scores in each security and reliability category. Every signal is checked automatically from public evidence about the published package, including repeated runs of it in an isolated sandbox, and we only credit what we can confirm. How we score → Why this is hard to score →
Supply Chain Security100
- No malware found by supply-chain analysis.Pass
- No known CVEs affecting this package version or its production dependencies.Pass
- No install/post-install scripts declared.Pass
- 0 of 1 dependencies flagged as unhealthy. View diagnostics → Pass
Provenance & Transparency45
- Source repository is publicly reachable at the declared URL. View diagnostics → Pass
- Provenance check failed: no build-provenance attestation is published. See how to fix → View diagnostics → Fail
- Clear OSI-approved license (MIT).Pass
- Actively maintained (last published 1 days ago).Pass
- Disclosure check failed: no security disclosure policy was found in the source repository. See how to fix → Fail
Schema Quality & AI Usability60
- AI-judged instruction clarity (excellent).Pass
- Context-footprint check failed: tool/resource definitions use about 2029 tokens (~405/item across 5 items; 5 tools + 0 resources), over budget; trim descriptions and params. See how to fix → Fail
- Usage-examples check failed: none of the tools include examples. See how to fix → Fail
Stability & Change Management27
- Stability observed for 8 of 30 days with no destabilising changes; credit accrues until the full window elapses.Partial
Tool Coverage100
- 100% of tools have a non-trivial description (not blank, and not just the tool's name).Pass
- 100% of tool parameters carry a description.Pass
Tool Safety100
- No prompt-injection markers were found in the server instructions, tool names or descriptions we captured.Pass
- We read all 5 captured tool definition(s), and no name or description among them implies an irreversible operation.Pass
- An AI judge read all 6 captured unit(s) of tool text and found none that tries to manipulate the model reading it.Pass
Capabilities100
- Implements a supported MCP spec version (2025-11-25); the latest is 2026-07-28.Pass
Every change we have recorded for this component, newest first. Security-relevant changes are always shown. ▲ marks a change for the better, ▼ a change for the worse; unmarked changes are neutral.
- 5 Nov 26 +1
No change was recorded against any check on this day. Stability & Change Management went from 23 to 27. That category is still filling its 30-day observation window: 7 days of observed history at the previous scan, 8 at this one. The score rises as the window fills, whether or not the server changes.
- 3 Nov 26 +1
No change was recorded against any check on this day. Stability & Change Management went from 17 to 20. That category is still filling its 30-day observation window: 5 days of observed history at the previous scan, 6 at this one. The score rises as the window fills, whether or not the server changes.
- 1 Nov 26 +1
No change was recorded against any check on this day. Stability & Change Management went from 10 to 13. That category is still filling its 30-day observation window: 3 days of observed history at the previous scan, 4 at this one. The score rises as the window fills, whether or not the server changes.
- 30 Oct 26 +1
No change was recorded against any check on this day. Stability & Change Management went from 3 to 7. That category is still filling its 30-day observation window: 1 days of observed history at the previous scan, 2 at this one. The score rises as the window fills, whether or not the server changes.
- 29 Oct 26 +15
- Tool safety: pass → unverified ▼ security
- Malware scan: unverified → pass ▲ security
- Stability: Stability not yet verified: we do not have a sandbox capture of the MCP schema this version of the package serves yet. security
- Schema quality: 342 → 405 ▼ functional
- Schema quality: 342 → 398 ▼ functional
- Capabilities: pass → unverified ▼ functional
- Tool coverage: 100 → unverified ▼ functional
- Stability: unverified → 0.03 ▲ functional
- First check of Schema quality: unverified functional
- Package version: 0.2.0 → 0.4.0 functional
- Package version: 0.2.0 → 0.3.0 functional
- Package version: 0.2.0 → 0.2.1 functional
- 28 Oct 26 53
First indexed and scored.
Diagnostic detail from the automated scan of this channel: what the scanner observed at each step, so you can see exactly where a check passed or failed. It is informational only and never changes the trust score.
Captured 5 Nov 2026 · Analysed npm/agent-referee-mcp@0.4.0
Provenance No attestation
The registry publishes no build provenance for this version, so there is nothing to verify.
| Result | No attestation |
|---|---|
| Ecosystem | npm |
Background: How many MCP packages publish verified provenance →
Dependencies 1 package
| Packages resolved | 1 |
|---|---|
| Tree resolution | Complete |
Background: SBOMs and build attestations, explained →
The tools this component advertises to a client, with an estimated token cost for each. Expand a tool to see its parameters and schema. The per-tool counts are indicative and are not scored directly; the schema's total context footprint is one signal in Schema Quality & AI Usability. A tool's description is untrusted text the model reads on every call, which is what makes this list a security surface and not just an inventory: how tool poisoning works →
demo_match Try a free demo tie-break ~209
Free, no wallet: runs a complete Agent Referee tie-break in one call, playing both seats, and returns the winner, the signed receipt and the checks. Use it to see how Agent Referee works or to test the connection. It is not a real decision between two agents (use tiebreak for that). Demo limit: 20 per day per IP address.
| Name | Type | Req | Description |
|---|---|---|---|
| seats | array | – | Exactly two seat labels, e.g. ["alpha", "beta"]. |
| tier | string | – | weighted (a fair or weighted coin, the default), rps (one game of rock-paper-scissors, ties replay) or rps-bo3 (best of 3). |
| title | string | – | Optional short label stored in the receipt, e.g. "Who drafts the summary". |
| weights | object | – | weighted tier only: the odds per seat as whole numbers, e.g. {"alpha": 70, "beta": 30}. Default 50/50. |
No output schema declared.
No examples provided.
saved_receipts List saved review receipts ~108
Free and local: lists the sealed-review receipts this server saved on this machine (newest first), or returns one in full by its review_id, plus today's paid spend. Use it to find a receipt again, for example after a review call timed out on your side but finished on the service.
| Name | Type | Req | Description |
|---|---|---|---|
| id | string | – | A review_id to fetch in full. Leave it out to list recent receipts. |
| limit | integer | – | How many recent receipts to list (default 20). |
No output schema declared.
No examples provided.
sealed_review Sealed review of your own output (free daily allowance, then paid) ~855
FREE OR PAID. Free: each address gets 3 single-reviewer reviews and 1 review-3 a day, inside a daily budget shared by everyone (used whenever no wallet is set up, or with free: true). Paid: $0.05, or $0.20 for review-3, in USDC on Base from the wallet the owner configured; charged only if the review completes, and capped per day by the owner. Multi-part (paid only, $0.60 for review-3x2, $0.70 for review-3x3): send the work as 2 or 3 pieces instead of one output, and three reviewers from three companies each read every piece. Independent AI reviewers check one agent's output against the task it was given and return a verdict (pass, needs_changes, fail, cannot_assess; review-3 can also be split), the problems found, and a signed receipt the agent can hand to whoever receives the work. Use it when someone else needs to trust the check, not for routine self-checks. The task and output are sent to the AI companies running the reviewers: leave out secrets and personal data. Can take up to a few minutes. The receipt is saved to disk automatically.
| Name | Type | Req | Description |
|---|---|---|---|
| criteria | array | – | Optional weighted criteria, e.g. [{"name": "accuracy", "weight": 3}]. Each reviewer scores 0-10 per criterion. |
| free | boolean | – | Use one of today's free reviews instead of paying (each address gets 3 single-reviewer and 1 review-3 a day, inside a daily budget shared by everyone, so they can run out). Without a wallet set up, r… |
| output | string | – | The exact work to review (up to 12,000 characters), for review and review-3. The receipt is tied to this exact text. |
| panel | – | – | Optional choice of reviewers. "rotate" picks one model per AI company at random; a list of model ids (1 for review, 3 for review-3) uses exactly those; a list of up to 5 such lists has one drawn at r… |
| pieces | array | – | For review-3x2 (exactly 2 pieces) and review-3x3 (exactly 3), instead of output: the work in parts that belong together, e.g. code and its tests. Up to 36,000 characters in all. Every reviewer reads… |
| task | string | yes | What the work was supposed to do: the instructions or spec it was produced from, with its requirements (10-4,000 characters). Reviewers judge the output against this. It is NOT a message to the revie… |
| tier | string | – | review: one reviewer (by default Claude Sonnet), $0.05. review-3: three reviewers from three AI companies (by default Claude Opus, GPT-6 Luna and Gemini 3.7 Flash), majority verdict, $0.20. review-3x… |
| title | string | – | Optional short label stored in the receipt, e.g. "Who drafts the summary". |
No output schema declared.
No examples provided.
tiebreak Settle a tie with another agent (free demo) ~430
Settle a deadlock with an agent from a different owner, when neither of you should decide alone (who goes first, which of two equal plans). Both agents call this tool with the same key and settings, each with its own my_seat; Agent Referee combines both agents' secret seeds (commit-reveal), so neither side nor the service can steer the result, and both get the same signed receipt. Blocks until the other agent joins (up to wait_seconds) and the result is in. Currently free demo matches only (receipts marked demo); paid tiers are coming next. Don't use it when both agents have the same owner: just decide in code.
| Name | Type | Req | Description |
|---|---|---|---|
| key | string | yes | A name both agents use for this decision, e.g. "task-912-who-drafts". Agree on it (and on every other setting) with the other agent first: the same key and settings put you both in the same match. 3-… |
| move | string | – | rps tiers only: your first move. Replays after a tie (and later games of rps-bo3) use random moves. Leave it out for random. |
| my_seat | string | yes | Your seat: one of the two seats. The other agent takes the other one. |
| seats | array | – | Exactly two seat labels, e.g. ["alpha", "beta"]. |
| tier | string | – | weighted (a fair or weighted coin, the default), rps (one game of rock-paper-scissors, ties replay) or rps-bo3 (best of 3). |
| title | string | – | Optional short label stored in the receipt, e.g. "Who drafts the summary". |
| wait_seconds | integer | – | How long to wait for the other agent to join (default 60). Waiting costs nothing. |
| weights | object | – | weighted tier only: the odds per seat as whole numbers, e.g. {"alpha": 70, "beta": 30}. Default 50/50. |
No output schema declared.
No examples provided.
verify_receipt Check an Agent Referee receipt ~309
Free and offline: checks an Agent Referee receipt (a sealed review or a match) against the service's pinned signing key, recomputes what can be recomputed, and lists any problems. For a sealed review someone handed you, pass the exact task and output you have, so the check confirms the review is about that exact work. Flags demo receipts.
| Name | Type | Req | Description |
|---|---|---|---|
| acks | array | – | Match: the signed step acknowledgments that came with the receipt (optional). Any contradiction between them and the receipt is reported. |
| output | string | – | Sealed review: the exact output you received. Without it, the check can't confirm the review is about the work you have. |
| pieces | – | – | Multi-part review (review-3x2, review-3x3), instead of output: every piece you received, [{name?, text}], with the names and in the order given. Multi-part judge match (judge-3x2, judge-3x3): the pie… |
| receipt | object|string | yes | The signed receipt: the object with payload, signature, algorithm and public_key (or that object as a JSON string). A saved record that contains a receipt works too. |
| task | string | – | Sealed review: the exact task you expect the review to be about. Strongly recommended. |
No output schema declared.
No examples provided.
How do I install the Agent Referee MCP server?
Agent Referee runs locally as an npm package, launched with npx -y agent-referee-mcp. Ready-made configuration for Claude, Cursor, VS Code, Codex and 5 more is on this page, copied from each client's own documentation.
npm · agent-referee-mcp
claude mcp add ccheez34-agent-referee -- npx -y agent-referee-mcp
What is the Agent Referee MCP server?
Agent Referee is an MCP server listed in the public MCP registry as io.github.ccheez34/agent-referee. Neutral tie-breaks between AI agents and sealed review of an agent's output, with signed receipts. This page covers its npm package (agent-referee-mcp).
Is the Agent Referee MCP server safe to use?
Agent Referee scores 72 out of 100 on VerifyMCP. We found no known CVEs affecting it as of 5 November 2026. It declares no install or post-install scripts. That is a record of what we were able to check automatically, not an endorsement. The category breakdown on this page shows every signal behind the number, including the ones we could not confirm.
What tools does the Agent Referee MCP server expose?
Agent Referee exposes 5 tools: demo_match, tiebreak, sealed_review, verify_receipt, saved_receipts. Their descriptions and schemas cost roughly 1,911 tokens of context every time the server is loaded.
Is the Agent Referee MCP server still maintained?
Agent Referee is still listed as active in the MCP registry. We last reached this channel on 5 November 2026. Those dates come from our own scans of the registry and the channel itself, not from anything the publisher announced.
What licence is the Agent Referee MCP server under?
Agent Referee declares the MIT licence, which is OSI-approved. That covers the source only, and says nothing about the cost of any service it calls.