MeigaHub MeigaHub
Login Create account
Real data · Recalculated every 24 h

MeigaHub's real benchmarks

Aggregated, anonymous metrics computed on real product usage. Nothing is staged: every figure comes from the production database and is recalculated automatically every 24 hours.

Last updated: 27/08/2026 04:15

Tool calls
41.752
72.8% success
Active users
10
with recorded tool calls
LLM models
29
distinct ones in production
Estimated savings
$37,35
vs 100% cloud

Split local vs cloud

Share of LLM calls answered by a self-hosted local model versus a cloud API (OpenAI/Mistral/etc).

Local (your own server) 6%
2.490 calls · 66.4% success
Cloud (external API) 94%
39.262 calls · 73.2% success

Success rate per tool

Top 12 tools with ≥30 recorded calls. Computed on model_task_outcomes.

Tool Calls % success Split
manage_leads 5.057 100%
fetch_readable 4.643 68.3%
find_local_leads 3.225 84.5%
manage_leads 2.999 89.9%
search 2.246 86.7%
manage_automation 1.858 1.2%
manage_leads 1.226 99.8%
manage_automation 1.208 98%
stealth_browser 1.156 53.7%
manage_leads 865 6.9%
manage_automation 760 67.9%
manage_leads 667 100%

Recorded usage volume

Aggregated counters from usage_tracking — historical totals of real actions.

searches_daily
16.523
web_fetches_daily
6.799
emails_daily
2.223
stealth_sessions_month
1.210
chat_messages_daily
709
desktop_actions_monthly
697
image_analysis_daily
370
public_chat_messages_monthly
99
server_downloads_monthly
18
api_fetches_daily
13

Estimated savings — methodology

We compare the real current cost (only the calls that actually went to the cloud) against a hypothetical scenario in which every one of the LLM calls had run in the cloud.

  • Estimated tokens per call: 1500 (input + reasoning + tool call).
  • Blended cloud price: $10/1M tokens (GPT-4o reference, 50% input / 50% output).
  • Local models are detected by name: they contain qwen / llama / mistral-7b / phi / gemma / .gguf.
If everything ran on cloud
$626,28
Cloud actually spent
$588,93
Saved by going local
$37,35

A conservative estimate. Real savings per user are usually higher, since GPT-4o charges input and output differently, small models add a monthly minimum, and cloud cost grows with context length. These figures exclude electricity and the amortisation of local hardware.

Real cases with auditable metrics

A technical breakdown of the cases shown on the home page. Real sectors and real metrics; names and identifying details withheld for privacy (anonymised with the user's consent). The figures can be reproduced from the product dashboard and from the usage_tracking, model_task_outcomes and automation_rule_runs tables.

CASE 1

Legal practice · REGCON prospecting

A 4-person practice in Galicia (anonymised). Goal: find companies required to file an equality plan under Royal Decree 901/2020 that do not yet appear in the Ministry of Labour's REGCON registry.

MetricValueVerifiable source
Pool companies checked against REGCON2.847company_pool.regcon_checked_at · 6 weeks
Companies with no equality plan (not in REGCON)1.483company_pool.in_regcon = false
Qualified leads created (with contact email)312leads.source = automation
External API cost (Brave Search)€18Brave invoice · 6 weeks
Cloud LLM cost€0Local Qwen3.5 9B server · model_task_outcomes
Hours of manual work avoided~80 hClient estimate (one manual REGCON check ≈ 2 min)
Technical pipeline: manage_leads pool_verify_regcon (province) → manage_leads pool_search only_verified_not_regcon → collect_leads with web enrichment (Brave API) → scheduled email campaign. All orchestrated by the agent on the user's local server.
CASE 2

Marketing agency · SEO blog on autopilot

A 6-person agency (anonymised). Goal: produce recurring SEO content without hiring an outside writer or losing hours to briefing and review every week.

MetricValueVerifiable source
Posts published in 90 days47blog_posts.published_at
Articles in Google's top 30 (primary query)18Google Search Console · 90-day snapshot
Human visits to posts (90 d)3.412page_views.is_bot = false
Local/cloud ratio when writing84% / 16%model_task_outcomes · task tool:write_and_publish_blog_post
Cloud LLM cost~€7/moOpenAI invoice (escalation after a local failure)
Human review time~2 h/weekClient estimate (vs ~12 h/week writing)
Technical pipeline: Daily cron automation rule → search for industry trends → write_and_publish_blog_post with automatic schema markup → cross-channel social publishing. A+ routing learns which local models do well at long-form writing and only escalates to cloud on a quality failure.
CASE 3

Engineering firm · watching BOE / BDNS / PLACSP / TED

An 8-person civil engineering firm (anonymised). Goal: watch public tenders, grants and European calls relevant to their field (civil works, water infrastructure, energy efficiency) every day without assigning a person to it.

MetricValueVerifiable source
Official sources monitored (daily intake)4BOE · BDNS · PLACSP · TED (official APIs)
Opportunities ingested (90 d)2.184opportunities.search_case_id
Opportunities scoring ≥70 (LLM matching)12opportunity_matches.score >= 70
Applications submitted with tailored documents3candidacies.status = submitted · adapted_documents
Cloud LLM cost€0100% local (Qwen2.5 14B for scoring + tailoring)
Human time spent~3 h/weekFinal review of applications (vs full-time manual watching)
Technical pipeline: Search case with boe_api, bdns_api, placsp_atom, ted_api sources → scheduled daily intake → SearchCaseMatchingService batched scoring → chat notification with opportunities scoring ≥70 → generate_adapted_document using the company's profiles (document folders in profile_folders).

Data policy: all three cases are real accounts with anonymised identities. The percentages and aggregates published here cannot re-identify the client. If you are a user and want your case featured (named or anonymised), write to us from contact.

Transparency and reproducibility

Create free account Local hardware requirements