Peer Review MCP Server
by raydollete
README.md
# Peer Review MCP Server
A provider-agnostic MCP (Model Context Protocol) server that asks a **weighted quorum of external LLM peers** the same question and returns a consensus answer with a machine-consumable **certainty score**. It supports any OpenAI-compatible or Anthropic-compatible API, stacks sources in cost tiers, and degrades visibly (never silently) when a source is unavailable.


## Features
- **Weighted quorum with a certainty score** — peers vote with configurable weights; every `peer_review` response carries a machine-consumable `certaintyScore`.
- **Provider-agnostic** — two wire formats (OpenAI `chat/completions`, Anthropic `v1/messages`) cover OpenAI, Gemini, OpenRouter, Ollama, vLLM, Anthropic, and most gateways.
- **Cost tiers with escalation** — tier 1 sources are consulted first; higher tiers only fan out on quorum shortfall.
- **Visible degradation** — failed or unavailable sources are itemized in the result, never silently dropped; an unmet quorum is a reported outcome, not an error.
- **No vendor SDKs** — all providers are reached over plain HTTPS `fetch`.
## Table of Contents
- [Features](#features)
- [Requirements](#requirements)
- [Installation](#installation)
- [Quick Start](#quick-start)
- [Configuration](#configuration)
- [Tools](#tools)
- [Troubleshooting](#troubleshooting)
- [Migrating from gemini-for-claude-mcp](#migrating-from-gemini-for-claude-mcp)
- [Development](#development)
- [Contributing](#contributing)
- [License](#license)
## Requirements
- Node >= 20
## Installation
```bash
git clone https://github.com/raydollete/peer-review.git && cd peer-review
npm install
npm run build # emits dist/app.js
```
## Quick Start
Create a config file (default path `./peer-review.config.json`):
```jsonc
{
"thresholds": { "tier1": 2, "tier2": 4 }, // weight targets — see "How tiers, weights, and thresholds interact"
"arbiter": "gpt-large", // source name used for consensus evaluation
"sources": [
{ "name": "gpt-large", "apiType": "openai", "baseUrl": "https://api.openai.com/v1", "model": "gpt-5.2", "apiKeyEnv": "OPENAI_API_KEY", "weight": 2, "tier": 1, "timeoutMs": 60000 },
{ "name": "claude", "apiType": "anthropic", "baseUrl": "https://api.anthropic.com", "model": "claude-fable-5", "apiKeyEnv": "ANTHROPIC_API_KEY", "weight": 2, "tier": 2, "timeoutMs": 120000 }
]
}
```
Then register the server in Claude Code:
```bash
claude mcp add peer-review \
--env PEER_REVIEW_CONFIG=/absolute/path/to/peer-review.config.json \
--env OPENAI_API_KEY=sk-... \
-- node /absolute/path/to/peer-review/dist/app.js
```
## Configuration
The server reads a JSON config file from `PEER_REVIEW_CONFIG` (default `./peer-review.config.json`). See `peer-review.config.example.json`.
### Config file fields
| Field | Type | Required | Description |
|---|---|---|---|
| `thresholds` | object | yes | Map of tier → positive integer weight target. Keys may be `"tier1"` or `"1"`. Every tier used by a source must have a threshold (validated at startup). |
| `arbiter` | string | yes | Name of the source used to evaluate agreement between peer responses. Must match a `sources[].name`. |
| `sources[].name` | string | yes | Unique source name (used in `query_peer`/`count_tokens` and in quorum reports). |
| `sources[].apiType` | `"openai"` \| `"anthropic"` | yes | Wire format: `openai` → `POST {baseUrl}/chat/completions`; `anthropic` → `POST {baseUrl}/v1/messages`. |
| `sources[].baseUrl` | URL | yes | API base URL. For `openai`, include the version segment (e.g. `https://api.openai.com/v1`). |
| `sources[].model` | string | yes | Model id sent to the provider. Never client-suppliable. |
| `sources[].apiKeyEnv` | string | exactly one of these two | Name of the environment variable holding a static API key. If the variable is unset the source is marked unavailable (the server still starts). |
| `sources[].apiKeyCommand` | string | exactly one of these two | Command executed (via `sh -c`) to mint a short-lived bearer token, e.g. `gcloud auth print-access-token` for Vertex AI. Output is trimmed and cached (see `PEER_CREDENTIAL_TTL_S`); a 401 triggers one re-mint + retry. |
| `sources[].weight` | positive int | yes | Weight this source contributes to quorum when it agrees with the consensus. |
| `sources[].tier` | positive int | yes | Cost tier. Tier 1 is consulted first; higher tiers only on shortfall. |
| `sources[].local` | bool | no | Marks a source served by the local inference backend. A caller that itself runs on that backend sets [`PEER_REVIEW_SKIP_LOCAL`](#environment-variables) so these sources are dropped from its requests (self-contention, not a second opinion). Locality, not model id. |
| `sources[].jsonMode` | bool | no | Asks the provider for a JSON-object reply on arbiter evaluations via `response_format`. OpenAI-compatible sources only — startup rejects it on `anthropic`. Off = request bodies unchanged. Set it on the arbiter source once you have confirmed its endpoint accepts `response_format`. |
| `sources[].maxOutputTokens` | int | no | Per-source override of `PEER_MAX_OUTPUT_TOKENS`, sent as `max_tokens`. Reasoning models bill hidden thinking against it before the answer, so give them a budget above their thinking length (32768 is a safe Anthropic value) and raise `timeoutMs` with it. A reply stopped at the ceiling fails the source with `cause: peer_truncated` rather than voting half-written text. |
Startup fails fast on: an unreadable/invalid config file, an unknown `arbiter` name, a source tier without a threshold, duplicate source names, a source named `caller` (reserved for the [`callerAnswer`](#peer_review) rating channel), a source declaring both or neither of `apiKeyEnv`/`apiKeyCommand`, a source declaring `jsonMode` with `apiType: "anthropic"`, or a timeout budget that cannot fit the first tier's wave plus its arbiter evaluation inside `PEER_REVIEW_DEADLINE_MS`. The first wave is sized to its slowest configured peer, plus a separate arbiter allowance, so a first verdict is always reachable. Higher tiers are not summed: at run time each tier's wave is capped at the time remaining minus one arbiter evaluation, and a tier with no time left is not launched and is itemized `skipped`. Credential material never appears in config, logs, or responses.
### Environment variables
| Variable | Default | Description |
|---|---|---|
| `PEER_REVIEW_CONFIG` | `./peer-review.config.json` | Path to the config file. |
| `PEER_TIMEOUT_MS` | `120000` | Per-source request timeout (includes retries). |
| `PEER_REVIEW_DEADLINE_MS` | `280000` | Hard overall deadline for one `peer_review` call; on expiry in-flight calls are aborted and the accumulated (possibly degraded) result is returned. MUST stay strictly below the declared `PEER_REVIEW_CALLER_TIMEOUT_MS` (next row) — above it the caller abandons the call at its own ceiling and reports only its generic timeout, so the partial result arrives after the caller has given up and no verdict anyone computed is read. Each tier's peer wave gets the smaller of its configured timeout and the time remaining minus the arbiter timeout. Also bounds `query_peer`. |
| `PEER_REVIEW_CALLER_TIMEOUT_MS` | `300000` | Declared MCP client tool-call timeout (ms) — the ceiling this server assumes. The harness registration owner sets the real value (e.g. omp's per-server `timeout` key); `PEER_REVIEW_DEADLINE_MS` must stay strictly below it. A deadline at or above it makes the server warn once at startup and report the mismatch (`cause: configuration`) on `list_peers` and every failure envelope while it holds. |
| `PEER_REVIEW_SKIP_LOCAL` | `false` | When true (`true`/`1`), every source flagged [`local`](#configuration) is excluded from the quorum for this server registration. Set it per-caller in that caller's MCP registration `env` when the caller itself runs on the local backend, so it does not contend with itself. The class-wide sibling of the per-request [`callerSource`](#peer_review) skip. |
| `PEER_MAX_OUTPUT_TOKENS` | `8192` | Max output tokens injected server-side into every peer call; a source may override it with [`sources[].maxOutputTokens`](#config-file-fields). Hidden reasoning is billed against this budget, so a reasoning model needs a budget above its thinking length or it answers with nothing and the source fails `cause: peer_truncated`. |
| `PEER_CREDENTIAL_TTL_S` | `3000` | Cache TTL (seconds) for tokens minted via `apiKeyCommand`. |
| `LOG_LEVEL` | `info` | pino log level (`fatal`\|`error`\|`warn`\|`info`\|`debug`\|`trace`). Logs go to stderr only. |
| `NODE_ENV` | — | `production` switches the logger from pretty-printing to JSON. |
### How tiers, weights, and thresholds interact
A `peer_review` call is a weighted vote. The config decides who votes, in what order, and how many votes win:
- **`weight`** — how much a source's agreement counts. Give larger weights to sources you trust more.
- **`tier`** — cost band. Tier 1 is consulted on every call; tier 2 only when tier 1 falls short; and so on. Put cheap, fast models in tier 1 and expensive frontier models above.
- **`thresholds`** — the agreeing weight required to declare quorum, keyed by *target* tier.
The one rule that is easy to miss:
> **A call has a single required weight: the threshold of its *target* tier.** The target tier is the highest configured tier unless the caller caps escalation with the `tier` input. Lower tiers' thresholds are not applied along the way — `thresholds.tier1` only ever binds a call made with `"tier": 1`.
So with the Quick Start config (`{"tier1": 2, "tier2": 4}`), a default call must accumulate agreeing weight **4** (tier 2's threshold). Escalation is lazy, not mandatory: if tier 1's agreeing weight alone reaches 4, tier 2 is never contacted.
Execution order for one call:
1. All sources of the lowest tier are queried **in parallel**.
2. The arbiter (at temperature 0) rates each response 0–1 against its own consensus answer. A source **agrees** when its rating is ≥ 0.7; only agreeing sources' weights count. Evaluation is triggered mid-tier as soon as the responses in hand could arithmetically meet the required weight — and again at the end of every tier, whether or not the required weight was reachable, so the accumulated answers are always rated before escalating.
3. If agreeing weight ≥ required weight, quorum is achieved: any still-in-flight requests are aborted (cost saved) and the arbiter's consensus answer is returned.
4. Otherwise the next tier fans out and the arbiter re-rates over **all** accumulated responses — earlier answers are never discarded.
5. Tiers that were never reached do not appear in `quorum.sources[]` at all.
**The time budget (the reserve model).** `sources[].timeoutMs` is a per-request ceiling, and the run budgets the tiers that escalation walks in two stages. **Startup hard-checks the first tier only:** its wave (its slowest source's effective timeout — an unset `timeoutMs` means the `PEER_TIMEOUT_MS` default, 120 000 ms) plus one arbiter evaluation must fit strictly inside `PEER_REVIEW_DEADLINE_MS`, so a first verdict is always reachable. **Later tiers reserve at entry:** when escalation reaches a tier, the run reserves one arbiter evaluation and clamps that tier's wave to the remaining budget (the smaller of the tier's configured timeout and what the reserve leaves). A tier that cannot start is not billed and not faked as down: it is itemized as `skipped` with 'Skipped: no time left before the deadline after the arbiter reserve'. The global deadline (the wall) is the backstop that bounds the whole call. The practical rule for the tiers escalation walks: the earlier tiers' slowest timeouts plus one arbiter evaluation each must leave the escalated tier time to launch — budget them against `PEER_REVIEW_DEADLINE_MS` and recompute whenever you change it (the shipped `peer-review.config.example.json` budgets 60 s + 60 s before a tier-2 entry under the 280 s default). A higher tier's own `timeoutMs` may exceed the remaining budget — that value is also its `query_peer`/preflight ceiling, and the quorum path clamps the wave at run time.
Sizing guidance:
- **Overprovision each tier.** A tier whose potential weight exactly equals the threshold escalates the moment any one source fails or dissents. Three weight-1 sources against a threshold of 2 tolerate one outlier; two sources tolerate none.
- **Prefer model diversity within a tier.** Two copies of the same model mostly vote together, so their combined weight overstates the independence of the "second opinion".
- **Keep the required weight reachable.** The target tier's threshold must be ≤ the summed weight of all sources in tiers ≤ target, or `quorum.achieved` can never be true.
- **Let lower tiers be able to win.** A tier can only short-circuit a call if the weight accumulated *through* that tier can reach the target tier's threshold. If it can't, that tier is pure overhead on default calls: its peers are billed, the end-of-tier arbiter rating is billed, and escalation happens regardless. With `{"tier1": 2, "tier2": 3}` over two weight-1 tier-1 sources, a default call can only ever reach weight 2 at tier 1 against a required 3 — so tier 1 never saves anything. Either raise the tier-1 weights, add a tier-1 source, or lower `tier2` so the cheap tier has a path to quorum. A tier that can't win is still useful if you want its answers *in the mix* for the later arbiter pass — just size it knowing that's what you're paying for.
- **Match tier to speed, and isolate slow reasoners.** A `tier` is a *cost/latency class*, not a redundancy slot: a tier's peers race, and the instant the accumulated agreeing weight meets the target threshold, every still-in-flight peer in that tier is aborted (step 3 above). So a slow reasoning model sharing a tier with a faster one is cancelled before it can answer *every time the faster one already suffices* — it never contributes, and it surfaces in `quorum.sources[]` as `cancelled`. Give a slow reasoner its **own** higher tier as the sole member, so it is awaited and folded into the re-rating when escalation reaches it. If you must place several reasoners in one tier, set that tier's threshold equal to the summed weight of all its members, so no faster subset can meet the bar and abort the rest.
- **Span providers, not just models.** Cross-*provider* diversity beats cross-model diversity within one provider: models from the same lab share training data and failure modes, so their agreement overstates the independence of the second opinion. Draw a tier's voters from different providers where you can.
- **Never place a `local` source alone in your top tier.** A caller running on the local backend sets `PEER_REVIEW_SKIP_LOCAL`, which drops every `local` source from the quorum. If a `local` model is the sole member of the escalation tier, that tier is empty for that caller and quorum can never escalate into it.
- **Pick a cheap, reliable arbiter.** The arbiter is called at least once per `peer_review` call plus once per escalation, and its tokens count toward `tokenUsage`. It can double as a regular peer, but note that it then rates its own answer alongside the others — check how much of your threshold that self-rating can supply, and prefer an arbiter that isn't also one of your heaviest voters. If the arbiter's own call fails, the whole evaluation falls back to `arbiterFailed: true` with `certaintyScore: 0` — so give the arbiter role to the source with your most dependable credential.
### Example: a three-tier stack
The walkthroughs below all use this config — two cheap flash-class voters in tier 1, a mid-price model in tier 2, a frontier model in tier 3:
```jsonc
{
"thresholds": { "tier1": 2, "tier2": 3, "tier3": 4 },
"arbiter": "gpt-mini",
"sources": [
{ "name": "gemini-flash", "apiType": "openai", "baseUrl": "https://generativelanguage.googleapis.com/v1beta/openai", "model": "gemini-3-flash-preview", "apiKeyEnv": "GEMINI_API_KEY", "weight": 1, "tier": 1, "timeoutMs": 45000 },
{ "name": "gpt-mini", "apiType": "openai", "baseUrl": "https://api.openai.com/v1", "model": "gpt-5-mini", "apiKeyEnv": "OPENAI_API_KEY", "weight": 1, "tier": 1, "timeoutMs": 45000 },
{ "name": "gemini-pro", "apiType": "openai", "baseUrl": "https://generativelanguage.googleapis.com/v1beta/openai", "model": "gemini-3-pro", "apiKeyEnv": "GEMINI_API_KEY", "weight": 2, "tier": 2, "timeoutMs": 60000 },
{ "name": "claude", "apiType": "anthropic", "baseUrl": "https://api.anthropic.com", "model": "claude-fable-5", "apiKeyEnv": "ANTHROPIC_API_KEY", "weight": 2, "tier": 3, "timeoutMs": 120000 }
]
}
```
A default call targets tier 3, so it needs agreeing weight **4**. A `"tier": 1` call needs only **2**. Note the arbiter is `gpt-mini`, deliberately *not* one of the Gemini sources: if `GEMINI_API_KEY` ever goes missing, evaluation still works (see scenario 4).
### What happens when a query comes in
#### 1. Easy factual query — quorum at tier 2, tier 3 never billed
`{ "prompt": "What does HTTP status 431 mean? One sentence." }`
Tier 1 fans out to `gemini-flash` and `gpt-mini`; both agree, but their combined weight (2) can't reach the required 4, so tier 2 is consulted. `gemini-pro` agrees too: 1 + 1 + 2 = 4 ≥ 4 — quorum. `claude` is never called and never appears in `sources[]`.
```jsonc
// data
{
"response": "HTTP 431 (Request Header Fields Too Large) means the server refused the request because its headers are too big.",
"certaintyScore": 0.98, // min(1, 4/4) × mean(1.0, 1.0, 0.95), rounded
"quorum": {
"achieved": true,
"tier": 2, // tier reached; requiredWeight stays the target tier's threshold
"requiredWeight": 4,
"agreeingWeight": 4,
"sources": [
{ "name": "gemini-flash", "model": "gemini-3-flash-preview", "status": "ok", "weight": 1, "agreement": 1 },
{ "name": "gpt-mini", "model": "gpt-5-mini", "status": "ok", "weight": 1, "agreement": 1 },
{ "name": "gemini-pro", "model": "gemini-3-pro", "status": "ok", "weight": 2, "agreement": 0.95 }
]
},
"tokenUsage": { "prompt": 812, "completion": 118, "total": 930 } // includes both arbiter rounds
}
```
#### 2. Low-stakes check — cap escalation with `"tier": 1`
`{ "prompt": "Is 'reciept' spelled correctly?", "tier": 1 }`
Now the required weight is tier 1's threshold (2). Both flash sources agree, quorum is achieved in one round, and tiers 2–3 stay uncontacted. This is the cheap path for queries where two small models agreeing is good enough. The ceiling holds when every tier-1 source answers — a disagreement-shaped miss is revisable, so escalating it would only cost more. An outage-shaped miss is not: if one tier-1 source errors and the attainable weight (the summed weight of sources that answered) falls below tier 1's required weight, the run escalates to the next configured tier to restore the capacity to reach the bar, while `requiredWeight` stays tier 1's 2.
```jsonc
// data
{
"response": "No — the correct spelling is \"receipt\".",
"certaintyScore": 0.95, // min(1, 2/2) × mean(1.0, 0.9)
"quorum": {
"achieved": true,
"tier": 1,
"requiredWeight": 2,
"agreeingWeight": 2,
"sources": [
{ "name": "gemini-flash", "model": "gemini-3-flash-preview", "status": "ok", "weight": 1, "agreement": 1 },
{ "name": "gpt-mini", "model": "gpt-5-mini", "status": "ok", "weight": 1, "agreement": 0.9 }
]
},
"tokenUsage": { "prompt": 296, "completion": 41, "total": 337 }
}
```
#### 3. Contested question — an outlier forces the full stack
`{ "prompt": "In Node 20, is structuredClone faster than a JSON round-trip for small flat objects?" }`
`gemini-flash` answers the opposite of everyone else and the arbiter rates it 0.2 — below the 0.7 agreement bar, so its weight is excluded (but still itemized). Agreeing weight climbs 1 → 3 → 5 as tiers 2 and 3 are pulled in, with the arbiter re-rating the accumulated set each round; quorum lands at tier 3.
```jsonc
// data
{
"response": "For small flat objects a JSON round-trip is usually still faster in Node 20; structuredClone wins once values contain types JSON can't represent.",
"certaintyScore": 0.9, // min(1, 5/4) × mean(0.95, 0.9, 0.85) — outlier's 0.2 not averaged in
"quorum": {
"achieved": true,
"tier": 3,
"requiredWeight": 4,
"agreeingWeight": 5, // gpt-mini 1 + gemini-pro 2 + claude 2; gemini-flash excluded
"sources": [
{ "name": "gemini-flash", "model": "gemini-3-flash-preview", "status": "ok", "weight": 1, "agreement": 0.2 },
{ "name": "gpt-mini", "model": "gpt-5-mini", "status": "ok", "weight": 1, "agreement": 0.95 },
{ "name": "gemini-pro", "model": "gemini-3-pro", "status": "ok", "weight": 2, "agreement": 0.9 },
{ "name": "claude", "model": "claude-fable-5", "status": "ok", "weight": 2, "agreement": 0.85 }
]
},
"tokenUsage": { "prompt": 2140, "completion": 486, "total": 2626 }
}
```
#### 4. Degraded run — missing credential, quorum unreachable
Same easy prompt as scenario 1, but `GEMINI_API_KEY` is unset, taking out both `gemini-flash` and `gemini-pro`. The remaining potential weight (`gpt-mini` 1 + `claude` 2 = 3) can never reach 4, so the call completes with `quorum.achieved: false` — still `success: true`, never an error. The degradation is itemized, not hidden.
```jsonc
// data
{
"response": "HTTP 431 (Request Header Fields Too Large) means the server refused the request because its headers are too big.",
"certaintyScore": 0.73, // min(1, 3/4) × mean(1.0, 0.95), rounded — shortfall caps the score
"quorum": {
"achieved": false,
"tier": 3,
"requiredWeight": 4,
"agreeingWeight": 3,
"sources": [
{ "name": "gemini-flash", "model": "gemini-3-flash-preview", "status": "unavailable", "weight": 1, "agreement": null,
"error": "Credential not resolvable; source skipped", "cause": "credential", "code": "CONFIGURATION_ERROR" },
{ "name": "gpt-mini", "model": "gpt-5-mini", "status": "ok", "weight": 1, "agreement": 1 },
{ "name": "gemini-pro", "model": "gemini-3-pro", "status": "unavailable", "weight": 2, "agreement": null,
"error": "Credential not resolvable; source skipped", "cause": "credential", "code": "CONFIGURATION_ERROR" },
{ "name": "claude", "model": "claude-fable-5", "status": "ok", "weight": 2, "agreement": 0.95 }
]
},
"tokenUsage": { "prompt": 640, "completion": 97, "total": 737 }
}
```
Had the arbiter itself been the casualty (e.g. `OPENAI_API_KEY` missing in this config), no agreement could be evaluated at all and the response would instead carry the highest-weighted successful answer with `"certaintyScore": 0`, `"quorum": { "arbiterFailed": true, ... }`, and a `quorum.arbiter` row `{ "name": …, "status": "failed", "error": { "code": …, "cause": "arbiter_unavailable", "detail": … } }` naming what its failure was. If instead **no source responded at all**, the call is the one `success: false` shape: `error.code: "EXTERNAL_SERVICE_ERROR"`, `error.message: "No peer source produced a response"`, and — since this change — `error.cause` naming the fault class (`credential` when every source is credential-unavailable, `peer_unreachable` when every host refused, `configuration` when every source was skipped) plus the full per-source roster in `error.context.sources[]`; see the [401 walkthrough](#walkthrough-401-from-every-source).
#### 5. Checking your own answer — `callerAnswer` the peers disagree with
`{ "prompt": "In JavaScript, does [10, 9, 100].sort() return [9, 10, 100]?", "callerAnswer": "Yes, sort() orders numbers ascending by default." }`
You already have an answer and want independent endorsement. The peers never see `callerAnswer` — they answer the prompt blind. Tier 1's two flash voters agree with each other but can't reach weight 4, so tier 2 is pulled in; `gemini-pro` agrees too and quorum lands at tier 2 (1 + 1 + 2 = 4). The arbiter separately rates your answer against the peer consensus and returns `callerAgreement` — low here, because default `sort()` is lexicographic (`[10, 100, 9]`), so your answer contradicts what the peers concluded.
```jsonc
// data
{
"response": "No — Array.prototype.sort() coerces elements to strings by default, so [10, 9, 100].sort() returns [10, 100, 9]. Pass a comparator, e.g. .sort((a, b) => a - b), for numeric order.",
"certaintyScore": 0.98, // peers-only: min(1, 4/4) × mean(1.0, 1.0, 0.95) — unaffected by callerAnswer
"callerAgreement": 0.1, // arbiter rated your answer against the peer consensus; it contradicts it
"quorum": {
"achieved": true, // the peers reached quorum with each other; your answer carries no weight
"tier": 2,
"requiredWeight": 4,
"agreeingWeight": 4,
"sources": [
{ "name": "gemini-flash", "model": "gemini-3-flash-preview", "status": "ok", "weight": 1, "agreement": 1 },
{ "name": "gpt-mini", "model": "gpt-5-mini", "status": "ok", "weight": 1, "agreement": 1 },
{ "name": "gemini-pro", "model": "gemini-3-pro", "status": "ok", "weight": 2, "agreement": 0.95 }
]
},
"tokenUsage": { "prompt": 934, "completion": 141, "total": 1075 }
}
```
`quorum.achieved: true` with `callerAgreement: 0.1` is the signal to catch: the peers are confident, but **not** with you — treat it as a flag to revise your answer. Had you been right, `callerAgreement` would sit near 1 alongside the same high `certaintyScore`. Note `caller` never appears in `sources[]` and its rating never moved `agreeingWeight`.
**Reading the result**: gate any automated decision on `quorum.achieved`, treat `certaintyScore` as the confidence dial (it discounts both shortfall and lukewarm agreement), scan `sources[]` for `error`/`unavailable` entries — a degraded quorum is always visible, never silent; every rated row carries `reason`, the arbiter's one-line explanation of its `agreement`; `agreement: 0` means the arbiter rated that answer as contradicting the consensus, while an `ok` row with `agreement: null` was not rated at all, and its `note` says why (the arbiter returned no rating for it, the arbiter failed, or the run ended before the arbiter rated it) — when a later arbiter pass fails, rows an earlier pass rated keep that rating with a `note` saying so; a `cancelled` row is a peer the run no longer needed once quorum was met, not a failure — its reason lives in `note` and it carries `weight: 0`, so do not count it as degradation — and, when you passed a `callerAnswer`, read `callerAgreement` as the separate "did the peers land where I did?" signal (scenario 5).
**A failed quorum withholds its prose.** When `quorum.achieved` is `false`, `response` comes back as `""` and the envelope carries `"responseWithheld": true` — text no quorum validated is unvalidated opinion, so the server no longer hands back synthesized consensus (or one peer's representative answer) for a caller to fold into its own reasoning. Everything needed to transcribe the verdict honestly is still returned: `quorum` with `requiredWeight`, `agreeingWeight`, and the full `sources[]` roster — including each rated source's one-line `reason`, so you can see why a peer scored low without its prose — `certaintyScore`, and `callerAgreement` when a `callerAnswer` was supplied. An achieved quorum is unchanged — the same `response` prose as before, with no `responseWithheld` key at all.
## Tools
All tools return the envelope `{"success": true, "data": ...}` or `{"success": false, "error": {"code", "message"}}`, JSON-stringified into a single text content block (compatible with the reference Gemini MCP). Error codes: `VALIDATION_ERROR`, `CONFIGURATION_ERROR`, `TIMEOUT_ERROR`, `EXTERNAL_SERVICE_ERROR`, `PEER_API_ERROR`, `PEER_RATE_LIMIT`, `TOOL_NOT_FOUND`, `INTERNAL_ERROR`. Model, temperature, and output limits are injected server-side — the input schemas are strict, so client-supplied `model`/`temperature` are rejected.
**Failure envelopes carry the cause, not just the code.** Alongside `code` and `message` (unchanged), every failure the server can attribute carries `error.cause` from the closed vocabulary in the [Troubleshooting table](#troubleshooting), and where applicable `error.source` (the peer or arbiter at fault), `error.detail` (one provider-agnostic line — HTTP status or the unwrapped transport `errno`, never provider body text or credential material), `error.fix` (the operator action), and, for a `peer_review` failure with at least one source consulted, `error.context` — the same weight math and `sources[]` roster the success path puts in `data.quorum`, so the no-success response is as diagnosable as the success one. An `INTERNAL_ERROR` is the redacted catch-all: its message is always the generic line, and its `error.context.correlationId` points at the full error in the server's stderr log (see the Troubleshooting table). `code` stays the coarse wire class for readers that switch on it; `cause` is what a repair guide should read. When a source fails as the arbiter, `data.quorum.arbiter` (on the arbiter-fallback success path) or `error.context.arbiter` (on the failure path) reports `{name, status: "failed", error: {code, cause, detail}}` — the `certaintyScore: 0` / `arbiterFailed: true` flags keep their exact meaning; a row is added, never moved.
**Every tool advertises MCP `annotations`.** All four declare `readOnlyHint: true` — none of them modifies the caller's environment — and `openWorldHint: true`, except `list_peers` (`openWorldHint: false`, it reads configuration only). A harness that gates concurrency on `readOnlyHint` (Claude Code does) may therefore run several calls issued in one message at once. Inputs, outputs, and quorum behavior are unchanged.
**Changes to `peer-review.config.json` require restarting the server** — it is read once at startup, and credential availability is evaluated at boot.
**Shipped envelope changes require a rebuild and a harness restart** — the harness spawns a long-lived child from `dist/app.js` (`npm run build` emits it), so a new `error.cause`/`fix` envelope reaches callers only after a rebuild and a harness restart or re-registration.
### `peer_review`
Weighted-quorum consultation. Input: `prompt` (1–100000 chars), optional `history` (`[{role: "user"|"model", content}]`), optional `tier` (highest tier to escalate to; defaults to the full stack), optional `callerAnswer` (1–100000 chars — your own answer to the prompt), optional `callerSource` (the name of a configured source the caller itself *is* — skipped as a redundant peer for this request, never called, itemized as `status: "skipped"`; an unknown name is a logged no-op). An unmet quorum is **not** an error — check `quorum.achieved` and `certaintyScore`.
```jsonc
// request
{ "prompt": "What is the capital of France? Answer with just the city name." }
// response data
{
"response": "Paris",
"certaintyScore": 1, // min(1, agreeingWeight/requiredWeight) × mean agreement of agreeing sources
"quorum": {
"achieved": true,
"tier": 1,
"requiredWeight": 2,
"agreeingWeight": 2,
"sources": [
{ "name": "gemini-a", "model": "google/gemini-3.5-flash", "status": "ok", "weight": 1, "agreement": 1 },
{ "name": "gemini-b", "model": "google/gemini-3.5-flash", "status": "ok", "weight": 1, "agreement": 1 }
]
},
"tokenUsage": { "prompt": 315, "completion": 45, "total": 360 } // aggregated across peers + arbiter
}
```
A third, slower tier-1 peer that was still answering when `gemini-a` and `gemini-b` met the bar would appear alongside them as:
```jsonc
{ "name": "slow-reasoner", "model": "glm-flash", "status": "cancelled", "weight": 0, "agreement": null, "note": "Cancelled: quorum was met before this source responded" }
```
`sources[].status` is one of:
- `ok` — the source answered and the arbiter rated it.
- `error` — the source was consulted and failed.
- `unavailable` — never called because its credential was unresolvable.
- `skipped` — excluded from this request via `callerSource`/`PEER_REVIEW_SKIP_LOCAL`; never called.
- `cancelled` — still in flight when quorum was met (the early-abort cancelled it) or the caller cancelled the call; a non-failure, not a peer defect — its reason lives in `note`.
`weight` is the source's effective contribution to the run: `0` on a `cancelled` row (a peer the run no longer needed contributed no vote), otherwise its configured weight — so `skipped` and `unavailable` keep theirs. `note` appears only on a `cancelled` row and carries why it was cancelled (e.g. "Cancelled: quorum was met before this source responded"); it lives outside `error`, so a reader counting failures off `error` MUST NOT count these rows. `agreement` is the arbiter's 0–1 rating (`null` if the source produced no rated response). If the arbiter itself fails, the response falls back to the highest-weighted successful answer with `certaintyScore: 0` and `quorum.arbiterFailed: true`.
**Including your own answer (`callerAnswer`).** When you already have an answer and want to know whether independent peers endorse it, pass it as `callerAnswer` instead of pasting it into `prompt`. Peers never see it — they answer the prompt blind, so their independence is preserved and `certaintyScore` stays an anchoring-free signal. The arbiter is instructed to derive the consensus **only** from the peer answers and then rate your answer against that consensus, so a confidently-wrong `callerAnswer` cannot pull the consensus toward itself. The response gains a `callerAgreement` field (0–1).
How it behaves:
- **Zero quorum weight.** `callerAgreement` is reported alongside the quorum but never counted into `agreeingWeight`, `quorum.achieved`, or `certaintyScore` — your own answer cannot self-certify. Read it as a separate signal: "did independent peers land where I did?"
- **Present only when supplied.** The key is omitted entirely from the response when `callerAnswer` was not passed. When it *was* passed, the key is always present.
- **`null` when unrated.** `callerAgreement` is `null` if the arbiter returned no rating for it — either because the arbiter omitted the `caller` entry from its ratings, or because there was no trusted evaluation at all (arbiter failed, or quorum ran with no ratable peer response — the same fallback paths that force `certaintyScore: 0`). A supplied `callerAnswer` therefore always yields the key, but not always a number.
- **Reserved name.** `caller` is reserved as the internal rating channel and is rejected as a configured `sources[].name` at startup.
```jsonc
// request
{ "prompt": "What is the capital of France?", "callerAnswer": "Paris" }
// response data — same shape as a normal peer_review, plus callerAgreement:
{
"response": "Paris",
"certaintyScore": 1, // unchanged by callerAnswer — derived from peers only
"callerAgreement": 1, // arbiter's 0–1 rating of "Paris" vs the peer consensus
"quorum": { "achieved": true, "requiredWeight": 2, "agreeingWeight": 2, /* … */ },
"tokenUsage": { "prompt": 315, "completion": 45, "total": 360 }
}
```
When your answer diverges from what the peers conclude, `callerAgreement` drops toward 0 while `certaintyScore` still reflects the peers' own convergence — e.g. `{ "callerAgreement": 0.15, "certaintyScore": 0.97 }` reads as "the peers strongly agree, but not with you."
### `query_peer`
Single-source escape hatch mirroring the reference `query_gemini`. Input: `prompt`, optional `history`, optional `source` (defaults to the lowest-tier highest-weight source).
```jsonc
// request
{ "prompt": "In one word, what color is the sky on a clear day?", "source": "gemini" }
// response data
{ "response": "Blue", "model": "google/gemini-3.5-flash", "source": "gemini", "finishReason": "stop",
"tokenUsage": { "prompt": 14, "completion": 1, "total": 15 } }
```
### `list_peers`
No input. Returns every configured source with availability (`available` = credential resolvable):
```jsonc
{ "sources": [ { "name": "gpt-large", "apiType": "openai", "model": "gpt-5.2", "weight": 2, "tier": 1, "available": true } ], "count": 1 }
```
While `PEER_REVIEW_DEADLINE_MS` is not strictly below the declared `PEER_REVIEW_CALLER_TIMEOUT_MS` — the MCP client tool-call timeout this server assumes — the response also carries `callerCeiling: { "deadlineMs": …, "callerTimeoutMs": …, "cause": "configuration", "fix": "…" }` naming both numbers and the two ways to fix the pair. Above the ceiling the caller abandons the call and reports only its own generic timeout, so no verdict anyone computed is read.
### `count_tokens`
Input: `text` (1–1000000 chars), optional `source`. Anthropic-compatible sources use the provider's count endpoint (`"method": "api"`); OpenAI-compatible sources use a `ceil(chars/4)` heuristic (`"method": "estimate"`).
```jsonc
{ "totalTokens": 42, "model": "claude-fable-5", "source": "claude", "method": "api" }
```
## Troubleshooting
### Failure causes
Every attributable failure names `error.cause` from this closed set. Read `cause`, not prose.
| cause | meaning | what to do |
| --- | --- | --- |
| `peer_unreachable` | the peer's host refused or failed the connection; `error.detail` names the unwrapped `errno` (`ECONNREFUSED`, `ENOTFOUND`, …) | check the source's `baseUrl` and network reachability from the server host, then retry |
| `credential` | 401/403 from the provider, or the credential is unset/unmappable | set the credential env var or repair `apiKeyCommand` for the named source in `peer-review.config.json`, then restart the server |
| `rate_limited` | the provider 429'd through its backoff budget | wait before retrying, or lower the weight/tier of this source |
| `peer_bad_response` | 404/5xx, or a body that is not valid JSON | check the model name and endpoint for the named source in `peer-review.config.json`, then retry once |
| `peer_truncated` | the provider stopped the reply at its own output ceiling (`stop_reason: "max_tokens"` / `finish_reason: "length"`), or returned no answer text inside it — the common shape for a reasoning model whose hidden thinking is billed against `max_tokens`; `error.detail` names the stop signal and the billed `completion` tokens | raise `maxOutputTokens` for the named source (or `PEER_MAX_OUTPUT_TOKENS` globally) so the budget clears the model's thinking length, and raise `timeoutMs` with it |
| `arbiter_unavailable` | the arbiter source could not be reached | check the arbiter source reachability and credentials in `peer-review.config.json` |
| `arbiter_malformed` | the arbiter answered in prose; JSON extraction failed even after the one re-ask | set `jsonMode` for the arbiter source or choose a JSON-reliable arbiter |
| `deadline` | the per-source `timeoutMs` or the global wall expired | raise `PEER_REVIEW_DEADLINE_MS` (staying strictly below `PEER_REVIEW_CALLER_TIMEOUT_MS`) or lower the per-source `timeoutMs` |
| `validation` | a tool argument failed schema validation; `error.fix` names the field | correct the named argument and retry |
| `configuration` | bad config file or env var, or the ceiling pair is misset | fix the named file or variable, then restart the server |
| `internal` | unclassified — the redacted catch-all | inspect the peer-review server stderr log for the `correlationId` in `error.context`; rerun with `LOG_LEVEL=debug` |
### Walkthrough: 401 from every source
All peers were consulted and every one answered 401. This is **not** reported as a network outage — `cause` is `credential`, the roster is intact, and no credential material ever appears in a caller-visible field:
```jsonc
{
"success": false,
"error": {
"code": "EXTERNAL_SERVICE_ERROR",
"message": "No peer source produced a response",
"cause": "credential",
"detail": "No consulted peer source produced a usable response; see error.context.sources[]",
"fix": "Set the credential env vars or repair apiKeyCommand for the unavailable sources in peer-review.config.json, then restart the server",
"context": {
"tier": 1,
"requiredWeight": 4,
"agreeingWeight": 0,
"sources": [
{ "name": "gpt-large", "model": "gpt-5.2", "status": "unavailable", "weight": 2, "agreement": null,
"error": "Authentication failed for \"gpt-large\" (HTTP 401)", "cause": "credential", "code": "CONFIGURATION_ERROR" }
]
}
}
}
```
Set the env vars (or repair `apiKeyCommand`) and **restart the server** — config and credential availability are evaluated once at boot.
### Walkthrough: the caller abandoned a slow review
Symptom: the harness reports its own generic "timed out" and never reads a verdict, while the server logs the completed answer *after* the caller left. Cause: `PEER_REVIEW_DEADLINE_MS` is not strictly below `PEER_REVIEW_CALLER_TIMEOUT_MS` (the declared MCP client tool-call timeout — omp's per-server `timeout` key). Fix either side: lower `PEER_REVIEW_DEADLINE_MS`, or raise the tool-call timeout in the harness registration and keep the declared value equal to it. The server never dies over this pair: it warns once at startup, `list_peers` reports the mismatch (`callerCeiling`, above), and every failure envelope while it holds carries the same fix line.
Per-call server-side logging names the request id, tool, tier, and duration, so "answered after the caller left" is visible in the log rather than guessed.
## Migrating from gemini-for-claude-mcp
This server fully supersedes [`gemini-for-claude-mcp`](#migrating-from-gemini-for-claude-mcp): Gemini becomes just one configured source, reachable via AI Studio or Vertex AI.
It is intended as a drop-in replacement: `query_peer` accepts the same `prompt`/`history` shape (`role: "user"|"model"` preserved) and returns the same payload fields as `query_gemini` plus `source`; `list_peers` covers `list_gemini_models`; `count_tokens` covers `count_gemini_tokens` (with the `method` caveat below). Configure Gemini as an ordinary source, either way:
**AI Studio (static key):**
```json
{ "name": "gemini", "apiType": "openai",
"baseUrl": "https://generativelanguage.googleapis.com/v1beta/openai",
"model": "gemini-3-pro", "apiKeyEnv": "GEMINI_API_KEY", "weight": 2, "tier": 1 }
```
**Vertex AI (short-lived OAuth token via gcloud):**
```json
{ "name": "gemini", "apiType": "openai",
"baseUrl": "https://aiplatform.googleapis.com/v1beta1/projects/YOUR_PROJECT/locations/global/endpoints/openapi",
"model": "google/gemini-3.5-flash",
"apiKeyCommand": "gcloud auth print-access-token", "weight": 2, "tier": 1 }
```
Caveat: `count_tokens` for Gemini uses the `ceil(chars/4)` estimate (`"method": "estimate"`) rather than Gemini's native count endpoint.
## Development
```bash
npm run typecheck && npm run lint && npm test # full suite
npm run dev # tsx watch mode
```
See [ARCHITECTURE.md](./ARCHITECTURE.md) for layer rules, the quorum algorithm, and runtime guards.
## Contributing
Issues and pull requests are welcome. Before submitting a PR, run the full check suite:
```bash
npm run typecheck && npm run lint && npm test
```
## License
Apache-2.0 — see [LICENSE](./LICENSE).
This server cannot be deployed
Maintenance
ActivityActive
ResponsivenessNo recent community issues