一键导入
ambient-minicpm-v-vision
Local MiniCPM-V visual-understanding provider for Ambient screenshot, UI, game-frame, and image evidence.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Local MiniCPM-V visual-understanding provider for Ambient screenshot, UI, game-frame, and image evidence.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Safely observe the Ambient Blockchain, plan Tool Oracle and x402 interactions, prepare Ambient program deployment work, and run local live-test evidence gates.
Create reusable TinyStyler style profiles from user text examples and apply TinyStyler-style text transfer through Ambient CLI.
Generate raster images through hosted APIs such as OpenAI, Google Nano Banana Pro, fal/FLUX, Replicate, Stability AI, and Ideogram.
Inspect a workspace and summarize notable files for fixture-plugin tests.
Review workspace files for the Pi package fixture.
Local faster-whisper STT provider for Ambient speech input validation.
| name | ambient-minicpm-v-vision |
| description | Local MiniCPM-V visual-understanding provider for Ambient screenshot, UI, game-frame, and image evidence. |
Use this package when Pi needs local visual evidence for a bounded image, screenshot, UI design, game frame, or sampled video frame. Do not use it as the primary reasoning model, and do not ask it to mutate files. The provider returns evidence that Pi/GLM can reason over.
Commands:
minicpm_vision_status
minicpm_vision_verify_runtime_manifest [--archive <local-runtime-archive>] [--binary <extracted-llama-server>] [--artifact-id <artifact-id>]
minicpm_vision_start [--wait-ms 120000]
minicpm_vision_analyze --image <local-image> [--image <reference-image>] --output-json <artifact.json> [--prompt <task>]
minicpm_vision_stop
Operational notes:
openbmb/MiniCPM-V-4_5-gguf:q4_k_m.openbmb/MiniCPM-V-4.6-gguf:q4_k_m; use only when explicitly comparing quality or latency.llama-server with --chat-template chatml; the request body disables thinking where supported, and falls back from strict JSON schema if the runtime rejects that sampler shape.AMBIENT_MINICPM_V_LLAMA_SERVER can bind an existing llama-server binary or a Desktop-installed workspace-owned runtime. If no runtime is bound or discovered, status reports unavailable.minicpm_vision_verify_runtime_manifest validates the managed-runtime manifest and optional local archive checksum. Desktop should use the recommended default managed download for pinned macOS arm64/Linux x64 runtime artifacts, or install a user-approved local archive into .ambient/vision/minicpm-v/runtime after archive/binary checksum verification. Desktop removes macOS quarantine from the managed copy only after checksum verification and records signing/Gatekeeper facts; Windows default download remains disabled until separate Windows evidence lands.AMBIENT_MINICPM_V_ENDPOINT or --endpoint can point to an already-running local OpenAI-compatible MiniCPM endpoint on localhost, 127.0.0.1, or [::1]. Remote endpoints are rejected until Ambient has a separate security-reviewed hosted path covering allowed hosts, user consent, media privacy, secret handling, request redaction, artifact retention, network egress controls, and UI copy.minicpm_vision_analyze accepts one local image path, or two local image paths for a current/reference comparison. Desktop should resolve attachments, browser screenshots, or Ambient media references to approved local files before calling this command. Do not pass arbitrary remote URLs.--output-json artifact. The stdout result is intentionally bounded and should not contain image bytes, secrets, or full raw model output.Safety and quality boundaries: