Its instructions
What it must do, in what tone, under which business rules. Written in plain language, versioned, exportable.
For the people who have to say yes
Everything your CIO, your DPO and your CFO will want to know before saying yes.
How a Nexfire agent works, what it can touch, what it cannot do without you, where your data goes, what it costs, and how you leave. No certification we do not hold, no promise the product does not keep.
An agent is six settings and a loop.
Its tools
Only the ones you give it: web search, file reading and writing, document production, media generation, and the connectors you have authorised. It does not see the others.
Its knowledge
One or more document bases, synced from your sources. It searches them by meaning and by keyword, and cites what it found.
Its model
The AI provider and model that make it reason, chosen per agent among fifteen providers. Switchable at any time. The Forge, which assembles an agent through conversation, uses your own AI key.
Its autonomy
Supervised or autonomous, plus the list of actions that require approval: save, delete, send, pay. High-risk actions always wait, whatever the mode.
Its memory
What it retains from one run to the next, inside your workspace, with the retention period you set.
Then, on every task:
- 1
It receives the task
Through chat, through a trigger (email, message, deal, page, meeting, webhook), through a schedule, through the API or from your website.
- 2
It plans and executes
It breaks the task down, calls its tools one at a time, reads your documents, queries your data. Every step is visible in the Studio as it happens, and kept afterwards.
- 3
It produces
A Word, Excel, PowerPoint or PDF deliverable, a media file, or an answer. The code it writes to build a file runs in an isolated sandbox, with no access to your network. The deliverable is visually reviewed before hand-off.
- 4
It stops
Before any sensitive action, it waits in the review queue. You approve, reject or correct. It resumes with your feedback.
- 5
It delivers
Into your Deliverables space, and through the channel you chose: email, Slack, Teams, WhatsApp, Telegram, Discord, webhook.
It only reaches what you have authorised, and you can take it all back.
Connector categories
- Communication
- Meetings
- Support
- CRM
- Finance
- Storage
- Marketing
- Productivity
- Analytics
- HR
- Development
Your applications: 120+, by authorisation
Connection through each application's Authorize button, with only the requested permissions. Nexfire stores no passwords. Access is revoked in one click, on our side or the vendor's.
Your documents
SharePoint, Google Drive, Notion, Confluence, Jira, or direct upload of text, Markdown, PDF, Word and Excel files. Scheduled sync. Connection credentials are encrypted in a vault and never shown in clear text.
Your databases
PostgreSQL, MySQL, SQLite, through a connection server you configure with an account whose permissions you choose. You can plug in your own tool servers, locally or over HTTP, with header, token or key authentication.
Your systems
An API with personal tokens scoped to read or write, fixed at creation, capped in number. Inbound integrations tested before activation. Signed webhooks, with token rotation and a call log.
Your meetings
A bot joins Meet, Teams, Zoom or Webex only after explicit consent in your workspace. It is visible to participants. You choose the processing region.
Your website
An assistant in three lines of code, restricted to the domains you declare, with revocable keys, a cost cap and abuse protection.
It does the work, you approve
The brake is in the engine, not in the settings.
Every high-risk action waits
Every tool carries a risk level. Above a threshold, execution pauses and waits for a human decision. That threshold cannot be switched off, even in autonomous mode. An admin cannot forget it, an agent cannot bypass it.
Send, pay, delete, save
Four families of supervised actions, adjustable agent by agent. Autonomous mode can only relax these, never the risk threshold.
It only spends within a mandate
An agent can only pay within a mandate you approve, with a monthly cap. Each workspace's AI budget is tracked in real time.
It asks when it does not know
A clarifying question interrupts execution and is addressed to you. It does not guess, it does not fill in the blanks.
Paid capabilities are enabled by an admin
Image, video, voice and music generation: off by default, enabled only by a workspace owner or admin.
The meeting bot asks permission
Sending a bot into a meeting for the first time requires explicit, timestamped consent in your workspace.
Designed and hosted in Europe
You choose who processes, how long we keep, and when we erase.
Hosting
Platform hosted in Europe. For organisations that require it, a version installed on your own servers, prepared on quotation and with hands-on assistance.
AI processing
The provider is chosen per agent. With Mistral, a French provider, your data never leaves the European Union. With your own keys, your contractual terms with the provider apply, not ours.
Retention
Execution retention set per workspace, 90 days by default, from 1 day to 10 years. Execution snapshots are encrypted at rest and purged when the period ends.
Secrets
API keys, connector credentials, tokens: encrypted in a vault, never shown in clear text, never written to logs.
Individual rights
Export and erasure of a user's data, self-service from their profile, to feed your records and answer within GDPR deadlines.
AI Act
A built-in questionnaire qualifies each use case and generates the matching PDF report, to attach to your compliance file.
What we do not claim. We display no certification we have not obtained. What we state above is verifiable in the product, and we show it during the technical review.
Everyone sees what concerns them. You see everything.
Workspaces
One workspace per team, per client or per project, with its agents, connectors, documents and budget. In the Nexfire cloud, execution data, the knowledge base and secrets are isolated per workspace at the database level itself (PostgreSQL row-level security), not only in code.
Roles
Owner, admin, member and viewer in each workspace; organisation-level administration above. Only admins enable paid capabilities and manage access.
Audit log
Every action by every agent and every user: who, what, when, with which data, at what cost. Viewable per workspace and across the organisation, exportable.
Observability
Running executions, success rate, duration, cost per agent and per model. Governance alerts when a threshold is exceeded.
FinOps
AI budget per workspace, consolidated tracking for the organisation, real cost per execution. Your own keys are never debited; ours are billed at the provider's list price, with no coefficient.
Policies
Autonomy level, supervised actions, allowed providers, set per workspace and applied to all of its agents.
Fifteen providers, chosen per agent, switchable without breaking anything.
- Anthropic
- OpenAI
- Mistral
- xAI
- DeepSeek
- Kimi
- Qwen
- Z.ai
- MiniMax
- Perplexity
- Groq
- Together
- OpenRouter
- Ollama, on-premise
Every model is rated for its reliability at agent work, and that rating is shown at selection time. You can enforce a provider per workspace, or let each agent use the best fit. Your own API keys are accepted and encrypted; with them, AI billing goes straight through your contract.
For visual and audio deliverables, image, video, voice and music models are chosen the same way, and are only enabled by an admin.
You get in within a minute. You leave with your agents and your data.
Who it is for
- Nexfire cloud
- Any company, from 0 to 99 EUR a month, plus the Enterprise offer
- On your servers
- Organisations with full sovereignty requirements
Hosting
- Nexfire cloud
- Europe
- On your servers
- Your infrastructure
Models
- Nexfire cloud
- Fifteen providers, your keys or ours
- On your servers
- Same, plus local models
Setup
- Nexfire cloud
- One minute, self-service
- On your servers
- Prepared on quotation, assisted installation
Support
- Nexfire cloud
- Help centre, support; named contact on the Enterprise offer
- On your servers
- Named contact
| Criterion | Nexfire cloud | On your servers |
|---|---|---|
| Who it is for | Any company, from 0 to 99 EUR a month, plus the Enterprise offer | Organisations with full sovereignty requirements |
| Hosting | Europe | Your infrastructure |
| Models | Fifteen providers, your keys or ours | Same, plus local models |
| Setup | One minute, self-service | Prepared on quotation, assisted installation |
| Support | Help centre, support; named contact on the Enterprise offer | Named contact |
Versions and tests before production
Every agent is versioned. Every workflow is compiled into a frozen version and checked before launch. You know what is running, and you can roll back.
Import, export, cloning
An agent exports to a single file, clones from one workspace to another, re-imports elsewhere. Your instructions and settings belong to you.
API and tokens
Everything the interface does can be done through the API, with scoped tokens. Nexfire fits into your tools; it does not replace them.
Explicit quotas
Agents, automations, documents, meeting hours, webhook throughput: every plan shows its limits, and the interface warns before they are reached.
Three deliberate choices, because they protect your company.
No computer control
An agent does not click inside your software on your behalf. It goes through authorised connections whose permissions are known and revocable. Less spectacular, and auditable.
No autonomy without a brake
There is no mode in which an agent sends, pays or deletes without a human having seen it. That mode will not be added.
No certification displayed without holding it
We show what we can prove. When a certification is obtained, it will appear here, with its date.
Questions from decision-makers
Not by Nexfire. For AI processing, the terms of the provider you choose apply; with your own keys, the terms of your own contract apply. We help you choose.
Exactly the connectors, documents and databases you assigned to it, with the permissions granted at authorisation time. Nothing else. You revoke an access in one click.
Every sensitive action was waiting for your approval; you reject it and the agent starts again with your feedback. Everything is in the audit log, with the data used.
Yes: AI budget per workspace, monthly cap on payment mandates, quotas per plan, visible cost per execution. With your own keys, AI is billed by your provider and Nexfire debits nothing; with ours, the provider's price applies with no coefficient.
Through workspaces and roles: owner, admin, member, viewer. Invitations are sent by email, and administrative actions are logged.
Yes, on quotation: we prepare a version of the platform installed on your servers, and we assist you in setting it up and running it.
You export your agents, retrieve your deliverables, and request erasure. No data is retained beyond the period you set.
Your use cases may be. The built-in questionnaire qualifies each of them and produces the report to attach to your file. We go through it together during the technical review.
Yes. A four-page sheet covers access, human control, data and deployment, for your supplier questionnaires and your tenders.
Forty-five minutes with an engineer, on your constraints, not on our slides.
Bring your use case, your tools and your security requirements. We show the product as it is, including what it does not do.
Download the security sheet (PDF, 55 KB)
Hosted in Europe · Sensitive actions always approved · No markup on AI