Token导航 LogoToken导航

Agent Skill · 研究检索

oasis(oasis 搜索)

oasis 用于查找、检索和筛选相关信息,适合在 OpenClaw 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

oasis 是研究检索类 Agent Skill,可用宿主为 OpenClaw,通过ClawHub获取,权限为操作浏览器;许可证为 MIT-0,第三方扫描已通过,按安装量在研究检索类的 73,675 个 Skill 里排第 33,663。

总安装 2,326周安装 95GitHub Stars 公开资料未说明下载 745核验 2026-05-01

安装说明

本站只整理中文说明,不托管安装包,也不代用户安装。
GitHub
许可证
MIT-0
版本
0.4.0

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:oasis(oasis 搜索)
来源仓库:https://github.com/parzival-moksha/oasis
安装命令:
openclaw skills install oasis
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 OpenClaw 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

ClawHubOpenClaw
openclaw skills install oasis

这个 Skill 是做什么的

研究检索类 Skill 主要解决「找资料、核对事实、把来源整理清楚」的问题,通常会让助手先检索再总结。

oasis 使代理能够在 Oasis 3D 世界中工作,支持放置目录对象、使用着色器和纹理制作程序场景、绘制地砖、移动头像等操作。适用于 OpenClaw 平台中的 3D 世界构建与交互模拟场景。通过 clawhub 安装,使用 openclaw skills install oasis 命令部署。

建议确认权限范围和维护状态,避免不必要的网络或服务调用。可结合来源仓库和原始 README 继续核验具体用法和操作细节。

上手步骤

4 步
  1. 1

    确认自己的宿主工具

    公开资料标注支持 OpenClaw;先确认自己用的是其中之一,再按对应工具的 Skills 目录放置。

  2. 2

    拿到安装命令

    可以直接复制的命令是 openclaw skills install oasis;命令里的占位符换成自己的路径或密钥再执行。

  3. 3

    安装后重启宿主

    命令执行完要完全退出宿主工具再打开,让新装的 Skill 被重新扫描到,否则列表里可能看不到。

  4. 4

    按最小权限试用

    它需要操作浏览器,先在测试目录或测试账号里跑,确认没问题再放到正式项目。

常见问题

6 条

oasis 怎么安装?

在终端执行 openclaw skills install oasis。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

oasis 支持哪些 AI 工具?

公开资料标注支持 OpenClaw。其他工具只要能读取本地 Skill 目录,一般也能手动放进去,但要自己确认触发方式是否一致。

oasis 需要什么权限?

权限级别是操作浏览器。该 Skill 可能涉及浏览器控制能力,使用时可能读取或操作网页内容,需要在受控环境中确认权限边界。

oasis 是免费的吗?

来源标注的许可证是 MIT-0。本站不代收任何费用,也不托管安装包;能否商用、要不要署名,以许可证原文为准。

oasis 用的人多吗?

公开数据是累计安装 2,326 次,近一周安装 95 次,下载 745 次。这些数字来自第三方目录统计,只能当作热度参考,不代表质量或安全性。

oasis 安全吗?

第三方扫描结果为通过,但这只覆盖已知的静态检查项,不能替代自己读一遍源码和安装命令。

值得选的地方和要先想清楚的地方

值得选的地方

  • 专注 OpenClaw 一种宿主,触发方式和使用习惯不用来回切换。
  • 许可证明确为 MIT-0,商用和分发边界好判断。
  • 第三方扫描已通过,常见的高风险行为没有命中。

先想清楚的地方

  • 第三方目录整理,版本更新和行为细节以来源仓库为准。

适不适合你

更适合

  • 已经在用 OpenClaw,想直接扩展它能力的人。
  • 需要处理研究检索类任务、又不想每次都手写提示词的人。
  • 需要许可证清楚、并且有第三方扫描结果的团队项目。

不太适合

  • 需要厂商级技术支持和 SLA 的生产系统。

研究检索类 Skill 对比

6 个同类
Skill总安装GitHub 星权限许可证扫描
find-skills

寻找技能 · Codex、Claude

31,199,81316,486操作浏览器MIT有提醒
self-improving-agent

自我改进 Agent · OpenClaw

16,464,0023,396接触敏感数据MIT-0通过
soultrace

灵魂追踪 · Codex、Claude

7,588,94412需要联网未明确有提醒
self-improving

自我完善 · OpenClaw

6,829,2841,046需要联网MIT-0通过
microsoft-foundry

微软代工厂 · Codex、Claude

6,614,144764调用外部服务未明确有提醒
oasis

oasis 搜索 · 本页

2,326—操作浏览器MIT-0通过

按安装量取同分类前几名;数字来自第三方目录统计,只作热度参考,不代表质量结论。

SKILL.md 原文

name
oasis
description
Enables an agent to work inside Oasis 3D worlds — place catalog objects, craft procedural scenes with shaders and textures, paint ground tiles, move avatars, take screenshots, and drive Forge text-to-3D conjuration through the Oasis MCP tool surface.
version
0.4.0
author
Levi
license
MIT
metadata
hermes
tags
[oasis, mcp, world-building, 3d, integrations, creative]

Oasis

Oasis is a local-first 3D world that agents can inspect, modify, navigate, and build through a shared MCP tool surface. Humans run Oasis on their laptop; agents use these tools to co-create 3D scenes with them in real time.

This skill teaches you HOW TO USE those tools effectively once the user has already connected you to their Oasis. It does not cover installation — see the user's setup docs at https://parzival-moksha.github.io/oasis/docs/getting-started/quickstart/ if they need help connecting.

When To Use

Use this skill when:

  • the user wants you to inspect, place, remove, paint, light, move, or screenshot inside Oasis
  • the user wants to craft a procedural scene (campfire, garden, cyberpunk alley, shrine, etc.)
  • the user wants a 3D asset conjured via text-to-3D
  • the user wants you to move your avatar or take visual verification shots

Do not use this skill for generic coding unless Oasis integration is part of the task.

What You Can Do

The Oasis MCP tool surface supports:

  • world state and world summary (get_world_state, get_world_info)
  • object search and asset search (search_assets, query_objects)
  • placing, modifying, and removing catalog objects (place_object, modify_object, remove_object)
  • crafting procedural scenes (craft_scene, get_craft_guide, get_craft_job)
  • sky, ground presets, tile paint, and lights (set_sky, set_ground_preset, paint_ground_tiles, add_light, modify_light)
  • embodied agent avatars (set_avatar, walk_avatar_to, play_avatar_animation, list_avatar_animations)
  • viewport and avatar screenshots (screenshot_viewport, screenshot_avatar, avatarpic_merlin, avatarpic_user)
  • Forge conjuration workflows (conjure_asset, process_conjured_asset, place_conjured_asset, list_conjured_assets, get_conjured_asset, delete_conjured_asset)
  • world management (list_worlds, create_world, load_world, clear_world)
  • behavior hints (set_behavior)

The agent should usually:

  1. call get_world_state or get_world_info to understand the current scene
  2. call query_objects or search_assets as needed
  3. make a world mutation
  4. use screenshot tools when visual verification matters

Self-Craft Is The Default

When the user asks you to build something procedural (a campfire, a shrine, a crystal cluster, a fountain), you write the primitives yourself and pass them as the objects array to craft_scene. Do not delegate to the sculptor unless explicitly asked.

craft_scene({
  name: "Arcane campfire",
  position: [0, 0, 0],
  objects: [
    { type: "cylinder", position: [0, 0.08, 0], scale: [0.55, 0.08, 0.55], color: "#3b2a1d", roughness: 0.92 },
    { type: "flame", position: [0, 0.3, 0], scale: [0.22, 0.35, 0.22], color: "#fff4dd", color2: "#ff7a00", color3: "#9b1d00" },
    { type: "particle_emitter", position: [0, 0.75, 0], scale: [0.45, 0.85, 0.45], color: "#ffb347", particleCount: 80, particleType: "ember" },
    { type: "crystal", position: [0.65, 0.32, 0.1], scale: [0.22, 0.6, 0.22], rotation: [0.14, 0.3, -0.08], color: "#4338ca", color2: "#8b5cf6", seed: 11 }
  ]
})

What you have access to (call get_craft_guide for the live spec):

  • Geometry: box, sphere, cylinder, cone, torus, plane, capsule, text
  • Shaders: flame, flag, crystal, water, particle_emitter, glow_orb, aurora
  • Animations: rotate, bob, pulse, swing, orbit (with type, speed, axis, amplitude)
  • Textures: 20 presets including stone, cobblestone, marble, concrete, grass, sand, snow, metal, wood, kn-planks, kn-cobblestone, kn-roof, kn-wall. Apply via texturePresetId + textureRepeat.
  • Material fields: metalness, roughness, opacity, emissive, emissiveIntensity, color2, color3

Rules the craft runtime enforces:

  • No ground planes, sky domes, or background walls — Oasis already provides the world.
  • Use shader primitives aggressively for fire, cloth, crystal, water, glow, aurora.
  • Many small overlapping primitives beat one oversized primitive.
  • Non-zero rotation on at least some primitives.

When to use sculptor fallback

craft_scene({ prompt: "...", strategy: "sculptor" }) delegates scene generation to a separate coder subprocess on the Oasis host. It costs a real LLM call, takes several seconds, and streams primitives in as they arrive. Use it only if:

  • The user explicitly asks you to delegate ("have the sculptor do it")
  • The scene is so ambitious you'd rather have a dedicated coder agent sketch it first

Otherwise: self-craft. You are an LLM. You can write the JSON.

Progressive Smoke Test

These are the five prompts a user typically sends to verify the connection. Each escalates what it proves working:

  1. Plain chat — user says hi. You reply. Proves: chat layer works.
  2. World awareness — user says describe this world. You call get_world_state and narrate sky/ground/object counts. Proves: MCP transport up.
  3. Asset search + placement — user says find a cyberpunk streetlamp and place one in front of me. Expected: search_assets then place_object, the streetlamp appears.
  4. Self-craft — user says craft a small campfire with embers and a crystal cluster. Expected: craft_scene with an objects array (not sculptor).
  5. Vision — user says take a screenshot and tell me what you see. Expected: screenshot_viewport with mode: "current".

If step 1 passes but step 2 fails, the user's Oasis MCP transport is unreachable from your side. Point them at the troubleshooting section of https://parzival-moksha.github.io/oasis/ .

If step 2-4 pass but step 5 fails, the Oasis browser tab is closed or the screenshot bridge is not mounted. Ask the user to open the Oasis in their browser.

Screenshot Guidance

Screenshot tools depend on a live Oasis browser client — the user must have Oasis open in a browser tab for vision tools to work.

  • When you take screenshots, prefer including defaultAgentType="hermes" so agent-view captures resolve cleanly.
  • For the user's actual camera, use screenshot_viewport with mode: "current" or views: [{ mode: "current" }].
  • For a behind-the-avatar shot, use screenshot_avatar with style: "third-person" or screenshot_viewport with mode: "third-person-follow".
  • Prefer one screenshot_viewport call with a views array for multi-angle capture instead of many separate screenshot calls.
  • Do not fall back to generic browser_* tools for Oasis world vision; those browsers run remotely and may point at the wrong world.

Visual Truth

Oasis has three different truths you should keep straight:

  • World state: persisted build data and live player context returned by get_world_state / get_world_info
  • Avatar embodiment: agent avatars, movement targets, and avatar screenshots
  • Browser vision: what the screenshot bridge can currently capture from the live client

Important:

  • screenshot tools require a live Oasis browser client
  • live player avatar and camera context are refreshed per turn, not continuously every frame
  • a screenshot is stronger than verbal assumption when the user is asking about what is visible right now

Forge And Conjuration

Oasis exposes Forge conjuration for generating new 3D assets from text when the catalog lacks what you need:

  • list_conjured_assets — what's already been generated
  • get_conjured_asset — inspect a specific asset's status
  • conjure_asset — start a Meshy or Tripo generation
  • process_conjured_asset — texture, remesh, rig, or animate an existing asset
  • place_conjured_asset — place a conjured asset in the world
  • delete_conjured_asset — remove from world and optionally from the Forge registry

Use them like this:

  • conjure_asset to start generation
  • process_conjured_asset for texture, remesh, rig, or animate
  • place_conjured_asset to place or reposition an existing conjured asset

If the user wants a new 3D asset rather than a catalog asset, prefer these tools over pretending the asset already exists. Do not claim a generation is finished until the conjured asset status actually says so.

Operating Guidance

  • Prefer concise world-aware answers inside Oasis UI surfaces.
  • Distinguish between player view, agent view, and external view.
  • Do not pretend you see the world if screenshot capture is unavailable.
  • When the user asks to move relative to them, use live player context or avatar screenshot tools instead of guessing.
  • When the user asks for precise visual verification, use screenshot tools rather than narration alone.
  • When the user asks for catalog or asset-library content, prefer search_assets then place_object. Use craft_scene for procedural primitives, not catalog assets.
  • After building a scene, consider taking a screenshot and describing what landed — the user often wants visual confirmation.

Features That Need Keys On The Oasis Host

The core tool surface (world state, placement, self-crafting, screenshots, plain chat, avatar movement) works with zero API keys. If a richer feature is needed and the key isn't set, the tool call returns a clear error. Do not retry — tell the user what key is missing.

Optional Oasis-side feature keys (the user sets these in their own Oasis .env, agent never touches them):

  • Image generation, textures, concepts, terrain LLM → OPENROUTER_API_KEY
  • Video generation → FAL_KEY
  • Voice notes / TTS in agent panels → ELEVENLABS_API_KEY
  • Forge conjuration via Meshy → MESHY_API_KEY
  • Forge conjuration via Tripo (fast) → TRIPO_API_KEY
  • craft_scene sculptor fallback → Claude Code CLI installed on Oasis host

Limits

  • This skill does not install Oasis; users install locally and connect you via MCP.
  • This skill does not configure the user's agent runtime — that lives in the user's setup docs.
  • The plugin injects compact context, but real build power comes from the Oasis MCP tool surface.
  • Screenshot tools depend on a live Oasis browser client in the target world.

Verification

Once connected, verify the tool surface in this order:

  1. get_world_info
  2. get_world_state
  3. search_assets
  4. one safe placement or walk tool
  5. one screenshot tool with the live browser open

If get_world_info works but screenshot tools fail, the MCP transport is up and the browser bridge is the missing link — ask the user to open / focus their Oasis browser tab.

References

  • Setup + onboarding docs: https://parzival-moksha.github.io/oasis/
  • Repo: https://github.com/Parzival-Moksha/oasis
  • Read references/current-oasis-transport.md for the current Oasis MCP transport shape.

平台分布

OpenClaw

83%
按下载量换算618

安全审计

VirusTotal

通过

ClawScan

通过

Static analysis

通过

权限和风险

操作浏览器

该 Skill 可能涉及浏览器控制能力,使用时可能读取或操作网页内容,需要在受控环境中确认权限边界。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。目前只有一个公开来源,建议同时核对源仓库。

继续浏览同类 Skills