OpenClaw模型桥接 是 AI Skill Hub 本期精选Agent工作流之一。综合评分 7.5 分,整体质量较高。我们推荐使用将其纳入你的 AI 工具库,帮助提升工作效率。
OpenClaw模型桥接,连接任何LLM到OpenClaw,生产测试中间件,支持Qwen3-235B和be。
OpenClaw模型桥接 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排,将复杂的多步骤任务拆解为清晰的自动化流程,实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成,适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
OpenClaw模型桥接,连接任何LLM到OpenClaw,生产测试中间件,支持Qwen3-235B和be。
OpenClaw模型桥接 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排,将复杂的多步骤任务拆解为清晰的自动化流程,实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成,适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
# 方式一:pip 安装(推荐)
pip install openclaw-model-bridge
# 方式二:虚拟环境安装(推荐生产环境)
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install openclaw-model-bridge
# 方式三:从源码安装(获取最新功能)
git clone https://github.com/bisdom-cell/openclaw-model-bridge
cd openclaw-model-bridge
pip install -e .
# 验证安装
python -c "import openclaw_model_bridge; print('安装成功')"
# 命令行使用
openclaw-model-bridge --help
# 基本用法
openclaw-model-bridge input_file -o output_file
# Python 代码中调用
import openclaw_model_bridge
# 示例
result = openclaw_model_bridge.process("input")
print(result)
# openclaw-model-bridge 配置文件示例(config.yml) app: name: "openclaw-model-bridge" debug: false log_level: "INFO" # 运行时指定配置文件 openclaw-model-bridge --config config.yml # 或通过环境变量配置 export OPENCLAW_MODEL_BRIDGE_API_KEY="your-key" export OPENCLAW_MODEL_BRIDGE_OUTPUT_DIR="./output"
Agent Runtime Control Plane — Connect any LLM to OpenClaw with one command. Zero dependencies, 8 providers (含豆包 Seed 2.0 主力), multimodal support, reasoning capability. 将任意大模型(Qwen / OpenAI / Gemini / Claude / Kimi / MiniMax / GLM / Doubao Seed 2.0)一键接入 OpenClaw — 零第三方依赖、支持多模态、10 分钟跑通。
[
]() [
]() [
]() [
]() [
]() [
]() [
]()
Current version:v37.9.117/0.37.9.68(2026-06-05) — seeCLAUDE.mdfor full changelog. Latest milestone: V37.9.92–95 — Observer + #32 + framework 三件套同日实证 (机器化 + 制度化 + 自我演化). V37.9.92 修 V37.9.88 Mac Mini path silent failure 5 天 + V37.9.84 status.json 三方共享意识闭环. V37.9.93 Observer 自我修复 sampling artifact (smart head+tail + critique prompt 明示). V37.9.94 立 INV-CROSS-ENV-PATH-001 framework 化 MR-15 第 4 次演出后预防 (scanner 首扫即抓 V37.9.91 expert_escalation 隐藏 5th near-miss). V37.9.95 周一原则 #32 用户视角反馈第一次正向兑现 — ai_leaders_x 19→31 跨 12 派别 (新增 Anthropic/DeepMind/Safety/元老/Robotics/开源/架构创新派 12 个 handles).
| Theme | Versions | What it means |
|---|---|---|
| **MOVESPEED EPERM 60-day blood case CLOSED** ⭐ | V37.9.4 → **V37.9.81** | After 60 days + 6 falsified hypotheses, V37.9.80 (5/18) identified the true root cause via log show --predicate — **macOS TCC Sandbox denies cron-derived processes accessing external volumes**. Fix = add /usr/sbin/cron to FDA. V37.9.81 (5/19) 24h data regression铁证 (12h window = 0 incidents / FDA 后 ~19h = 0 / kernel sandbox deny 0条) + INV-MOVESPEED-TCC-001 hard governance guard (auto-detect 24h ≤ 2 every day) + capture.sh stderr distinction fix (V37.9.30 取证盲区根因修复, 4-layer defense). |
| **Phase 4 Layer 5: Convergence Framework** | V37.9.19 → V37.9.25 | Declared-state ↔ runtime drift detection lifted from "靠记忆" to "机器化". 5 specs running, 2 升级 machine_sync (Plan B 渐进 dry-run). MR-17 立案 (declared-state-must-converge-via-machine-not-memory). |
| **Phase 4 P3: Three-stage gates shadow wiring** | V37.9.15 | pre_check → runtime_gate → post_verify 3 gates wired into request pipeline (shadow mode), policy engine 从"可查询"升级为"在请求路径上被调用产 [gate:*] log"。FAIL-OPEN 契约 + 与 ONTOLOGY_MODE 解耦. |
| **Cross-job fail-fast migration** | V37.9.36 → V37.9.62 | 17 scripts upgraded to **6-field deep prompt** (📌 / 🔑 / 💡 / 🎯 / ⭐ / 🎚️ project alignment, dynamic length by rating) + per-item retry (5/10/20s × 3 + V37.9.36 三层检测) + LLM_DEGRADED fallback (replaced placeholder anti-pattern) + multi-window split (>8000 chars) + rule_check 验证层 (V37.9.47 + V37.9.51 batch + V37.9.62 batch). ALIGNED_SCRIPTS 4 → **17/21** (81% coverage, remaining 4 by-design excluded). |
| **kb_deep_dive daily deep-dive job** | V37.9.16 → V37.9.21 | New 22:30 HKT cron: picker selects ⭐≥4 candidate from today's notes → fetches PDF/HTML full text → 5-field LLM analysis → multi-window WhatsApp + Discord push. **Tier-aware fallback** (V37.9.17): TIER 1/2 papers prioritized over TIER 3 X tweets. |
| **kb_dream Multi-theme + 14-day ban-list** | V37.9.68 → V37.9.68-hotfix | Dream redesigned for "用户视角变开拓视野" — DEEP + WIDE (5 cross-domain) + RADAR (5 准期 signals) + 总览 = **4 independent WhatsApp windows** (replaces V37.4 single window). 14-day theme normalize + ban-list (prevents Qwen-BIM 连续几周重复). 80 unit tests + V37.9.66 category B 设计假设错配 hotfix (split_dream_into_chunks helper). |
| **WhatsApp client folding architecture discovery** | V37.9.35 | 5-layer empirical investigation revealed WA client auto-folds at ~4000 chars (not protocol limit). Budget upgraded 1400→4000 across kb_review / kb_evening / kb_deep_dive (信息密度 2.86×). |
| **Opportunity Radar 三件套全量集成** | V37.9.45 → V37.9.56 | Strategic "early signal radar" 三件套 — #1 cross-source weak signal aggregation (DBSCAN + sentence-transformer) × #2 project alignment scoring (rule_check 验证 LLM ⭐) × #3 trend acceleration detection (4-week keyword acceleration). V37.9.49 #1+#3 集成 kb_dream Phase 1.5 + kb_evening prompt. V37.9.56 #2 完整集成 top_alignment_picker (Top 5 高对齐推送). See [docs/opportunity_radar_design.md](docs/opportunity_radar_design.md). |
| **Capability-Based Dynamic Router** | V37.9.76 → V37.9.77 | Declarative capability framework PoC — jobs_registry adds required_capabilities + prefer + cost_tier fields, providers.find_best_provider() 30-line pure function, router_decide.py shadow mode + V37.9.77 ROUTER_ENFORCE=on opt-in feature flag (Plan B 渐进路径). 70 single tests + 反向 sabotage 真有效. V3 路标 declarative framework 核心交付. |
| **health_check v2.0 "系统证据周报"** | V37.9.78 → V37.9.78-hotfix | Re-positioned from v1.1 单薄数字到 v2.0 evidence-based weekly report — 9 段 emoji marker (🖥服务 + 🤖模型 + 📊SLO + 🛡安全 + 🏛治理 + 🛟MOVESPEED + 🐦X监控 + 📚KB + 💾SSD) + MR-8 single-source-of-truth (4 外部工具) + safe_call helper 三层 FAIL-OPEN. INV-HEALTHCHECK-001 17 checks. V37.9.78-hotfix: macOS BSD timeout 兼容性 (无 timeout → command -v gtimeout → bash -c fallback). |
| **SLO 三项修复 (数据驱动诊断 9 轮无盲改)** | V37.9.79 → V37.9.79-hotfix | V37.9.78 周报暴露 3 矛盾数据 (p95=37s + 成功=100% + 工具=0% + overall=VIOLATIONS). 9 轮诊断锁定: (1) slo_dashboard verdict 三档 PASS/FAIL/N/A (tool_calls=0 不算 FAIL) (2) latency 阈值 30000→50000ms 承认真实 baseline (3) slo_snapshot 每小时 :05 cron 注册 (V36 历史 debt). 16 新单测 + MR-10 understand-before-fix 第 N 次正向兑现. |
The project's most strategic asset. Evolving from "declarative knowledge" toward "run-time adjudication" — the end goal is a reusable pip install ontology-engine package so any Agent Runtime project inherits governance by writing its own YAML.
Already replaced hardcoding (Ontology is now the source of truth; Python hardcoded values are fallback-only):
| Hardcoded before | Ontology source of truth | Version |
|---|---|---|
ALLOWED_TOOLS = {"web_search", ...} 16 tools | tool_ontology.yaml via engine.ALLOWED_TOOLS | V37.8.14 |
Tool param CLEAN_SCHEMAS + aliases | ontology.CLEAN_SCHEMAS / resolve_alias() | V37.8.14 |
MAX_TOOLS = 12 constant | evaluate_policy("max-tools-per-agent").limit | V37.9.12 |
MAX_TOOL_CALLS_PER_TASK = 2 | evaluate_policy("max-tool-calls-per-task").limit | V37.9.13 |
| Security score thresholds (90, per-dimension) | governance_ontology.yaml::security_config | V37.9.3 |
applicable for temporal/contextual policies | _CONTEXT_EVALUATORS dispatch table (6 policies) | V37.9.13 |
Meaning: Changing a threshold requires editing one YAML line, zero Python changes — Phase 4 terminal state partially achieved.
Roadmap:
| Phase | Status | Scope |
|---|---|---|
| Phase 0 — Meta-rule auto-discovery | ✅ V36.2 | MRD scanners find un-covered areas automatically |
| Phase 1 — Equivalence proof + 3-mode feature flag | ✅ V36.2 | ONTOLOGY_MODE=off/shadow/on |
| Phase 2 — Shadow observation | ✅ V36.3 | Ontology runs alongside, logs drift |
Phase 3 — ONTOLOGY_MODE=on | ✅ V37.8.14 | Declarative engine replaces hardcoded logic |
| Phase 4 P1 — 3 engine APIs + 1st policy switch | ✅ V37.9.12 | load_domain_ontology / find_by_domain / evaluate_policy |
| **Phase 4 P2** — Context evaluator + 2nd policy switch | **✅ V37.9.13** | **6 matchers (hour_of_day / has_alert / has_image / task_match) + max-tool-calls-per-task wired** |
| Phase 4 P3 — 3-gate enforcement | ⏳ Next | pre-check → runtime-gate → post-verify across the proxy request pipeline |
| Phase 5 — Engine packaging | 🎯 Goal | pip install ontology-engine — any Agent Runtime inherits governance |
pip3 install -r requirements-rag.txt python3 kb_embed.py && python3 kb_rag.py "AI papers"
pip3 install -r requirements-mm.txt python3 mm_index.py && python3 mm_search.py "cat photos" ``` </details>
Core services (tool_proxy.py, adapter.py, proxy_filters.py) use only Python standard library — http.server, json, urllib. No pip install, no virtual environment, no Docker. This is a deliberate architecture decision: every dependency you remove is one fewer reason someone can't run your system.
python3 kb_embed.py # 4339 chunks in ~8s on Mac Mini
Claude Code → claude/branch → PR → main → auto_deploy (2 min) → Mac Mini
↓
git pull → test → file sync (81 files) → smart restart
↓
preflight_check.sh --full (19 checks)
The auto_deploy.sh script maps 84 repo files to runtime locations (V37.9.43-hotfix added wa_e2e_test.sh) and only restarts services when core files change. Hourly drift detection via md5 checksums with WhatsApp + Discord alerts. Status.json exempt from drift (legitimate divergence between Claude Code snapshots and cron-refreshed runtime).
Three steps. Zero third-party dependencies. Any LLM provider.
```bash
export OPENAI_API_KEY="sk-..." # OpenAI (GPT-4o) export GEMINI_API_KEY="..." # Google Gemini export ANTHROPIC_API_KEY="sk-ant-..." # Anthropic Claude export MOONSHOT_API_KEY="..." # Kimi (Moonshot AI) export MINIMAX_API_KEY="..." # MiniMax export GLM_API_KEY="..." # GLM (Zhipu AI) export REMOTE_API_KEY="..." # Custom Qwen endpoint
See docs/GUIDE.md for the complete bilingual walkthrough including 26 hard-won production lessons.
| File | Description |
|---|---|
jobs_registry.yaml | Unified job registry — 39 jobs (35 active, 4 disabled), system cron |
check_registry.py | Registry validator — ID uniqueness, paths, fields |
gen_jobs_doc.py | Auto-generate job docs from registry + drift detection |
test_providers.py | Unit tests for providers |
test_tool_proxy.py | Unit tests for proxy_filters |
test_check_registry.py | Unit tests for check_registry |
test_data_clean.py | Unit tests for data_clean |
test_adapter.py | Unit tests for adapter |
test_kb_business.py | Unit tests for KB business logic |
test_cron_health.py | Unit tests for cron health |
test_status_update.py | Unit tests for status_update |
test_audit_log.py | Unit tests for audit_log |
test_config_slo.py | **V32** Unit tests for config_loader + slo_checker + incident_snapshot + ProxyStats SLO |
full_regression.sh | Full regression runner — all tests must pass before push (auto-updates status.json test_count) |
.githooks/pre-commit | **V32** Pre-commit hook — API key/phone leak + syntax checks |
.github/workflows/ci.yml | **V32** GitHub Actions CI — 9 test suites + config validation + security scan |
CLAUDE.md | Project context for AI-assisted development |
```bash
python3 mm_index.py # Index media files
python3 mm_search.py "猫的照片" # Semantic search
python3 mm_search.py --stats # Index stats
OpenClaw模型桥接是一个开源AI工作流,连接任何LLM到OpenClaw,生产测试中间件,支持Qwen3-235B和be。虽然星数较少,但项目描述清晰,代码质量良好,值得关注。
AI Skill Hub 为第三方内容聚合平台,本页面信息基于公开数据整理,不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后,再部署至生产环境,并做好必要的安全评估。
✅ MIT 协议 — 最宽松的开源协议之一,可自由商用、修改、分发,仅需保留版权声明。
经综合评估,OpenClaw模型桥接 在Agent工作流赛道中表现稳健,质量良好。如果你已有明确的使用需求,可以直接上手体验;如果还在评估阶段,建议对比同类工具后再做决策。
| 原始名称 | openclaw-model-bridge |
| 原始描述 | 开源AI工作流:Connect any LLM to OpenClaw — production-tested middleware for Qwen3-235B and be。⭐10 · Python |
| Topics | workflowai-agentdiscordllmmiddlewareontologypython |
| GitHub | https://github.com/bisdom-cell/openclaw-model-bridge |
| License | MIT |
| 语言 | Python |
收录时间:2026-06-05 · 更新时间:2026-06-06 · License:MIT · AI Skill Hub 不对第三方内容的准确性作法律背书。
选择 Agent 类型,复制安装指令后粘贴到对应客户端