Wisepanel

io.github.ikoskela由 GitHub 账号 ikoskela 发布v1.0.1更新于 Nov 10, 2026

Multi-agent deliberation across Claude, Gemini, and Perplexity. Stream and publish.

概览

AI 生成的概览

运行多智能体审议:由前沿模型组成的小组从对立角色辩论一个问题,助手可实时查看并获取结果。

可以做什么
  • 启动、轮询、取消和列出审议,可设置小组规模、模型组、轮次与上下文
  • 在问题进入小组前重写问题,去除有偏见的措辞
  • 获取已完成审议的完整结果,包括存留下来的立场与未解决的分歧
  • 将完成的审议发布到公开的 Wisepanel Commons
适用场景
适用于出错代价高昂的高风险决策,例如架构选型、迁移、安全与隐私权衡、供应商选择,并希望看到反对自己直觉的最强论据。不适合有已知答案的问题;它比单次查询更慢、更贵。
运行要求
需要本地 Node.js 运行时,通过 npx 运行该 npm 包(仅 stdio,无 HTTP 端点)。需要 Wisepanel 账户,并以环境变量 WISEPANEL_API_KEY 提供 API 密钥(不是请求头)。可选 WISEPANEL_API_URL 覆盖 API 地址。需要访问 Wisepanel API 的网络连接。
安装前请注意
API 密钥属于机密:在 WISEPANEL_API_KEY 中直接填写,不要粘贴到聊天或截图中,泄露后应吊销。审议会产生费用;推理审计约 1.45 倍,联网搜索约 3.25 倍,最大规模小组可达最小规模的 5 倍。发布会使审议内容公开可见。context_file 路径会从本机读取文件。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Wisepanel,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

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_prompt rewrites 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:

bash
claude mcp add wisepanel --scope user \  --env WISEPANEL_API_KEY=wp_sk_ExampleOnly0000-replace-with-your-own-key \  -- npx -y wisepanel-mcp

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 http will 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.:

json
{  "mcpServers": {    "Wisepanel": {      "command": "npx",      "args": ["-y", "wisepanel-mcp"],      "env": {        "WISEPANEL_API_KEY": "your-api-key"      }    }  }}

Configuration

VariableRequiredDefault
WISEPANEL_API_KEYyes—
WISEPANEL_API_URLnohttps://api.wisepanel.ai

Troubleshooting

'url' is not a valid URL — the server was added with --transport http. Remove it and re-add using the stdio command above:

bash
claude mcp remove wisepanel --scope user

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.

ParameterTypeDescription
questionstring (required)The topic for the panel to deliberate
topologystringPanel size — see Topology. small (6 agents, default), medium (12), large (30)
model_groupstringSee Model groups. Default smart
roundsnumberPolyhedron traversals (1-5). Default 1 — see Rounds
contextstringAdditional framing context
context_filestringPath to a file used as context, for payloads too large to pass inline. Concatenated after context if both are given
compressionstringContext compression: none, moderate, aggressive (default)
short_responsesbooleanRequest concise panelist responses. Default false
show_and_audit_reasoningbooleanReasoning-quality scaffolding + cross-agent audit. Server default is on — omit to accept it, pass false to opt out. ~1.45x cost
web_search_enabledbooleanLet agents verify factual claims via native provider web search. Requires smart. Default false. ~3.25x cost, ~6.5x combined with audit

Topology

Agents sit on the polyhedron's edges, so the agent count is the edge count:

topologyPolyhedronVerticesAgentsResponses per round
smalltetrahedron46~12
mediumoctahedron612~24
largeicosahedron1230~60

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.

StructureUniform influenceFast propagationCost
Hub-and-spoke✗ one position frames everything✓linear
Chain / round-robin✗ anchoring, last-word advantage✗linear
All-to-all✓✓O(n²)
Polyhedral edges✓✓linear

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:

GroupRelative costUse when
smart1× (baseline)default; current flagships (Opus 5, GPT-5.6 Sol, Gemini 3.1 Pro Preview)
cheap / fast~¼×small models; fast optimises latency, cheap optimises cost — same tier
mixed< 1×random across all providers; cheaper on average, quality varies seat to seat
informed~1×search-capable models incl. Perplexity Sonar; the answer turns on current facts
largevarieslargest context windows — for big context payloads, not better answers
anthropic-fable~2×Claude Fable 5 on every seat; only when maximum capability is explicitly wanted

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.

ParameterTypeDescription
questionstring (required)The question to rewrite, as the user wrote it

Returns one of three outcomes. The original question is echoed back in every case, so you can always fall back to it:

outcomeMeaningBilled
transformedRewritten. Response includes rewrittenyes
no_change_neededAlready unbiased — use the originalno
fail_closedNo safe rewrite produced — use the originalno

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

1. wisepanel_start    -> returns run_id2. wisepanel_poll     -> (repeat) returns panelist responses as they arrive3. On completion, poll includes publish_available: true4. wisepanel_publish  -> publishes to Commons, returns public URL

Environment Variables

VariableRequiredDescription
WISEPANEL_API_KEYYesYour Wisepanel API key
WISEPANEL_API_URLNoAPI base URL (defaults to https://api.wisepanel.ai)

Development

bash
git clone https://github.com/ikoskela/wisepanel-mcp.gitcd wisepanel-mcpnpm installnpm run dev

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.

来源:README.md,提交 bfd4910

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v1.0.1最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分