Omair's Hermes · Dashboard

Usage

Tokens & Sessions

Total Sessions

211

24.2M tokens processed

Input Tokens

22.3M

cache reads: 374.1M

Output Tokens

1.9M

5,443 tool calls · 5,492 API calls

Cron Jobs

14

14 active · 5,285 runs total

Token Usage — Last 30 Days

Sessions by Platform

Models

Brain Configuration

Model	Sessions	Tokens
glm-5.1	87 sessions	6.2M
glm-5.2	80 sessions	4.5M
glm-5-turbo	27 sessions	11.3M
gemini-2.5-flash	10 sessions	1.4M
glm-5	5 sessions	362.2K
glm-4.7	1 sessions	74.7K
gemma4:latest	1 sessions	355.4K

Activity

Recent Sessions ▼

Session	Source	When	Duration	Tokens	Tools
⏰ TRG Watch — Regular (9AM + 6PM PKT) · Jul 03 04:03	cron	21m ago	3min	21.7K	10
💬 Fix stale send_message references in Hermes skill files	whatsapp	2h ago	—	60.3K	28
⏰ Dojo Overnight Improvement (6am PKT) · Jul 03 01:05	cron	3h ago	6min	62.0K	38
⏰ Essay Backfill (Slow Retry) · Jul 03 01:11	cron	3h ago	12min	6.9K	5
⏰ TRG Watch — Regular (9AM + 6PM PKT) · Jul 02 13:03	cron	15h ago	4min	20.5K	10
⏰ Adil Uncle Debit Card (Thu 2pm) · Jul 02 09:01	cron	19h ago	—	11.5K	0
💬 Hermes creds.json Numbers Explained #2	whatsapp	20h ago	—	570.5K	115
💬 Hermes creds.json Numbers Explained	whatsapp	20h ago	11min	281.7K	32
⏰ Content Pipeline Health Check · Jul 02 08:01	cron	20h ago	—	2.9K	1
⏰ Adil Uncle Debit Card (Thu 1pm) · Jul 02 08:01	cron	20h ago	—	10.0K	0

Automation

Cron Job Health

Job	Description	Schedule	Type	Runs	Last Run
ok WhatsApp ensureJid Patch Watchdog	🔧 Fixes WhatsApp JID mapping bug automatically	`every 360m`	script	×83	1h ago
ok TRG Watch — Regular (9AM + 6PM PKT)	📊 Monitors TRG stock announcements twice daily	`0 4,13 * * *`	skill	×39	18m ago
ok TRG Watch — Jul 1 Escalation (hourly 6-11 PM PKT)	📊 Monitors TRG stock announcements twice daily	`0 13-18 1 7 *`	skill	×6	1d ago
ok Fallback Alert Watchdog	🚨 Alerts when model falls back from primary	`every 10m`	script	×2300	3m ago
ok SeggySaid Video Monitor	🎬 Downloads + transcribes new SeggySaid videos	`every 360m`	script	×51	13m ago
ok Yasar Ahmad Video Monitor	🎬 Downloads + transcribes new Yasar Ahmad videos	`every 360m`	script	×50	2h ago
ok Dashboard Refresh (hourly)	🔄 Regenerates this dashboard hourly	`every 60m`	script	×228	1h ago
ok Dojo Overnight Improvement (6am PKT)	🥋 Overnight skill analysis + improvement suggestions	`0 1 * * *`	LLM	×12	3h ago
ok Dashboard Uptime Monitor	—	`every 5m`	script	×2289	5m ago
ok Content Pipeline Health Check	—	`0 8 * * *`	LLM	×7	20h ago
ok Essay Backfill (Slow Retry)	—	`0 1 * * *`	LLM	×8	3h ago
ok Memory Reconciliation	—	`0 5 * * 1`	script	×1	3d ago
ok Session Capture Review	—	`/30 * * *`	script	×210	21m ago
error Memory System Test Gate	—	`0 4 * * 1`	script	×1	4d ago

Self-Improvement

🥋 Dojo Findings 🔵 6 ▼

Last analyzed: 2026-07-03 01:03 PKT

93.6%

Success Rate -1.7%

64

Sessions

1,156

Tool Calls

74

Total Errors

2

User Corrections

Trend: 11:39: 72.2% → 15:22: 79.5% → 01:03: 79.6% → 01:03: 80.1% → 11:19: 93.5% → 01:03: 95.4% → 01:07: 95.4% → 01:08: 95.6% → 01:03: 96.0% → 01:04: 95.6% → 01:04: 95.9% → 01:03: 95.3% → 01:03: 93.6%

🔵6 Pending

Tool Performance Breakdown

browser_vision

0.0%1 calls, 1 errors

vision_analyze

0.0%2 calls, 2 errors

browser_navigate

25.0%4 calls, 3 errors

patch

74.5%47 calls, 12 errors

search_files

80.9%47 calls, 9 errors

cronjob

85.2%27 calls, 4 errors

write_file

86.7%30 calls, 4 errors

browser_console

87.5%8 calls, 1 errors

execute_code

89.1%64 calls, 7 errors

terminal

91.0%565 calls, 51 errors

session_search

94.1%17 calls, 1 errors

process

94.9%39 calls, 2 errors

web_search

95.4%108 calls, 5 errors

read_file

97.4%76 calls, 2 errors

web_extract

100.0%69 calls, 0 errors

▶ WARNING vision_analyze / browser_vision — 100% failure rate (3/3 calls, content filter) general

Z.AI content filter returns HTTP 400 error code 1301 for all image analysis requests. Same pattern as previous 3 days — this is a persistent provider-level restriction on book cover / content images.

Impact: Vision-dependent workflows are non-functional. Agent wastes tool calls on guaranteed failures.

🔵 PENDING

Fix: 1. Skip vision_analyze entirely for images likely to trigger content filters (book covers, social media images) 2. Add pre-check that catches content filter errors and falls back to text-only description 3. This is a provider issue — no agent-side fix possible

Evidence

Books for High Testosterone Men List (whatsapp)

Error code: 400 - contentFilter level 2, code 1301

▶ WARNING browser_navigate — 75% failure rate (3/4 calls) general

ERR_HTTP_RESPONSE_CODE_FAILURE from target websites during the 'Books for High Testosterone Men List' session. All 3 failures from a single session trying to navigate to various URLs.

Impact: Browser-based workflows unreliable for certain sites. Session wasted 10+ tool calls on failed navigation attempts.

🔵 PENDING

Fix: 1. Use web_extract as primary method; only use browser_navigate when JavaScript rendering is required 2. Add 1 retry with exponential backoff; if second attempt fails, fall back to web_extract 3. Already documented in common-agent-pitfalls.md #4

Evidence

Books for High Testosterone Men List (whatsapp)

Navigation failed: net::ERR_HTTP_RESPONSE_CODE_FAILURE

▶ INFO cronjob — 85.2% success rate (4/27 failures) hermes-cron-jobs

3 failures from 'no_agent=True requires a script' error (persistent pattern), 1 failure from stale job_id reference across sessions.

Impact: Cron creation workflows break mid-task. The no_agent error triggers same-tool-failure loops (3 retries each).

🔵 PENDING

Fix: 1. Already documented in common-agent-pitfalls.md #2 and #8 2. Frequency increased from 3→4 this week despite documentation — agent may not be reading pitfalls doc consistently 3. Consider adding validation check in cron creation flow

Evidence

Segi Said Video Transcription & Platforms (whatsapp)

create with no_agent=True requires a script

Hermes creds.json Numbers Explained #2 (whatsapp)

Job with ID or name 'a94bc902fd59' not found

▶ INFO execute_code — 10.9% failure rate (7/64 failures) general

Python tracebacks in sandbox scripts: import errors, type errors, and API timeout issues.

Impact: Low impact — sandbox errors are isolated and don't affect other tools.

🔵 PENDING

Fix: Already documented in common-agent-pitfalls.md #6. Continue using try/except wrappers in scripts.

Evidence

Various cron/whatsapp sessions (multiple)

Traceback in sandbox script

▶ INFO terminal — 91.0% success rate (51/565 failures) general

Largest by volume but rate is healthy. Breakdown: timeouts, command-not-found, Python tracebacks. Many are from monitoring scripts and one-off analysis commands.

Impact: Low relative impact — 91% success on the highest-volume tool. Most errors are in non-critical scripts.

🔵 PENDING

Fix: Already documented in common-agent-pitfalls.md #5 and #6. Use background mode for long-running scripts.

▶ LOW User correction: WhatsApp delivery to wrong contact (Abdullah) hermes-dojo

User reported Dojo reports going to Abdullah instead of Omair. Root cause: previous Dojo runs used curl to localhost:3000/send with wrong chatId from creds.json.

Impact: Privacy/confidentiality issue — reports sent to wrong person. Already fixed.

🔵 PENDING

Fix: 1. Always use send_message tool with target='whatsapp' 2. Documented in common-agent-pitfalls.md #9 3. SKILL.md already updated to prohibit curl-based delivery

Evidence

Segi Said Video Transcription & Platforms (whatsapp)

You are sending dojo reports to abdullah on WhatsApp when i should be getting these here

Actionable Recommendations

→

Vision content filter is persistent (5th day) — consider pre-emptively skipping vision_analyze for image categories likely to be flagged

→

Cron no_agent error frequency increased despite documentation — agent may need the check enforced at tool level rather than doc level

→

Success rate dip from 95.3%→93.6% is primarily from low-volume tool failures (vision, browser) — high-volume tools remain stable

External Services

Integrations

🔍active

Langfuse Observability

cloud.langfuse.com
Live tracing of every LLM call, tool invocation, latency, and token cost. Waterfall views per session.

🧠active

Mem0 Memory Provider

app.mem0.ai
Semantic memory — no char limits, automatic fact extraction, cross-session recall.

☁️active

Cloudflare Pages

omair.pages.dev
CDN hosting for dashboard, library, books, and transcripts. 0.07s TTFB via Dubai edge.

🔎active

Tavily Search

tavily.com
Real-time web search with answer synthesis. Returns structured results with snippets. Free tier, 1K searches/month.

🛡️active

Security Hardening

Bundled security-guidance plugin + custom security skill. File-write warnings, prompt injection defense, credential hygiene.

📋active

Kanban Task Board

SQLite-backed multi-agent task board with auto-decomposition. Access via:
  1. Through me: say "create a task" or "what's on the board?"
  2. Web UI: ssh -L 9119:localhost:9119 ubuntu@VPS then open localhost:9119
  Auth: omair / khaadim2026

🧩active

MCP Consolidated

13 servers → 5 (2 merged + 3 standalone). Google Suite (30 tools) + Local Services (11 tools). Saves ~500MB RAM.

🔗active

Composio Connect

dashboard.composio.dev
1,623 tools across 11 apps. Free tier: 20K calls/month.

Available apps: Gmail, GitHub, Notion, Slack, Linear, Jira, HubSpot, Twitter, Airtable, Firecrawl, Perplexity

To connect a new app:
  1. Go to dashboard.composio.dev → Connected Accounts
  2. Click the app you want → OAuth login → authorize
  3. Done! Say "send a Slack message" or "search my Notion" in WhatsApp

To add MORE apps (Slack, Notion, etc. not listed?):
  Say "add [app name] to Composio" in chat — I'll create a new MCP server config with that toolkit

Usage: 0 / 20,000 calls this month | Server ID: 244e9609

🔐active

Off-box Backup (DR)

GitHub + Backblaze B2
age-encrypted daily backup of the brain (.env/config/scripts), mem0, and state.db to two off-box vendors. 3-2-1 disaster recovery.

🔀active

Groq Dual-Path Routing

Local proxy (:8788) tries api.groq.com direct, auto-fails-over through Cloudflare WARP on a 1010 ASN-ban. Powers off-CPU voice STT + the Groq LLM fallback — neither path is a single point of failure.

Content

Library Status

Scontent

SeggySaid

155 videos · 155 essays · 65.1K words
100.0% essay coverage
📕 Book · 📖 Transcripts

Ycontent

Yasar Ahmad

774 videos · 773 essays · 354.5K words
99.9% essay coverage
📕 Book · 📖 Transcripts

📚library

Master Landing

omair.pages.dev/library
Complete cross-linked library with book + transcript pages

Procedures

Skill Inventory — 56 Skills ▼

airtable (1)

productivity Airtable REST API via curl. Records CRUD, filters, upserts. 21d ago

api-integration (1)

software-development > 8d ago

architecture-diagram (1)

creative Technical diagrams: dark SVG architecture/cloud diagrams (HTML), Mermaid for markdown/Obsidian, and 5h ago

arxiv (1)

research Search arXiv papers by keyword, author, category, or ID. 21d ago

ascii-art (1)

creative ASCII art and video: text banners (pyfiglet), message art (cowsay), decorative borders (boxes), imag 5h ago

blogwatcher (1)

research Monitor blogs and RSS/Atom feeds via blogwatcher-cli tool. 21d ago

claude-design (1)

creative Design one-off HTML artifacts (landing, deck, prototype) AND author formal DESIGN.md token spec file 5h ago

coding-agent-delegation (1)

autonomous-ai-agents Delegate coding tasks to external CLI agents (Claude Code, Codex, OpenCode). Covers one-shot print m 7d ago

comfyui (1)

creative Generate images, video, and audio with ComfyUI — install, launch, manage nodes/models, run workflows 21d ago

content-pipeline-health (1)

devops Build, maintain, and health-check content pipelines (SeggySaid, Yasar Ahmad books) — essay generatio 13h ago

custom-mcp-servers (1)

autonomous-ai-agents Build in-house MCP servers using FastMCP + stdlib to replace broken, premium-locked, or missing thir 9d ago

database-operations (1)

software-development > 9d ago

deployment (1)

devops > 7d ago

gif-search (1)

media Search/download GIFs from Tenor via curl + jq. 21d ago

git-operations (1)

software-development > 9d ago

github-operations (1)

github Complete GitHub workflow: auth, repo management, PR lifecycle, issues, code review, and codebase ins 7d ago

google-workspace (1)

productivity Google Workspace operations via MCP and API: Calendar events, Drive file management, Docs/Sheets/Sli 7d ago

hermes-agent (1)

autonomous-ai-agents Configure, extend, or contribute to Hermes Agent. 1d ago

hermes-cron-jobs (1)

productivity Create, configure, and debug Hermes cron jobs — toolset/MCP integration, scheduling, delivery, and c 2h ago

hermes-model-config (1)

autonomous-ai-agents Switch, upgrade, and validate Hermes LLM models and providers. Manage fallback chains, add new provi 9d ago

hermes-operations-dashboard (1)

autonomous-ai-agents > 8d ago

himalaya (1)

email Himalaya CLI: IMAP/SMTP email from terminal. 21d ago

huggingface-hub (1)

mlops HuggingFace hf CLI: search/download/upload models, datasets. 21d ago

humanizer (1)

creative Humanize text: strip AI-isms and add real voice. 21d ago

jupyter-live-kernel (1)

data-science Iterative Python via live Jupyter kernel (hamelnb). 21d ago

kanban-orchestrator (1)

devops Decomposition playbook + anti-temptation rules for an orchestrator profile routing work through Kanb 21d ago

kanban-worker (1)

devops Pitfalls, examples, and edge cases for Hermes Kanban workers. The lifecycle itself is auto-injected 21d ago

llm-wiki (1)

research Karpathy's LLM Wiki: build/query interlinked markdown KB. 21d ago

maps (1)

productivity Geocode, POIs, routes, timezones via OpenStreetMap/OSRM. 21d ago

market-data (1)

productivity Look up stock prices, PSX/KSE market data, financial news, and company fundamentals for Pakistani an 9d ago

memory-management (1)

productivity > 3d ago

ml-evaluation (1)

mlops ML evaluation and experiment tracking: benchmark LLMs with lm-eval-harness (MMLU, GSM8K, HumanEval) 5h ago

notion (1)

productivity Notion API + ntn CLI: pages, databases, markdown, Workers. 21d ago

obsidian (1)

note-taking Read, search, create, and edit notes in the Obsidian vault. 8d ago

ocr-and-documents (1)

productivity Extract text from PDFs/scans (pymupdf, marker-pdf). 14d ago

openhue (1)

smart-home Control Philips Hue lights, scenes, rooms via OpenHue CLI. 21d ago

p5js (1)

creative Creative animations and browser art: p5.js sketches (gen art, shaders, interactive, 3D), Pretext tex 5h ago

plan (1)

software-development Plan mode: write an actionable markdown plan to .hermes/plans/, no execution. Bite-sized tasks, exac 21d ago

polymarket (1)

research Query Polymarket: markets, prices, orderbooks, history. 21d ago

popular-web-designs (1)

creative 54 real design systems (Stripe, Linear, Vercel) as HTML/CSS. 5h ago

powerpoint (1)

productivity Create, read, edit .pptx decks, slides, notes, templates. 21d ago

requesting-code-review (1)

software-development Pre-commit review: security scan, quality gates, auto-fix. 14d ago

research-paper-writing (1)

research Write ML papers for NeurIPS/ICML/ICLR: design→submit. 21d ago

security-hardening (1)

software-development Security best practices and threat model for Omair's Hermes Agent setup on Oracle Cloud VPS with aut 9d ago

songsee (1)

media Audio spectrograms/features (mel, chroma, MFCC) via CLI. 21d ago

songwriting-and-ai-music (1)

creative Songwriting craft, Suno AI music prompts, and programmatic AI music generation (AudioCraft/MusicGen, 5h ago

spike (1)

software-development Throwaway experiments to validate an idea before build. 21d ago

systematic-debugging (1)

software-development 4-phase root cause debugging: understand bugs before fixing. 7d ago

systemd-ops (1)

devops Manage Hermes systemd services, gateway restarts, systemd units, and VPS operations on Oracle Cloud 1d ago

teams-meeting-pipeline (1)

productivity Operate the Teams meeting summary pipeline via Hermes CLI — summarize meetings, inspect pipeline sta 21d ago

test-driven-development (1)

software-development TDD: enforce RED-GREEN-REFACTOR, tests before code. 7d ago

trg-watch (1)

research > 3h ago

video-transcription (1)

media Download and transcribe videos from any platform (YouTube, Facebook, TikTok, Instagram, etc.) using 3d ago

web-retrieval (1)

research > 2d ago

whatsapp-messaging (1)

productivity Send WhatsApp messages to contacts and groups — direct bridge access, contact resolution, and outbou 2h ago

xurl (1)

social-media X/Twitter via xurl CLI: post, search, DM, media, v2 API. 21d ago

Memory

Persistent Stores

🧠 Mem0 Cloud unlimited · semantic

Mem0 Cloud (app.mem0.ai) — Semantic search, auto fact extraction, cross-session recall. No char limit.

Local MEMORY.md (prompt injection) 4,982/5,000 chars

99.6% used · injected into system prompt each turn

Local USER.md (prompt injection) 2,204/2,500 chars

88.2% used · injected into system prompt each turn

Connections

Platform Status

1connected

Whatsapp

Active and receiving messages

2connected

Discord

Active and receiving messages

Topology

System Architecture

Messages flow from WhatsApp/Discord/Email through the gateway to GLM-5.2, which reaches consolidated MCP servers, Composio, and cloud services — all self-hosted on Oracle VPS.

📱

💬

Discord

📧

Email

▼

Hermes Gateway

systemd · allowlisted · 3 platforms

▼

🧠 GLM-5.2

Persona · Tool-use · Reasoning · Vision (GLM-4.5v)

Fallback: 5.1→5-turbo→4.7→4.5-air→gemini→groq

▼ ▼ ▼

📦 Consolidated MCP (5 servers)

google_suite.py · 30 tools

local_services.py · 11 tools

clickup · docintel · context7

13→5 consolidation · -500MB RAM

🔗 Composio Connect

1,623 tools · 11 apps

Gmail · GitHub · Notion

Slack · Linear · Jira

HubSpot · Twitter · Airtable · Firecrawl · Perplexity

OAuth-based · no browser needed

☁️ Cloud Services

🧠 Mem0 · semantic memory

📊 Langfuse · observability

🔎 Tavily · AI search

☁️ Cloudflare Pages · hosting

All on free tiers

▼

Google APIs

Calendar · Drive · Docs · Sheets · Slides · Tasks · Contacts

1 OAuth client · 1 token · 0600 file

Oracle VPS

systemd services · ~/hermes · ARM64

● Teal = in-house · self-hosted ● Gold = third-party managed ● Red = brain & senses

Trust & Data

Security Posture

🏠 Data on your turf

Everything runs on your Oracle VPS. No third-party server sits between the agent and Google APIs. All adapter code is auditable in ~/.hermes/custom-mcp/

🔐 Minimal OAuth scopes

One Google OAuth client with only the scopes each feature needs — Calendar, Drive, Docs, Sheets, Slides, Tasks, Contacts. Nothing broader. No admin/Workspace powers.

🔑 One token, well guarded

A single refresh token in a 0600 file behind a shared auth helper. No secrets in config.yaml. API keys isolated in .env.

🔒 Allowlisted access

WhatsApp, Discord, and Email are allowlisted to your number and addresses only. No one else can talk to it. Password gate on dashboard (khaadim2026).

🛡️ Security hardening active

security-guidance plugin warns on dangerous file writes. Custom threat-model skill covers prompt injection, data exfiltration, and credential hygiene. ClickUp: no destructive actions.

⚠️ Known gaps

No host firewall (ufw) yet. No fail2ban. Leaked Vercel/Discord tokens need rotation. No offsite backups. Home dir has 1.4GB phone backup clutter.

What's Next

Roadmap & Limitations

✅ Recently Shipped (June 2026)

GLM-5.2 upgrade
From 5.1, with reasoning model

MCP consolidation
13→5 servers, -500MB RAM

Composio Connect
1,623 tools, 11 apps

Mem0 + Tavily
Semantic memory + AI search

Surge→CF migration
All 6 pages on Cloudflare

Dashboard v2
Charts, Dojo findings, architecture

💡 Ideas to Explore

Host firewall (ufw)
Block rpcbind/Syncthing exposure

Token rotation
Regen Vercel/Discord/Composio keys

Gmail API
Labels, threads, search, triage

Offsite backups
Encrypted, automated, remote

Dojo improvements
Fix cronjob + retry loop findings

Home dir cleanup
1.4GB phone backups to archive

▲ Current Limitations

Unverified OAuth app — Google shows a one-time consent warning and a 100-user cap (you're the only user)

Google Keep unavailable — API is Workspace-only, won't work with a personal account

Maps-lite is driving-only — OpenStreetMap routing covers car routes; walking/transit not wired

External tools (web_extract, browser_navigate) depend on target site availability — mitigated by Tavily + Composio

Composio apps need per-app OAuth — tools exist in schema but return "not authenticated" until connected

Hermes Dashboard

Tokens & Sessions

Brain Configuration

Recent Sessions ▼

Cron Job Health

🥋 Dojo Findings 🔵 6 ▼

Integrations

Langfuse Observability

Mem0 Memory Provider

Cloudflare Pages

Tavily Search

Security Hardening

Kanban Task Board

MCP Consolidated

Composio Connect

Off-box Backup (DR)

Groq Dual-Path Routing

Library Status

SeggySaid

Yasar Ahmad

Master Landing

Skill Inventory — 56 Skills ▼

airtable (1)

api-integration (1)

architecture-diagram (1)

arxiv (1)

ascii-art (1)

blogwatcher (1)

claude-design (1)

coding-agent-delegation (1)

comfyui (1)

content-pipeline-health (1)

custom-mcp-servers (1)

database-operations (1)

deployment (1)

gif-search (1)

git-operations (1)

github-operations (1)

google-workspace (1)

hermes-agent (1)

hermes-cron-jobs (1)

hermes-model-config (1)

hermes-operations-dashboard (1)

himalaya (1)

huggingface-hub (1)

humanizer (1)

jupyter-live-kernel (1)

kanban-orchestrator (1)

kanban-worker (1)

llm-wiki (1)

maps (1)

market-data (1)

memory-management (1)

ml-evaluation (1)

notion (1)

obsidian (1)

ocr-and-documents (1)

openhue (1)

p5js (1)

plan (1)

polymarket (1)

popular-web-designs (1)

powerpoint (1)

requesting-code-review (1)

research-paper-writing (1)

security-hardening (1)

songsee (1)

songwriting-and-ai-music (1)

spike (1)

systematic-debugging (1)

systemd-ops (1)

teams-meeting-pipeline (1)

test-driven-development (1)

trg-watch (1)

video-transcription (1)

web-retrieval (1)

whatsapp-messaging (1)

xurl (1)

Persistent Stores

🧠 Mem0 Cloud unlimited · semantic