
Wisepanel
io.github.ikoskela由 GitHub 帳號 ikoskela 發布v1.0.1更新於 Nov 10, 2026
Multi-agent deliberation across Claude, Gemini, and Perplexity. Stream and publish.
概覽
執行多代理審議:由前沿模型組成的小組從對立角色辯論一個問題,助理可即時查看並取得結果。
- 可以做什麼
- 啟動、輪詢、取消與列出審議,可設定小組規模、模型群組、輪次與上下文
- 在問題送進小組前改寫問題,移除帶有偏見的措辭
- 取得已完成審議的完整結果,包括存留的立場與未解決的分歧
- 將完成的審議發佈到公開的 Wisepanel Commons
- 適用情境
- 適用於出錯代價高昂的高風險決策,例如架構選擇、遷移、安全與隱私取捨、供應商評選,並希望看到反對自己直覺的最強論據。不適合已有已知答案的問題;它比單次查詢更慢、更貴。
- 執行需求
- 需要本機 Node.js 執行環境,透過 npx 執行該 npm 套件(僅 stdio,沒有 HTTP 端點)。需要 Wisepanel 帳號,並以環境變數 WISEPANEL_API_KEY 提供 API 金鑰(不是標頭)。可選用 WISEPANEL_API_URL 覆寫 API 位址。需要連線至 Wisepanel API 的網路存取。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 Wisepanel,將其新增到工作區。
- 為需要使用其工具的對話啟用該服務。
Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。
其他 MCP 客戶端
參照 儲存庫 中的啟動說明。
README
Wisepanel MCP Server
The decision-intelligence layer between frontier models and high-consequence decisions.
Wisepanel takes a question, builds a panel of AI agents around it, and has them argue it out. You get back the positions that survived the argument, the reasoning behind each one, and the disagreements that never resolved.
This MCP server exposes that to Claude Code and any MCP-compatible client.
Why
A single model gives you one answer, fluently, whether or not it is right. That is fine for most questions. It is a bad property for the ones where being wrong is expensive.
The failure usually isn't ignorance. A model commits to a framing early and then argues for it, so you never see the objection, the alternative, or the assumption doing the work. Ask again and you get the same framing in different words. Ask three models separately and you get three confident answers with no way to choose between them.
How it works
Wisepanel builds disagreement in deliberately, at three levels.
One — each agent is handed a conflict to resolve. Roles are derived from your question, and each is defined by two forces that genuinely oppose each other: cost against access, speed against safety, proven against new. The agent can't champion one side. It has to reach a position that answers both, so it arrives with something worked through rather than a talking point.
Two — every agent resolved a different conflict, so their positions don't match. The agent holding cost against access lands somewhere the one holding speed against safety does not. These aren't two sides of an argument. They are several honest resolutions of the same question that disagree about what mattered most, and between them they cover the ground the question actually occupies.
Three — the structure makes them contend. Agents are placed on the edges of a polyhedron, so each one works at two vertices — two separate conversations. At its second vertex an agent is not merely a participant but a delegate from the first, instructed to represent what its co-participants concluded there alongside its own position. Every vertex therefore hears whole conversations it was not part of, argued by someone who was.
That last part is what makes a small panel go further than its headcount. Six agents means twelve seats, and each seat imports another discussion — so a point raised anywhere reaches the entire panel within a few hops, with no aggregator, no summarizer and no bottleneck. Speaking order is balanced, so no agent frames the discussion first or takes the last word. Details.
Isn't this just asking three models?
Pasting the same question into Claude, GPT and Gemini is a real technique, and it works for a real reason: different labs train on different data with different methods, so their priors genuinely differ. Wisepanel does the same thing — roles are spread across Anthropic, OpenAI, Google and Perplexity by default, so no single lab's blind spots go unchallenged.
But the model is only one of the places bias enters. There are four, and doing it by hand reaches one.
Your framing goes to everyone unchanged. You paste the same words three times, so you
sample three training substrates against a single reading of the question. When the question
carries an assumption — and questions about decisions usually do — you get three confident
answers to the wrong question.
wisepanel_magic_prompt rewrites the framing before the panel sees
it.
Each model answers as itself. You get Claude's median take, GPT's median take, Gemini's median take, and medians cluster. A model asked a neutral question gives a balanced answer; it will not volunteer the strongest case against your plan, because that isn't what it was asked for. Assigning a role changes what the model is optimising for, which produces arguments none of them offer unprompted.
Bouncing answers between models makes anchoring worse, not better. Feed A's response to B and B now reasons inside A's framing — models tend to accept a stated position and refine it rather than discard it and start over. So the sequential version is more biased than three independent queries, and whichever model you happened to open first sets the terms. Wisepanel balances speaking order and spreads the conversation across vertices precisely so no single position gets to be the one everyone reacts to.
The reconciliation lands on you. Three answers arrive; nothing has compared them. You do that work yourself, with your own priors, usually at the end of a long day on a decision you already lean one way about. A panel does the contending first and hands you what survived it.
Then there is the part that doesn't scale by hand. Three models is three samples. A panel is 6 to 30 roles chosen to span the question, each holding an opposition, each carrying a second conversation to its other vertex — twelve seats at the smallest size. You are not going to hand-run that, and you are certainly not going to do it consistently on every decision that deserves it.
Where doing it by hand wins: it's free, it's immediate, and you keep complete control of the wording. For most questions that is the right trade. This is for the ones where it isn't.
Checks around the argument
- The question is checked for bias first.
wisepanel_magic_promptrewrites loaded framing, embedded assumptions and false binaries before the panel sees them. A biased question produces a confident answer to the wrong thing. - Reasoning is auditable. Agents attribute claims, surface assumptions and flag each
other's gaps — on by default. See
show_and_audit_reasoning. - Claims can be checked against sources. Optional native web search verifies dates,
citations, figures and rules instead of trusting recall. See
web_search_enabled.
When to use it
When being wrong is expensive — architecture calls you'll live with for years, migrations, security and privacy trade-offs, vendor selection, anything where you want the strongest case against your instinct before you commit. It is slower and costs more than a single query. That is the trade you are making.
Don't use it for questions with a known answer, or where you would not act differently given a dissenting view.
Runs stream live, so you watch the argument develop rather than waiting for a verdict. Completed deliberations can also be published to the Wisepanel Commons.
Quick Start
Get your API key at wisepanel.ai/settings, then:
Paste the key exactly as shown on the settings page — the whole wp_sk_… string and
nothing else. Quotes around it are optional and harmless. Do not add a Bearer prefix:
the server sends Authorization: Bearer <your-key> itself, so including it yields
Bearer Bearer wp_sk_… and auth fails.
Restart Claude Code and run /mcp — wisepanel should show as connected.
✔ Connected only means the server process launched. Your API key isn't checked
until the first call, so a bad key still shows as connected. To confirm auth
actually works, run a deliberation and check that it returns a run_id.
This is a stdio server, not a remote one. There is no HTTP endpoint —
claude mcp add --transport httpwill not work no matter what URL you give it. Everything after the--is the command that launches the server locally.
Other MCP clients (manual config)
Add to your client's config file — ~/.claude.json for Claude Code, or the
equivalent for Cursor, Windsurf, Claude Desktop, etc.:
Configuration
Troubleshooting
'url' is not a valid URL — the server was added with --transport http.
Remove it and re-add using the stdio command above:
WISEPANEL_API_KEY environment variable is required — the key didn't reach
the server process. Pass it with --env as shown, not as an Authorization
header; headers apply to remote servers only.
API 401 / not authenticated despite a valid key — check the stored value with
claude mcp get wisepanel. It must be the bare wp_sk_… string. A Bearer prefix, a
trailing space, or a partial paste are the usual causes.
Not authenticated — verify the key is active at wisepanel.ai/settings. Keys are secrets: never paste them into chat, issues, or screenshots. If one leaks, revoke and reissue it.
Tools
wisepanel_start
Start a deliberation. Convenes a panel of AI models to debate a question from assigned perspectives. Returns run_id immediately.
Topology
Agents sit on the polyhedron's edges, so the agent count is the edge count:
Time and cost scale with agent count — large is 5× small. Escalate when a question needs
more genuinely distinct perspectives, not when you want a better answer from the same ones.
Why edges rather than vertices. Every edge of a Platonic solid is equivalent under the solid's symmetry group, and speaking order is balanced so no agent consistently anchors or consistently gets the last word. There is no hub and no privileged seat. Graph diameter stays small — 1, 2 and 3 respectively — so an insight raised anywhere reaches the whole panel in a few hops. Because each agent sits on an edge, it is simultaneously a participant and a bridge: propagation is a side effect of participation, with no messenger or aggregator role.
All-to-all buys the same reach and uniformity at quadratic cost. Edge assignment on a regular polyhedron is the structure that gets both at linear cost.
Model groups
Cost is relative to smart, the default:
Single-provider groups (openai, anthropic, google, perplexity) pin every seat to one
vendor, which removes cross-vendor diversity — usually the point of a panel.
Rounds
Agents sit on the edges of the polyhedron, not the vertices. Each agent connects two
vertices (conversation nodes) and contributes at both endpoints every round — so
rounds: 1 already produces roughly num_agents × 2 responses.
Rounds are full polyhedron traversals, not chat turns. rounds: 1 is already substantial
deliberation. Use 2+ only when agents need to react to other agents' completed positions —
e.g. a binary strategic decision with sharply opposing arguments.
wisepanel_magic_prompt
Rewrite a question to remove framing that would bias the panel toward a predetermined
answer — loaded wording, embedded assumptions, false binaries — while preserving intent.
Optional pre-step before wisepanel_start.
Returns one of three outcomes. The original question is echoed back in every case, so you can always fall back to it:
Show the user both versions and let them choose. The rewrite can shift emphasis in ways they may not want, so it should never be substituted silently. This mirrors the web app, where the transform runs only on an explicit click, behind a cost confirmation, with revert available.
Billed separately from the deliberation, and only when the text actually changes.
wisepanel_poll
Long-polls a running deliberation (waits up to 15s for new events). Returns panelist responses as they arrive.
wisepanel_result
Retrieve full results of a completed deliberation. Only needed if you didn't poll it live.
wisepanel_cancel
Cancel a running deliberation.
wisepanel_publish
Publish a completed deliberation to the Wisepanel Commons. Makes it publicly viewable and shareable.
wisepanel_list_runs
List all deliberation runs in the current session.
Typical Flow
Environment Variables
Development
Patent pending
Wisepanel's multi-agent deliberation architecture — including the polyhedral topology and the reasoning-audit and verification subsystems — is the subject of pending US patent applications assigned to QuROI, Inc.
License
MIT — see LICENSE.
The MIT license covers the client in this repository only. It grants no license, express or implied, to any patent, or to the Wisepanel platform and the methods it implements.
來源:README.md,提交 bfd4910
工具
0版本歷史
1- v1.0.1最新Sep 16, 2026
評分
已滿足適用檢查項的 50%
該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。
- 近 30 天內可連通25 分
請保持端點能夠回應 MCP 握手;SourceWeft 每 30 天探測一次。
- 已宣告認證方式10 分
請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。
- 已列出工具15 分
請讓伺服器在無憑證時回應 tools/list,或發布可在沙箱中無憑證啟動的軟體包。
- 描述必需15 分
- 圖示10 分
- 文件15 分
- 授權條款10 分