技能详情（站内镜像，无评论）
许可证：MIT-0
MIT-0 ·免费使用、修改和重新分发。无需归因。
版本：v0.1.0
统计：⭐ 0 · 84 · 0 current installs · 0 all-time installs
⭐ 0
安装量（当前） 0
🛡 VirusTotal ：良性 · OpenClaw ：良性
Package：alanarchy/clawswarm-consensus
安全扫描（ClawHub）
- VirusTotal ：良性
- OpenClaw ：良性
OpenClaw 评估
ClawSwarm's code and runtime instructions match its stated purpose (orchestrating many LLM agents and running a consensus engine), but you should supply and protect an LLM API key and be careful about any custom API endpoints you configure.
目的
Name/description describe a multi-agent prediction and consensus tool; included scripts (swarm_runner.py and consensus.py), README, and config reference implement exactly that behavior. Required binaries and permissions in the registry are minimal and consistent with running Python scripts and making HTTP calls to LLM providers.
说明范围
SKILL.md and examples instruct running swarm_runner.py and consensus.py and to set an API key environment variable (e.g., GROQ_API_KEY). Runtime actions are limited to reading a provided config file, making outbound HTTP requests to LLM provider endpoints, and piping results through the local consensus engine. It does not attempt to read arbitrary host files or credentials. Note: the registry metadata lists no required env vars, but the exampl…
安装机制
There is no install spec — this is instruction + packaged Python scripts. No installers, downloads, or extracted archives are used. Dependencies are normal Python libs (numpy, requests, pyyaml) and are declared in README; risk from install mechanism is low.
证书
The code expects an LLM API key (examples use GROQ_API_KEY) and may accept an api_key in the config; however the skill metadata declares no required env vars or a primary credential. Requesting an LLM API key is proportionate to the stated purpose, but the mismatch between metadata and the code (and the ability to set arbitrary base_url) should be noted.
持久
Skill is not always-enabled and does not request elevated or persistent system privileges. It runs as a normal user process, spawns a local subprocess for consensus.py, and does not modify other skills or system-wide configs.
综合结论
This skill appears to do what it says: run many LLM agents and compute a consensus. Before installing or running it: - Provide an LLM API key (e.g., GROQ_API_KEY, OPENAI_API_KEY) via environment variable or the config; the registry metadata currently does not list required env vars, so manually ensure you set the key. - Do not point 'base_url' to untrusted servers — the tool sends full prompts (including any context you include) to the configu…
安装（复制给龙虾 AI）
将下方整段复制到龙虾中文库对话中，由龙虾按 SKILL.md 完成安装。
请把本段交给龙虾中文库（龙虾 AI）执行：为本机安装 OpenClaw 技能「ClawSwarm」。简介：Multi-agent swarm prediction with consensus engine. Use when: running multiple …。
请 fetch 以下地址读取 SKILL.md 并按文档完成安装：https://raw.githubusercontent.com/openclaw/skills/refs/heads/main/skills/alanarchy/clawswarm-consensus/SKILL.md
（来源：yingzhi8.cn 技能库）SKILL.md
---
name: clawswarm
description: "Multi-agent swarm prediction with consensus engine. Use when: running multiple AI agents to predict prices, values, or outcomes and aggregating their predictions into a consensus. Supports any LLM provider (Groq, OpenAI, Ollama), configurable agent roles/temperatures, and a statistical consensus pipeline (MAD outlier filtering, adaptive anchoring, bias correction, weighted median aggregation). Triggers: swarm prediction, multi-agent consensus, collective intelligence forecasting, price prediction with multiple agents."
---
# ClawSwarm
Multi-agent collective intelligence framework. Run N agents with different analytical perspectives, aggregate predictions through a statistical consensus engine.
## Quick Start
### 1. Create a config file
```yaml
target:
name: "Gold"
current_price: 5023.1
unit: "USD/troy oz"
context: "RSI: 40.8 | MA5: 5084 | MA10: 5120"
agents:
- role: "Macro analyst focusing on geopolitical risk"
count: 50
temperature_range: [0.4, 0.7]
- role: "Technical RSI/MACD momentum trader"
count: 30
temperature_range: [0.45, 0.6]
- role: "Mean reversion auditor"
count: 20
temperature_range: [0.35, 0.55]
api:
provider: groq
model: llama-3.3-70b-versatile
api_key_env: GROQ_API_KEY
delay_ms: 1200
consensus:
max_deviation: 0.15
```
### 2. Run the swarm
```bash
python3 scripts/swarm_runner.py --config swarm.yaml
```
Output: JSON with `final_price`, `median_price`, `confidence`, `bull_ratio`, and all individual predictions.
### 3. Run consensus standalone
Pipe any predictions array to the consensus engine:
```bash
echo '{"predictions":[{"price":100.5,"confidence":70},{"price":99.8,"confidence":60}],"anchor_price":100.0}'
| python3 scripts/consensus.py
```
## Architecture
```
Config (YAML/JSON)
↓
Swarm Runner (swarm_runner.py)
├─ Agent 1 → LLM API → prediction
├─ Agent 2 → LLM API → prediction
├─ ...
└─ Agent N → LLM API → prediction
↓
Consensus Engine (consensus.py)
├─ Bias correction
├─ MAD outlier filtering
├─ Anchor-distance filtering
├─ Multi-method aggregation (weighted 40% + median 35% + trimmed mean 25%)
├─ Adaptive anchoring (dispersion → anchor strength)
└─ Clamping
↓
Final consensus prediction + confidence + bull/bear ratio
```
## Key Concepts
**Agent diversity**: Each agent gets a different role prompt and temperature. More diversity = better consensus.
**Consensus engine**: Not a simple average. Uses MAD (Median Absolute Deviation) to filter outliers, adaptive anchoring to stabilize results when predictions are dispersed, and multi-method aggregation for robustness.
**1 agent or 1000**: Works with any count. Single agent bypasses consensus. 5+ agents get full pipeline.
## Config Reference
See `references/config-reference.md` for full field documentation and example configs.
## Scripts
| Script | Purpose |
|--------|---------|
| `scripts/swarm_runner.py` | Orchestrate multi-agent predictions |
| `scripts/consensus.py` | Standalone consensus engine (pipe JSON in) |
## Dependencies
- Python 3.8+
- `numpy` (for consensus engine)
- `requests` or `urllib` (for API calls)
- `pyyaml` (optional, for YAML configs; JSON always works)