An AI team, with governance
It is not a chatbot. It is a full architecture: the orchestrator, the specialists, the memory, the cost controls and all the technical complexity, invisible to you.
Architecture
Four pieces, one conversation
One director, twenty specialists, an identity per agent and a memory that does not forget. All behind a single conversation.
Amadeus, the director
The CEO’s single point of contact. He takes the prompts, routes them to the right specialist and consolidates the result. He applies low-spending mode outside working hours, and manages quotas and kill-switches.
Departmental agents
20 departmental agents across finance, sales, marketing, people, operations, legal, IT, quality, analytics, customer support, strategy, engineering, product, communication, events, R&D, support, projects, procurement and cybersecurity. Each with its own tools and semantic memory.
IDENTITY.md per agent
Every agent has a Markdown file holding your brand voice, FAQs, products and processes. It is updated in small, controlled changes, never loses what it learned, and every change is recorded. Agents read their IDENTITY.md at the start of each conversation.
Semantic memory
The agents remember earlier conversations, the decisions taken and the agreements closed. The memory is isolated per company with Postgres RLS and a second middleware layer.
- 1
- A single director: Amadeus
- 20
- Departmental agents
- 4
- Models behind aliases
- 100%
- Data hosted in the EU
Governance
Quotas, modes and automatic kill-switches
You will never get a surprise invoice. Shara measures every token, every call and every cost, and cuts off automatically if something runs away.
STU, Shara Token Units
Every plan includes a monthly STU quota. When it runs out, the next conversation is not interrupted: it moves to overage, billed at your plan’s rate.
First: 50 M STU/month, overage 5 €/M
Pro: 140 M STU/month, overage 5 €/M
Max: 280 M STU/month, overage 5 €/M
Rollover of up to 20 % if you use 80 % or more
Automatic kill-switches
Automatic hard-limit cut-offs. If an agent tries to go past these thresholds, Shara stops the run and notifies the admin.
A single run is cut before it escalates
A spend ceiling on any one-hour window
A hard daily brake, no surprise invoice
Low-spending mode
Outside working hours, it spends less
Low-spending mode kicks in automatically outside the working hours you set: the premium models step down to Sonata / Prelude to cut cost by up to 70 %, without cutting the service.
Models
Models behind aliases, provider-independent
We keep the real LLM model names under NDA. You work with four aliases. If we switch provider for better quality or cost, your integration does not break.
Prelude
Quick work (classification, FAQs, short drafts)
Sonata
Everyday production (email, proposals, analysis)
Symphony
Hard work (strategy, contracts, deep research)
Concerto
Specialised (a secondary provider for specific tasks)
Self-paying engine
Every call costs what it costs
Every LLM call goes through an engine that measures the real cost, applies your plan’s margin and returns the final charge in cents. No opaque flat rates.
That is what removes the conflict of interest: if a conversation is long because the agent needs it to be, the cost goes up. If Amadeus delegates to a cheaper model, the cost goes down. You see the whole breakdown in the dashboard.
final charge = the model’s real cost + your plan’s margin // no flat rate, no surprises // every cent visible in the dashboard
Memory
They learn your company, patch by patch
An editable identity per agent, and a semantic memory that does not lose the thread. Your agents know your business and get better with every job.
IDENTITY.md per agent
An editable Markdown file with your company’s tone, vocabulary, processes, products and FAQs. Each update is a small, controlled change, never loses what it learned, and lands in the audit log with the author, the date and the change. Agents read it at the start of every run.
- Brand tone and vocabulary
- Processes, products and FAQs
- An audited change history
Semantic memory across conversations
The agents remember agreements closed, decisions taken, clients and projects, without you repeating the context. Indexed by embeddings with a configurable TTL and isolation per company.
- Remembers agreements and decisions
- Indexed by embeddings
- Isolated per company (RLS)
Want to know more?
Tell us what your company needs and we will help you design the AI team that fits you best.