Orchestrator.top · the platform, as it runs

Orchestrator.top:
we orchestrate your AI.

Servers, databases, models, agents, software: we make them play together, in-house, and we run them day and night.

3,647agent missions this month
51databases run in production
501scheduled jobs
124agents at work
0people at the keyboard
the orchestrator Trust Layer, which reviews everything that leaves the harness: the checks that verify everything a dot: a part of the platform a line: a real link

The drawing is the real map of our platform, the one our brain keeps of itself: each dot is a part — a server, a database, a function, a scheduled task, an agent, a skill, a lesson learned — and each line, a real link between two parts. Over 4,000 parts, over 7,000 links: everything is connected, and everything goes through the orchestrator at the center, held by the harness — the checks that verify everything — and nothing leaves without going through the Trust Layer, which reviews everything. The same platform installs on your machines, and belongs to you.

Orchestrator.top / what we orchestrate

What we already orchestrate:
our own platform.

Our agents run our servers, our databases, our applications and the software they rely on — voice, transcription, a driven browser — with no one at the keyboard. They can even move an application from one server to another without stopping it: almost a cloud, but yours.

What is orchestrated

Servers added, removed, replaced over the months9
Databases and applications 51 of them running66
Software deployable in one move voice (Kokoro), transcription (Whisper), browser (Chromium)…20
Application moves carried through from one server to another49

What our agents did with it over the last thirty days — 3,647 missions

Monitoring rounds tasks, alerts, backups, certificates, security1 150
Repairs bring a database back, recreate a service, fix a configuration1 058
Diagnoses find out why a site or a service stopped responding658
Research, writing and decisions summarise a watch, draft a report, pick the right agent781

3,410 missions carried through. And it is not only DevOps: our agents also sell for our own websites.

Act 1 · The problem / 1 of 3

AI is a commodity.
It is not in your hands.

Open models — published openly by Mistral, Meta, IBM or Alibaba — are nearly as good as the best, at a fifth of the price. Companies already use 22 times more of them than a year ago. For the first time, top-tier AI can run in-house.

The best open model against the best closed model — November 2026
QUALITY — HUMAN VOTES, ARENA LEADERBOARD best closed model, rented 100% best open model 97% within 3% AVERAGE PRICE — DOLLARS PER MILLION TOKENS best closed model, rented $0.70 best open model $0.13 5 times cheaper WHERE IT RUNS best closed model, rented at its provider's best open model wherever you want — including in-house in-house

Yet most companies still rent their AI from a provider: it changes without notice, you cannot see what it keeps, and when it degrades, you wait.

Sources: Stanford AI Index 2026 (Arena leaderboard, March 2026); Menlo Ventures, enterprise AI report (November 2026). A token is roughly three quarters of a word.

Act 1 · The problem / 2 of 3 — what the law says

And you are the one accountable.

To be useful, AI reads your contracts, your customers, your internal exchanges. Every request sent to a provider is a data export — and the law holds the company responsible, never its provider.

Up to €35M or 7% of worldwide turnover. Transparency mandatory since August 2026; high-risk systems regulated from December 2027AI Act
Up to €20M or 4% of worldwide turnover. The data controller is you — including for what your AI provider does with your dataGDPR
A US provider must hand over the data it is asked for, wherever it is stored. Microsoft France acknowledged it under oath before the French Senate on 10 June 2025CLOUD Act
Health, finance, essential operators: health data hosted in Europe (HDS), digital resilience of banks and insurers (DORA), network security (NIS2)sectors

Keeping AI in-house means keeping your data under your own law.

Sources: Regulation (EU) 2024/1689 and its digital omnibus (2026); GDPR, article 83; CLOUD Act (2018); Microsoft France hearing at the French Senate, 10 June 2025.

Act 1 · The problem / 3 of 3 — in production

In production, one agent in two fails.
Ours succeed 93.5% of the time.

ONE AGENT ATTEMPT — the same unit as the market report
the market6,259 agents · third-party report
56,6 %
us3,647 attempts · 30 days
93,5 %

The rest never goes unnoticed: 4.5% are deliberate stops — an agent going in circles is halted by our safeguards — and the others are missions cut short by an update or by a model that did not answer. Each one has its cause written down.

Why the market fails: running AI in production takes six trades — servers, database, models, memory, backups, governance — sold separately, to be assembled yourselfassembly
Failures whose cause carries a named code, not a mystery100 %

What sets us apart: the models we use are not better — they are the same as everyone else's. What changes is that every failure carries a code, every code feeds a rule, and every rule prevents the next failure.

So why does almost nobody have it? Let's look at who, worldwide, claims to answer that.

Act 2 · The market / Who holds which layer

Running AI takes eight layers.
Who holds which?

PlayerInfra­structureBackupComputeHarnessGover­nanceToolingAgentsQuality & proof
Microsoft Azure, Copilot
Capgemini, Sopra Steria integrators
Mistral models, compute
Scaleway European cloud
Mirantis AI infrastructure
ServiceNow workflows
Salesforce Agentforce
Sierra, Decagon customer-service agents
LangChain developer tools
Orchestrator.top run by agents

strong   partial   absent — based on each player's public offering

The giants cover a lot, with their cloud or their teams. Specialists go deep on one layer. We hold all eight with agents — and we are among the strongest on two: the harness and the proof — combined in the Trust Layer, our quality product.

Act 2 · The market / Where the value goes

AI becomes a commodity.
Good operations become scarce.

What commoditizes: AI itself

Models, inference, writing, coding, summarizing, translating. Cheaper every month, available everywhere, including on your own machines.

What becomes scarce: good operations

Running AI on your premises, keeping it up at night, repairing it without you, proving what it did — and keeping it under your control, so nobody can cut off your tap.

Value has left the model and moved toward what runs it, well, on your premises. Hence our motto: Operations is everything.

Act 2 · The market / What we do

Open models have become excellent.
We put them to work in your company.

We do not build models: Mistral, Meta, IBM and Alibaba publish them openly. Our trade is everything around them that makes them actually work — every day, securely, inside your company.

What it takesWhat we do
AI modelsWe install open models on your machines and pick the right one for each task. A large paid model is called only as a last resort.
Machines, and their upkeepWe set up and run the servers, databases, backups and updates, on the machines you choose.
Business softwareA suite of 10 applications — mail, customers, tasks, invoices… — written to be driven by agents, plus 20 tools ready to install.
A memory that learnsEvery problem solved becomes a knowledge card the agents read before acting: 10,923 cards, used by 64% of missions.
A quality harnessTrust Layer, our quality product: an agent's work is reviewed by other agents before it leaves, and anything irreversible waits for your approval: 1,027 reviews since July, across 64 delivered applications.
GuidanceYou describe your company to the conductor; it proposes a plan, you approve it, and it sets up the team of agents that fits you.

What it changes for you: one partner, one contract, and nothing to assemble.

Act 3 · How it works / An agent at work

Agent = model + harness + your memory.

The model reasons, the agent acts, the harness checks, the memory retains — and you stay in control. The model is interchangeable; the harness and the memory of your trade cannot be copied.

One mission, end to end
1 · MISSION 2 · REMEMBER 3 · REASON 4 · ACT 5 · CHECK 6 · LEARN You decide · stay in control The agent acts · 124 employees, 125 experts The model reasons · router, local first The harness checks · rules, review The memory retains · 10,923 cards you hand overa mission rereads the cardsand the history proposes a planand picks the tool executes, withits tools every commandpasses the rules checksand review your approvalif irreversible the resultbecomes a card ↺ the next mission starts from what was learned

The brain does not reason in place of the model: it gives it what the company already knows. And every mission makes it richer: 64% of missions already rely on a card.

Act 3 · How it works / Our products

Our products are written for our agents.
Our agents, for our products.

Classic software is made to be clicked by a human. Ours expose moves an agent calls directly, over MCP and API: that is why our agents work fast, with no third-party connector.

The productA move the agent callsThe AI employee whose job it is
Mail OneBoxFlow"reply to a customer" send_emailYara · customer support
CRM "create an opportunity" create_opportunityYassine · B2B sales
Tasks Senzo"turn an email into a task" email_to_taskSofia · operations director
Invoicing "issue an invoice" finalize_invoiceOmar · finance analyst
Presentations "generate a presentation" generate_deckHugo · multimedia creative
Campaigns "send a campaign" send_bulk_emailElsa · marketing director

And also: Vault (passwords), Jury (applications), Directories (enrichment), Team (the AI team). 10 applications, 66 moves in all.

Agents move from one product to the next — from mail to CRM, from CRM to invoice. Every new client need becomes an application in the suite; and if you are attached to your own tools, their moves become the agent's.

Act 3 · How it works / Team

You talk to the conductor.
It builds your team and puts it to work.

The conductor is our assistant: an AI you talk to as you would to a partner. Within minutes you have a team — and it gets to work on your projects.

You talkYour business, your projects, what eats your time — as you would to a partnerinterview
It understandsIt asks the right questions and spots what a team can take ondiagnosis
It proposesA plan: which roles, for which projects. You approve or correct itplan approved
It recruitsThe AI employees arrive already trained in their trade, in your team roomteam in place
The team worksAn AI director runs the team and sets priorities; everyone moves forward and keeps you postedat work
You receiveDeliverables: studies, contracts, offers, copy. You approve or send back, the team correctsdeliverables
Agents at work, spread across 74 roles124
Experts in the House of Experts and the 1,730 pieces of trade knowledge they share125

A real example, on one of our own companies — a shuttle service launching in Spain: the team delivered a sales strategy, a draft subcontracting agreement, a price grid and copy for the website. Each deliverable was approved or sent back by the manager, and everything it produces stays with you.

Act 4 · The proof / beyond DevOps

Our agents also sell
for our own websites.

What we built for our own companies — our shuttle websites, 24h00.info, our events platforms — has become a sales module:

Run the websitesOne tool to manage WordPress sites and front-ends. About thirty domains run on our fleet.
MeasureEvery site's visits, counted each night, and the journeys that lead to a sale.
WatchA daily watch on each site's topic: what people say, what people search for.
Write every dayContent created every day on the topic, reviewed before publication — turned down when it is not good enough.
ImprovePages, copy and journeys adjusted to what actually sells.

We use this module every day for our own sales. We offer to plug it into yours.

Act 4 · The proof / One example, in slow motion

3:14 a.m. An instance goes down.
Nobody gets woken up.

signal: outage detected → repaired → verified typical case
  1. 03:14:07sensorclient instance incomplete — 3 services missing
  2. 03:14:07diagnosiscause recognized from memory — seen 4 times beforememory
  3. 03:14:08governanceaction classified reversible → execution authorized61 ms
  4. 03:14:08backuptaken before intervention, encrypted, off-site2 s
  5. 03:14:10repairthe 3 services are recreated and restarted41 s
  6. 03:14:51verificationintegrity checks — all green3 s
  7. 03:14:54logwho decided, on what basis, with what outcometimestamped
  8. 03:14:54closed47 s · 0 external calls · 0 people at the keyboard

This is just one example: infrastructure operations. The same sensors, the same harness and the same memory also run sales, invoicing and support.

Act 4 · The proof / Trust Layer, 1 of 2 — the rules

Our rules are written
— and wired.

The Trust Layer, our quality product, starts with rules. A rule is a lesson learned from a real failure: "never do this — here is what happened last time". It is written for the agent, and wired: a checkpoint verifies it is followed, at every step of the work.

a request 1 At start-up rereads the goal and its rules 2 Before each command refuses danger e.g. wiping prod 3 Before delivery 1,712 checks ≈ 17 min, every time 4 The review agents review it pass, or block 5 Irreversible steps wait for your OK one word, by phone → DELIVERED, WITH ITS PROOF
Rules — the guardrails "never do this"155
Skills — the know-how "here is how it's done"1 402
Precise questions asked of the code by the 1,712 checks, before every delivery69 898
Deliveries stopped at review on our own work164

Why it matters to you: an instruction you politely ask an AI to follow, it eventually forgets. A checkpoint doesn't — and when it refuses, the work doesn't ship.

Act 4 · The proof / Trust Layer, 2 of 2 — the verdict and its proof

Trust Layer: no work ships without proof.

Checks1,712 automated checks, 69,898 verifications, replayed at every deliverygreen, or nothing ships
ReviewChecking agents review the work — the higher the risk, the stricter the reviewscore /100
VerdictAccepted, sent back, or blocked — with the file that justifies it1,027 issued
Follow-upsWhat was flagged and never fixed comes back, againup to 104 reminders
TrialsAgents are re-evaluated on frozen cases, continuously874 over 30 days
Verdicts rendered on our own work 519 accepted · 344 sent back · 164 blocked1 027
Defects tracked until closed, never lost along the way318 · 256 closed

The Trust Layer is sold on its own: plug in your own agents, and every piece of work gets a verdict and the file behind it. The company's question is not “can it do the job”, it is “who reviewed it, against what scale, and what can I show my auditor”.

What sets us apart: Everyone tests what they build. Almost no one can hand you the file that says a piece of work was judged, by whom, against what scale, and why it passed.

Act 4 · The proof / governance

Decisions arrive where you already work.

Approval channel — Slack, email, webhook or Telegram
Morning brief: 12 instances swept, integrity verified. Nothing irreversible, I'll keep you posted.
Database migration ready and tested. Sensitive action → I'm waiting for your GO.
GO ✓
Applied. Checks green, backup taken before switchover. Full report in the dashboard.
Actions subject to human approval before being carried out since August 16324
Granted the others waited, or were refused133
Irreversible actions taken without approval0
Outbound calls logged — host and path, never the request since August 15416 720

The figure that matters is not 324. It is 133. The system asks far more often than it is told yes — and it waits.

Approval goes to your company's channel: Slack, email, internal webhook — or Telegram. The executive answers in a word, where their team already talks.

Act 4 · The proof / memory first, the model last

Every outage solved becomes a card.
Every card saves a call to the model.

Knowledge cards — a problem encountered, its cause, its solution 1,086 added this month10 923
This month's missions that drew on these cards 2,353 out of 3,64764 %
Reasoning calls handed to a model, per day ≈ 600 in September, ≈ 90 today÷ 7

The model is not the first reflex: it is the last resort. First a rule, then memory, then the models installed at your company — and only after that, someone else. What repeats becomes a written routine, and a written routine calls no one. The outside tier can be switched off, and everything that leaves your walls is logged.

Act 4 · The proof / infrastructure and compute

Your stack, on your metal — or ours.

Servers, databases, HTTPS, sovereign DNS, domains — set up, maintained and repaired without you. On your machines, on ours, or with the provider of your choice: you decide, the bill stays in your name.

Databases we maintain in production51
Compliance check of each database, continuously average score90/100
Applications repatriated from a foreign vendor25
Successful backups over the last 7 days, the latest today 0 failures236
Specialised hardware needed to run the models a server, or an ordinary desktopnone

What sets us apart: Others sell you the machine. We give you back the machine and its operation.

Act 4 · The proof / standards

Built to industry standards.

Thirty-seven frameworks, thirty-six domains, filed under NIST CSF 2.0. They are not logos on a slide: they are the product's blueprint.

Our mechanisms filed under the six functions of NIST CSF 2.0
IDENTIFY 448 functions inventoried 7,916 components mapped PROTECT 155 wired rules secrets in vault, egress filtered DETECT 15 sensors running continuously issues handled continuously RESPOND automatic remediation 4,282 proposals RECOVER 236 backups / 7 d restores replayed GOVERN 324 approval requests · 416,720 outbound calls logged · the quality review

Our rule: a framework is counted only when an automated check verifies it — writing the name of a standard is not enough.

What sets us apart: anyone claiming compliance with everything has not shown you their map.

Act 5 · What you do with it / the catalog

Eight stacks. Five you can take on your own, starting today.

Most companies already have a piece. Infrastructure, compute, tooling, agents and the Trust Layer can be taken on their own. Backup, harness and governance come with the platform, run by us.

What you take on your own, and what comes with the platform
VERTICALLY — ONE FACTORY Infrastructure set up and maintained Backup & recovery encrypted, off-site Compute local first Harness 155 wired rules Governance 324 approvals requested Tooling 20 modules Agents 124 in service Trust Layer — quality & proof 1,027 verdicts ✓ standalone with the platform ✓ standalone with the platform with the platform ✓ standalone ✓ standalone, via Team ✓ on its own

Act 5 · What you do with it / Where to start

Where to start?
Seven ways to work with us.

A team of agentsYou describe your business to the conductor; it builds the team and puts it to work. Everything it produces stays yours.
Judge your agents' workPlug your agents into the Trust Layer: every piece of work gets a verdict — accepted, sent back or blocked — and the file behind it.
Repatriate an applicationYour application runs at a foreign vendor: we move it onto a machine of yours — database, accounts, files, domain. 25 done already, with no service interruption.
Start a projectA complete backend set up on your server or ours, in seven minutes — backed up, monitored, repaired without you.
Keep your servers runningYou keep your servers and your provider; we bring the operations — monitoring, repairs, backups.
Install a toolError tracking, secrets vault, customer support, dashboards… 20 modules installed at your company, with no vendor subscription.
Train your teamsNot “how AI works”, but what you can hand it, and how to keep it in check.

Act 5 · What you do with it / The first step

One workflow. Thirty days.
One proof file.

We don't start with a migration. We take a single one of your workflows — the one that costs you the most manual work — and set it up on your infrastructure, supervised and measured.

Week 1The workflow is mapped: inputs, outputs, and what will have to be provenscoping
Week 2The infrastructure is set up on your machine — backed up, monitoredfoundation
Weeks 3-4The agents work the workflow, under your approval, everything is loggedoperations
ExitThe file: what was executed, authorized, judged — and what remains openproof

What you keep, even if you stop there: the infrastructure is at your company and belongs to you, your data never left, and the proof file is yours. There is nothing to hand back, and nothing to uninstall elsewhere.

What sets us apart: most AI trials end with a demo. This one ends with a file you can put in front of a committee.

Act 5 · What you do with it / the catalog, and what surrounds it

Sovereign doesn't mean closed.

Twenty modules, installable at your company in one step, that you own. And around them, nothing that locks you in.

GlitchTiperrors · Sentry
Loki + Grafanalogs
Uptime Kumauptime
Metabasebusiness intelligence
Chatwootcustomer support
OpenProjectproject management
OpenSEOSEO
ntfynotifications
Vaultwardensecrets vault
Giteagit forge
Supabasecomplete backend
Capscreen recording
DoclingPDF reading
SearXNGweb search
Whispertranscription
Kokorospeech synthesis
Ollamalocal models
Hermesagent
OpenClawpersonal agent
Browserautomated Chromium
The harnessquality & proof
Your infrastructure, or the one you want. Your machines, ours, or a provider — AWS, OVH, Hetzner, Scalewayyour choice
Open models, or others. Open models run on your machines; if you want an external model, it plugs inyour choice
Your existing tools stay. Everything talks over APIs — we add to them, we don't replace theminteroperable
And you can leave. Open building blocks, exportable data, no closed proprietary partreversible

What sets us apart: the sovereignty sold by others is often another lock-in. Here, what you set up, you keep — even the day you leave us.

Act 6 · The vision / The AI stack — and the layer where we stand

Operations is everything.

L'exploitation est tout.

The model is a commodity: available everywhere, at a collapsing cost. Running it in production remains a craft — continuity, recovery, governance, tooling, traceability. That is the craft we operate, and it comes with the platform.

Your data is protected. Encrypted backups every day, off-site copies, proven restorationevery day
You validate before production. Staging environment: real database, real executionisolated staging
You equip your teams. Error monitoring, secrets vault, customer support, business intelligence, search20 modules
You stay in control. Every sensitive action waits for your approvalwithout exception

The whole thing runs on your infrastructure. Demonstration of a complete deployment on request: seven minutes, on a real server.

1 / 21 Book a meetingorchestrator.topshort version↑↓ navigateC short version