Gonka Second Opinion — MCP Server

This MCP server (io.github.bystray/gonka-mcp-server) provides a “second opinion” for AI agents by running one prompt across several live Gonka models and assigned debate roles. It returns divergent viewpoints in a single MCP call, without switching providers. It exposes 20 tools and is identified by the slug io-github-bystray-gonka-mcp-server.

🛠️ Key Features

Single call: one prompt across multiple live Gonka models + roles
Includes a debate roles mechanism (e.g., skeptic / proponent / pragmatist)
OpenAI-compatible mention in topics
20 tools

🚀 Use Cases

Checking or validating a decision using independent model perspectives
Generating a panel of viewpoints without changing inference providers

⚡ Developer Benefits

Easier evaluation workflow: one MCP call for multiple perspectives
Works within MCP tooling; supports streaming (badge text present in excerpt)

⚠️ Limitations

The description does not specify role coverage beyond examples or any control over model selection beyond what Gonka provides

Gonka Second Opinion — MCP Server

Get a second opinion from independent LLMs — in one call, for a fraction of a cent.

gonka-second-opinion runs your prompt across the independent models live on the Gonka network and across assigned debate roles (skeptic / proponent / pragmatist …), in a single MCP call. Your main model stays in charge — you just get a panel of divergent views to check a decision against, without switching providers.

MCP

Live at: https://mcp.gogonka.com/mcp

One question → three independent opinions

Example: gonka_second_opinion runs one prompt across independent models and roles (proponent/skeptic/pragmatist) and returns attributed, divergent answers for your model to synthesize.

$One 3-model second opinion costs about 0.0014 cents — a fraction of a cent per panel.$

The example is illustrative — the live model set rotates over time; call get_available_models for today’s lineup. Per-panel cost is typical/measured; check get_pricing for live rates.

Why this instead of a plain LLM connector

A panel, not a single answer. Independent models trained by different labs surface non-overlapping failure modes. Assigned roles add behavioral diversity on top. You get disagreement you can act on, not one confident guess.
Cheap enough to always use. On the big providers a 3-way panel costs real money, so nobody runs one. On Gonka it's a fraction of a cent — a second opinion becomes a default habit, not a luxury.
Zero setup, works in a sandbox. The server makes the gateway call for you. An agent with no outbound network and no way to edit its own config can still use it over the already-open MCP channel. Free trial by default; paste your own key to run on your balance.

Install (30 seconds)

Add to your MCP client config (Claude Desktop / Cursor / …):

json

{
  "mcpServers": {
    "gonka-second-opinion": {
      "url": "https://mcp.gogonka.com/mcp"
    }
  }
}

Restart the client — the tools appear automatically. No API key required to start (a free trial key is issued per caller).

Custom clients:

bash

curl -X POST https://mcp.gogonka.com/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/list","id":1}'

Use your own key (after you register)

Registered on the JoinGonka gateway (gate.joingonka.ai)? Then run every call on your own balance — no trial budget, no per-IP limit. First get your key:

Sign up at the JoinGonka gateway (welcome bonus: 50M nGNK, no credit card).
Open your gateway dashboard → API keys → copy your JoinGonka gateway key (starts with jg-).

Then tell this MCP server to use it — pick the way that matches your client:

A. Client with an API-key / header field (Cursor, Claude Code, LibreChat, most custom clients)

Put the key in the server's Authorization header — it arrives as Bearer … and every call switches to your balance automatically:

json

{
  "mcpServers": {
    "gonka-second-opinion": {
      "url": "https://mcp.gogonka.com/mcp",
      "headers": { "Authorization": "Bearer jg-YOUR_KEY_HERE" }
    }
  }
}

If your client shows a plain API key / Bearer token box instead of raw headers, just paste your jg-… JoinGonka gateway key there.

B. Client with no key/header field (e.g. claude.ai Custom Connectors)

Mint a personal connector URL that carries your key for you:

Go to https://mcp.gogonka.com/personal
Paste your jg-… JoinGonka gateway key → you get a URL like https://mcp.gogonka.com/k/<token>/mcp
Add that URL as the MCP connector (in claude.ai: Add custom connector → URL). Nothing else to enter — no key field needed.

Keep the personal URL secret — it stands in for your key. The same key always mints the same URL.

How it's handled: the key (or personal token) is read from the request only, forwarded to the JoinGonka gateway, and never stored or logged. An unrelated Authorization header (OAuth, a proxy's own token) is ignored, so it can't break the free trial path. Remove the header/URL and the server falls back to the free trial.

The flagship tools

`gonka_second_opinion(prompt, system?, perspectives?, max_tokens?)`

Run one prompt across the live models and/or a set of roles, in parallel.

Leave perspectives empty → one opinion per live model.
Pass perspectives (short role/stance labels, max 5) → one opinion per role, each answered candidly from that viewpoint, rotated across the live models. Examples: ["for","against","neutral"], ["security expert","product manager","end user"].
Returns { opinions:[{model, perspective?, response}], synthesis_instructions, trial_budget|cost }. Attribute each opinion to its model + role when you present them.

`gonka_chat(prompt, system?, model?, max_tokens?)`

One completion through the server — for when you want a single cheap answer, or a sub-task run without giving the agent its own network/key. model accepts "auto", a nickname, or an exact id; anything not live right now is routed to a model that is.

Both tools auto-detect trial (free, per-IP budget) vs registered (your own jg-… key) mode. On trial exhaustion they return a signup link + welcome bonus to relay to the user — no fabricated keys, no altered numbers.

Also included

Cost intelligence — is Gonka actually cheaper for your workload? (8 tools)

Gonka live inference price per 1M tokens and how many times cheaper than GPT-5.5 — see get_pricing for the real-time number.

Reflects the last cached render (GitHub proxies images); call get_pricing for the real-time number. All figures float with the GNK/USD rate — never trust a hardcoded one.

get_pricing — current $/1M tokens, GNK/USD rate, live ratios vs OpenAI/DeepSeek/Anthropic
compare_providers(provider) — side-by-side vs openai / anthropic / deepseek / mistral / gemini
calculate_savings(monthly_spend_usd) — exact monthly/annual savings for your spend
suggest_model_for_task(task_description, …) — model + cost recommendation
get_available_models — the models live on the network right now, with live pricing
get_trial_key — free gc- key instantly (no signup); expires in 2h, first use extends to 24h
get_signup_link — permanent account + welcome bonus + copy-paste SDK snippets
register_on_gonka(…) — personalized cost-analysis pitch (computes savings; does not create an account)

Documentation knowledge graph — search 1000+ Gonka concepts (10 tools)

query_graph(question, …) — topic search over the concept graph
search_docs(query, …) / read_doc(filename, …) / list_docs() — full-text docs
get_node(label) / get_neighbors(label, …) — concept detail and relations
get_god_nodes() / get_community(id) / get_graph_stats() — structure overview
find_shortest_path(source, target, …) — conceptual path between two ideas

MCP prompts & resources

Prompt gonka_start(task) — seed a cost-comparison conversation
Prompt calculate_my_savings(monthly_spend_usd) — seed a savings estimate
Resource gonka://pricing-guide — structured live pricing data + decision matrix

Honesty & hygiene

Live, not hardcoded. Prices float with the market and the model set rotates — the server quotes both live (get_pricing, get_available_models), and the badges above are rendered from live data, not typed into this file.
Panel ≠ magic. A second opinion is worth exactly as much as the models are independent. The panel uses the genuinely different models live on the network at the time of the call; it does not claim a bigger fleet than is actually serving.
OpenAI/Anthropic-compatible. Two config changes, zero code changes: base_url=https://gate.joingonka.ai/v1 (OpenAI) or https://gate.joingonka.ai (Anthropic).
Tool hygiene. All tools ship typed input schemas with MCP annotations (readOnlyHint, idempotentHint, openWorldHint), enum/range constraints, and clear descriptions. Inference tools are per-IP rate-limited in trial mode.
Privacy. HTTPS/TLS 1.2+. No keys in query strings (flagged by middleware). Anonymized logging (IP, User-Agent, tool name); inference prompts truncated; caller keys never logged or returned.

Live health grade: https://wmcp.sh/mcp/grade/mcp.gogonka.com

task_description	string	required	What task the model should perform (e.g. 'chatbot', 'code generation', 'summarization').
monthly_budget_usd	number	optional	Current monthly API spend in USD (0 = unknown). Optional.
current_provider	string	optional	Current LLM provider for cost comparison.

prompt	string	required	The question to put to every opinion (required).
system	string	optional	Optional base system instruction applied to all.
perspectives	any	optional	Optional list of short role/stance labels (max 5). Each becomes one independent opinion.
max_tokens	integer	optional	Max completion tokens per opinion (kept low — this fans out).
share	boolean	optional	Default False. Set True ONLY when the user explicitly wants a shareable public link. It creates a PUBLIC (unlisted) web page at gogonka.com/o/<id> showing the question and answers, and returns `share_url`. Warn the user the page is public before sharing. Publication is refused if the text looks like it holds an API key/secret.
synthesis	string	optional	Optional. When sharing (share=True), pass YOUR short comparison of the opinions (where they agree / differ, your takeaway) — it is shown on the page as the asking agent's take. Leave empty if you have none; the page still shows a text-similarity signal.

monthly_spend_usd	number	optional	User's current monthly LLM spend in USD.
current_provider	string	optional	Current provider (openai, anthropic, deepseek).
user_query	string	optional	What the user said (for context, echoed back — not sent anywhere).

question	string	required	Natural-language question or topic, e.g. "how do I deposit GNK".
depth	integer	optional	How many hops to traverse from the matched concept in the knowledge graph. Higher = more context, more tokens.
token_budget	integer	optional	Approximate max size of the returned text.

filename	string	required	Exact or partial .md filename as returned by query_graph(), search_docs(), or list_docs() (e.g. "hardware-specifications.md" or "hardware-specifications").
max_chars	integer	optional	Maximum characters to return; longer files are truncated.

label	string	required	Concept name (or a close substring of it), e.g. "collateral".
relation_filter	string	optional	Only return edges whose relation label contains this substring (case-insensitive). Empty = no filter.

query	string	required	One or more keywords, e.g. "min_amount escrow". Prefer fewer, more specific words over a full sentence.
max_results	integer	optional	Maximum number of files to return.
context_chars	integer	optional	Size of the excerpt shown around the match, in characters.

source	string	required	Starting concept name, e.g. "trial key".
target	string	required	Destination concept name, e.g. "gateway".
max_hops	integer	optional	Give up if the path is longer than this many edges.