Gateway Meter

Providers

configured in gateway

Working

✓

verified with API call

Need Keys

⚠

3 broken keys

Total Tokens

⚙

0 requests to date

🔍

OpenSEO Bridge

Open AI & MCP page

OpenSEO UI

HTTP 200

Docker :3004

MCP handshake

HTTP 200

127.0.0.1:3004/mcp

OpenClaw MCP

not linked

SEO orchestrator allowlist updated

DataForSEO

key needed

Required for paid SEO data calls

Autonomous reports

162 offload

/srv/seo-offload

🧠

OpenClaw Brain Update - Aug 23, 2026

162 fleet online

OpenClaw chat fix

Main agent default moved to atlas-coding/deepseek-ai/deepseek-v4-flash.
Default smoke test returned READY with 1M context and no overflow.
Main tool surface reduced from full MCP flood to repo/web/browser/server essentials.
Compaction tightened: 28K reserve, 1K recent tokens, 2 recent turns.

Local model status

162 upgraded to 11 Xeon cores, 41 GiB RAM, 32 GiB swap, and 21 TB storage with about 14 TB free.
Ollama has 43 models; 40 chat-capable models are now registered in OpenClaw at a real 16K context.
qwen2.5-coder:3b passed the private API smoke test and is the responsive local-fleet default.
Heavier 8B-30B models are available for background work; 70B remains CPU-only and very slow.
Kimi K3 2.78T is connected through a repaired private tunnel, but remains experimental, serialized, and capped at four output tokens.

API chat website

Model cache added under /var/www/llm-chat/models-cache.json.
Provider-prefixed routing enabled for Atlas, OpenRouter, Zhipu, and local Ollama.
Atlas Coding Plan shows 15 working models; local Ollama shows 2 models.
Public chat backend test passed through Atlas DeepSeek V4 Flash.

SEO autonomy notes

The six-hour SEO autopilot is now local-first through ollama-162/qwen2.5-coder:3b, with automatic cloud fallback.
The Aug 23 live run completed successfully in about 77 seconds and recorded self-hosted-162/qwen2.5-coder:3b.
seo-orchestrator stays on Atlas/DeepSeek for large-context tool runs and public-content quality.
OpenSEO is linked/visible, but paid SEO data still needs DataForSEO credentials.
Model growth, crawl archives, and reports should use 162; its 21 TB disk has about 14 TB free.

162 Local Model Fleet - Live Test Report

Completed Aug 23, 2026 through the same OpenClaw and API Chat paths used by the UI.

3 of 3 workflows passed

Test	Result	Measured performance	Answer / finding
OpenClaw: 162 Local Model Fleet	PASS No cloud fallback	142s total; 7,494 tokens at 16K context	Returned a concise sitemap/indexing priority. The first shared-lane cold run timed out at 257s; the isolated retry passed.
API Chat: Qwen 2.5 Coder 3B (162)	PASS HTTP 200	49.36s first token; 120.91s total under concurrent CPU load	Produced a valid Toronto mortgage broker title tag. Best current local choice for interactive coding and SEO.
API Chat: Qwen3 8B (162)	PASS HTTP 200	32.87s first token; 263.02s total	Returned three structured checks covering schema, crawlable HTTPS links, and title/description/canonical tags. Better detail, but about 4-5 seconds per token: background use only.

Reliability repairs applied

Interactive traffic now uses a dedicated Ollama service and tunnel on 127.0.0.1:11163.
Background downloads and large jobs remain on 127.0.0.1:11162.
Local chat output is capped at 512 tokens with connection and request timeouts.
Disconnected PHP streams are cancelled instead of occupying the model queue.

Qwen3 adapter repair

The OpenAI-compatible endpoint ignored the non-thinking flag and streamed hidden reasoning.
Local Qwen3 API Chat requests now use Ollama's native chat endpoint with thinking disabled.
The bridge translates native output back to standard API Chat SSE; other models and providers are unchanged.
Queue isolation fixes reliability, but CPU remains shared, so heavy concurrent work still increases latency.

📝

Action Items — What Needs Improvement

Google AI Studio Gemini generation is blocked while prepaid credits are depleted

Add credits/billing at ai.studio/projects or replace the key with one from a funded project

PRIORITY

DeepSeek The key authenticates, but completions return insufficient balance

Top up the account at platform.deepseek.com

PRIORITY

Groq Fast free-tier inference

Add GROQ_API_KEY in the OpenClaw systemd provider override

OPTIONAL

Cerebras Fast free-tier inference

Add CEREBRAS_API_KEY in the OpenClaw systemd provider override

OPTIONAL

NVIDIA NIM Nemotron, Llama, Qwen and vision models

Add NVIDIA_API_KEY in the OpenClaw systemd provider override

OPTIONAL

Mistral Mistral and Codestral models

Add MISTRAL_API_KEY in the OpenClaw systemd provider override

OPTIONAL

Provider Audit

Tested live 2026-10-14 03:04

deepseek

api.deepseek.com 0 models

Error HTTP 401

Authentication Fails, Your api key: ****97bf is invalid...

openrouter

openrouter.ai 462 models

Depleted HTTP 429

Provider returned error...

zhipu

open.bigmodel.cn 11 models

Working HTTP 200

glm-4-flash

google-ai-studio

generativelanguage.googleapis.com 50 models

Error HTTP 404

This model models/gemini-2.5-flash-lite is no longer av...

groq

api.groq.com 0 models

No Key

cerebras

api.cerebras.ai 0 models

No Key

nvidia-nim

integrate.api.nvidia.com 0 models

No Key

mistral-free

api.mistral.ai 0 models

No Key

atlas-coding

127.0.0.1 119 models

No Balance HTTP 402

API error...

local-ollama

127.0.0.1 0 models

Local / internal

ollama-162

127.0.0.1 0 models

Local / internal

Model Browser

sample models from live APIs

Zhipu 8 models

Model ID	Name	Input	Cost
glm-4.5	glm-4.5	text	UNKNOWN
glm-4.5-air	glm-4.5-air	text	UNKNOWN
glm-4.6	glm-4.6	text	UNKNOWN
glm-4.7	glm-4.7	text	UNKNOWN
glm-5	glm-5	text	UNKNOWN
glm-5-turbo	glm-5-turbo	text	UNKNOWN
glm-5.1	glm-5.1	text	UNKNOWN
glm-5.2	glm-5.2	text	UNKNOWN

Total shown: 8 models across 1 working providers

Token Usage

by provider

📈

No token usage data yet

Make API calls through the proxy (port 9089) to see usage here.

Recent Requests

last 0

No requests yet

Gateway

Current Uptime 0s

Last Started -

PID -

Boot ID -

Restarts (10) 0

Last Restart N/A

Config Version -

Proxy Port 9089 (token tracking)

Recent Activity

No activity recorded