This section documents what was installed, tested, and
catalogued in the legacy OpenClaw inventory. Runtime status above now comes only from 162; old 192 gateway findings remain here as migration evidence until each tool is reconciled.
Providers (8 configured)
OpenRouter working
17 free models installed (incl. Lyria 3 Pro / Clip preview, $0). Live chat verified.
Zhipu working
8 models (GLM-4.5 → 5.2 + glm-4v-flash vision). Live chat verified.
DeepSeek exhausted
Key valid, $0 balance (402). Needs top-up, not broken.
Google AI Studio exhausted
Key valid, free quota depleted (429).
Groq / Cerebras / NVIDIA NIM / Mistral no key
Configured in gateway, awaiting signup API keys.
👁 Eyes (Vision) — working
Vision models flagged with "input":["text","image"] in openclaw.json so the agent can receive screenshots/images.
Zhipu · glm-4v-flash
Multimodal. Used in the E2E eye test below — described a screenshot correctly.
OpenRouter · nvidia/nemotron-nano-12b-v2-vl:free
Free vision-language model.
OpenRouter · nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free
Free omni/reasoning vision model.
OpenRouter · google/gemma-4-31b-it:free
Free Gemma 4 vision model.
✋ Hands (Browser / Actions) — working
OpenClaw's built-in browser tool launches real Chromium via browser.executablePath.
Repaired 2026-08-01: the entire ~/.cache/ms-playwright cache had been removed, so the configured
chromium-1228/chrome-linux64/chrome path pointed at nothing and no browser could launch — while this panel still
claimed "working". Chromium build 1155 was reinstalled, executablePath (and the research profile) repointed at
/root/.cache/ms-playwright/chromium-1155/chrome-linux/chrome, and a real headless launch verified against example.com.
| Capability | Command | Result |
|---|
| Open page | openclaw browser open https://example.com | launched Chromium, returned live page |
| Screenshot | openclaw browser screenshot | captured PNG |
| Vision E2E | screenshot → glm-4v-flash | "Example Domain … for documentation examples" |
🧠 Brain (Local LLM) — working
Updated 2026-08-31: FairChoice SEO now calls Ollama directly on the 162 host over loopback.
The route is private, has no per-request API charge, and keeps generated recommendations in review-only mode.
| Check | Result |
|---|
| Primary SEO model | phi4-mini:latest |
| Fallback SEO model | granite4.2:8b |
| Private API route | http://127.0.0.1:11434/v1 |
| Available local fleet | 46 verified installed models, including Qwen 3, DeepSeek R1, Granite, Mistral, embedding and vision models |
The free bridge is OpenAI-compatible and private to the server/tunnel path. Third-party SEO datasets still require their own provider credentials.
MCP Servers — 25 configured, 21 verified healthy
Audited 2026-07-30 by JSON-RPC handshake against every entry in openclaw.json.
See the live MCP Server Health panel above for per-server detail.
| Group | Servers | State |
|---|
| Verified healthy | search-console-mcp (98 tools), librecrawl (37), lightpanda (31), github-mcp-server (30), firecrawl-mcp (26), playwright-mcp (24), mcp-browser-use (23), mdla-hands (17), browser-use (16), browser-tools-mcp (14), mdla-marketing (8), specification.website (6), tavily-mcp (5), agent-search (4), wet (4), scrapling (3), fairchoice-seo-mcp (3), exa-mcp-server (2), teachable-scraper (1), jina-reader (1), docker-services-mcp (1) | 21 up · 350 tools |
| Blocked on a credential | agentql-mcp (AGENTQL_API_KEY), gsc-mcp (GOOGLE_APPLICATION_CREDENTIALS), brightdata-mcp (API_TOKEN), serpapi-mcp (SerpApi key + rewire from stdio to HTTP) | 4 awaiting keys |
Fixed across this audit: playwright-mcp (forced into HTTP mode by --port), github-mcp-server (pointed at the CLI bin), librecrawl (rebuilt from source as a real HTTP MCP), browser-use (switched to its own maintained MCP server) and jina-reader (repointed to the real published package). Four registrations that were never MCP servers — crawlie, browser-harness, mdla-research-bridge and selenium-mcp — were removed; the underlying tools remain installed.
Primary 162 dashboards
GSC, sitemap, reports, tasks and review-only local AI recommendations.
Public sitemap is served locally and audited by the scheduled sitemap-indexing agent.