EzyAgents documentation

EzyAgents is a platform of AI agents that automate your business operations: support, sales, finance, HR, marketing and more. An orchestrator, the Manager, plans the work and distributes it across ten specialized agents, each equipped with its own tools and connectors.

You choose the AI provider and the models that power your agents: Claude (Anthropic), ChatGPT (OpenAI), Mistral, Grok (xAI), AWS Bedrock EU or your own keys. EzyAgents is not tied to any model vendor.

Your agents do more than answer: they run missions over time, schedule routines, send emails, prospect on LinkedIn, make calls, launch ad campaigns and produce real files (Excel, PowerPoint, Word, PDF).

Every sensitive action goes through human oversight (HITL) that you configure, and your data is hosted and encrypted in Europe.

Deploy your first agent in under 10 minutes:

• Create an account (14-day free trial, no credit card).
• Choose a business agent (Support, Sales, Finance, HR…).
• Connect your tools in one click (Gmail, Outlook, Slack, HubSpot…).
• Define your approval rules (HITL), then go live.

A workspace ready to use in 20 minutes

The path we recommend for a pilot team, in the order where each step makes the next one useful:

• Knowledge base → Add a document → Link: paste a Notion page, a Confluence space, a GitHub repository, a Drive folder, SharePoint, Box or Dropbox. The source is recognized and stays in sync.
• Connectors → Slack: the bot replies in a thread, requests approvals and asks its questions where the team works.
• My agents → activate a business agent; its tools, connectors and guardrails are ready; the Conversation tab is a real sandbox.
• Manager (Pro plan) → a request in natural language ("chase this month's unpaid invoices, with approval before each send"): the Manager plans, delegates, and every sensitive action goes through Approvals.
• On your phone: add ezyagents.fr to your home screen and turn on notifications in My account to approve from your phone.
• In Chrome: install the extension from the Chrome Web Store and click "Sign in with Ezyagents"; your agents are in the side panel, on any web page.
• Settings → General: workspace instructions (tone, signature, rules) apply to every agent; Billing & usage: monthly budget and per-run cap.

EzyAgents does not lock you into one AI provider: you choose the one that powers your agents, and the models they use. Anthropic (Claude), OpenAI (ChatGPT), Mistral, xAI (Grok) and AWS Bedrock EU are available with no API key to provide; you can also plug in your own keys. With no particular setup, your agents start on Claude, then Auto mode picks the right model for each task from your chosen provider.

Available providers and models

From the Pro plan, you pick the primary provider in Settings → Models & AI, along with a fallback order: if it goes down, your agents automatically switch to the next provider in the chain. Requests go through EzyAgents' managed infrastructure and stay within your plan's included AI usage. On the Trial and Solo plans, your agents run on Claude (Anthropic).

Automatic selection (Auto)

By default, Auto mode picks the model for each task without ever switching provider: a fast tier for triage and short follow-ups, a standard tier for most of the work, a powerful tier for the heaviest tasks. The most advanced models (Claude Fable 5.1, GPT-6 Astra) are never picked automatically: their cost makes them an explicit choice.

From the Pro plan, you can also pin a specific model to an agent, from the conversation's model selector. At 80% of the month's AI usage, the selector recommends switching back to Auto. On Enterprise, routing can be fine-tuned per agent, with a dedicated model for each tier.

Your models, your keys (BYOK, from the Pro plan)

From the Pro plan (outside the trial period), you can use your own providers and models by bringing your API keys, including the latest models from each vendor. Two modes are available:

• Hosted by EzyAgents (from Pro): simply enter your Anthropic, OpenAI, Mistral or xAI key; we store it in isolation in our proxy and route your requests to the chosen model (e.g. GPT-6 Astra, GPT-6.1 Sol, Mistral Large 3, Grok 4.7). Nothing to install.
• Self-hosted (from Pro): point EzyAgents to your own compatible gateway (LiteLLM, Azure OpenAI, Bedrock, Ollama…) with a URL and a key you control. No model call goes through our servers.

Supported providers: Anthropic (Claude), OpenAI, Mistral, xAI, AWS Bedrock, Azure OpenAI and Ollama. A fallback chain can be configured to switch providers automatically. With your own keys, token billing happens directly with your provider.

No data is used to train models. To keep inference within the EEA, choose Mistral or AWS Bedrock EU (from the Pro plan, using our keys), or connect Bedrock in your own AWS account (BYOK).

Each plan includes a monthly volume of conversations. A "conversation" is a mission entrusted to an agent (a series of exchanges and actions until it is closed).

Included AI usage

Every plan includes fair-use AI usage, visible in Settings → Billing & usage. An alert is sent at 80%; at 100%, agents pause until the next month. You can set a lower monthly budget and a per-run spending cap: a run that reaches it stops and keeps its work. With your own keys (BYOK, hosted or self-hosted), tokens are billed by your provider and the plan's AI usage does not apply; the conversation quota stays that of your plan, except in self-hosted mode.

Voice, images and videos included each month

Beyond what is included, these uses consume credits: 5 credits per call minute, 1 per dictation minute, 4 per image, 1 per 1,000 characters read aloud, 25 per AI video clip. With your own accounts (Twilio, Higgsfield, a Gemini or OpenAI key for images), they consume no credits.

Credits

Credits also pay for sales enrichment by the Sales agent: 1 credit per 10 prospects found, 1 per verified work email, 3 per personal email, 10 per mobile number, nothing when there is no result. They are sold in packs that never expire:

The Manager is your point of contact and the team's conductor, available on every plan from the trial onwards. On Solo and during the trial it keeps the shared base (email, calendar, messaging, files, search) and hands business work to your active agent. Describe the expected outcome; it builds a work plan with success criteria, asks the right questions, then delegates each step to the competent business agent, up to 50 simultaneous tasks per organisation.

Before wrapping up, it checks every success criterion, produces a summary and submits any remaining sensitive actions for your approval. Delegated tasks appear in the "Background tasks" panel, where you can stop them.

Project teams

Since September 2026 a project can bring several agents together: from two agents on, the Manager leads the project, spreads the work between them, tracks progress and returns one consolidated file. Members hand over to each other without losing the project context. Every agent now has an avatar (shape and colour chosen in its profile); the Manager is recognisable by its white drop.

Each agent ships with its tools, connectors and guardrails. They all share a common base: read and send emails, post in Slack and Teams, create calendar events, search Drive, query the knowledge base and the web, produce files and schedule routines.

• Customer Support: email triage, tickets, knowledge base search, approved replies, escalation, calls.
• Sales SDR: LinkedIn and SIRENE prospecting, enrichment, email and LinkedIn sequences, prospecting pipeline, calls, e-signature.
• Finance: invoice follow-ups, reconciliation, monthly accounting report, SIRENE search, calls.
• HR & Office: onboarding, employee questions, leave and scheduled reminders, e-signature, calls.
• Marketing: brand content, email campaigns, competitive intelligence, ad campaigns, short videos.
• Communications: LinkedIn, X and Instagram posts, AI visuals, campaign plans, short videos, approved publishing.
• Tech / Engineering: security audits of code and GitHub, debugging in an isolated terminal, approved pull requests, runbooks and incidents.
• Analyst (Excel & Slides): real Excel and PowerPoint files, data analysis, reporting.
• Legal / Compliance: contract review, NDAs and T&Cs, GDPR register, deadlines, e-signature, SIRENE search.
• Procurement / Operations: supplier sourcing, quotes, IBAN anti-fraud checks, renewals, e-signature, calls, SIRENE search.

On Trial and Solo, the chosen agent uses the tools of its own role. From the Pro plan, every agent can be equipped with the tools of every role.

Give an agent a goal that runs over time ("book 10 qualified meetings this month"): it turns it into a mission with an objective, 3 to 12 steps, success criteria, indicators and a cadence. It moves the mission forward day after day and flags what blocks it: a missing connector, a decision to make or information to provide.

Missions are tracked in each agent's Workspace tab, where you can pause, resume or archive them; pausing a mission suspends its routines. Available on every plan.

Scheduled tasks

Ask for a routine in a conversation ("every Monday, review unpaid invoices") or create it from the Scheduled tasks menu: daily, on business days, weekly, hourly or on demand. Each task can be run right away, paused or edited, by you or by the agent.

Workflows

The Workflows menu chains several steps and several agents, in simple mode or in advanced mode with parallel branches.

The "+" menu in the message box lets you attach files (up to 5 per message: PDF, Excel, Word, PowerPoint, CSV, text, images), start a deep research or open a temporary chat.

• Deep research: the agent explores the web and your sources in parallel, then delivers a sourced report.
• Real files: Excel, PowerPoint, Word, PDF or CSV produced in an isolated execution environment, to download or open in Google Drive.
• Plan and questions: on a long task, the agent shows its plan (steps, success criteria, deliverables) and can ask you a question before going on.
• Projects: group conversations, sources (Drive or SharePoint links, web pages, notes, files, saved replies) and custom instructions, with project-only memory if needed.
• Memory: your agents remember your preferences from one conversation to the next, and what they learn about your company (contacts, projects, facts) is shared with the whole team of agents. The "Sources" chip under a reply shows the memories used; My account → My memory lets you correct or forget them.
• Published pages: an agent can publish a report or a page at a stable address, restricted to the workspace or public if the administrator allows it (Documents → Published pages).
• Images: image generation by Google (Nano Banana) or OpenAI, to turn on in Settings → Models & AI (off by default, as these providers are outside the EU).
• Voice: dictation with the microphone in the message box, and a "Listen to the reply" button under each reply.
• Temporary chats and sharing: a temporary chat does not use memory and is erased after 30 days; public link sharing comes on top of sharing with workspace members. Both options are off by default and can be turned on in Settings → General.

Following your agents' work

The Activity button at the top of the screen shows what is running now, pending sends, what is scheduled and what finished in the last 24 hours; every run can be stopped. The "To do" section on the home page and in the Manager gathers everything waiting for you: approvals, agent questions, missions, pending LinkedIn sends and failed sends.

The Support, Sales, Finance, HR and Procurement agents make outbound calls: the agent steers the conversation toward a precise goal, with a natural voice in French or English.

• By phone: with your own Twilio account (no credits) or with telephony provided by EzyAgents (5 credits per minute beyond the included minutes).
• Web call: the person receives a link and talks to the agent from their browser, with no number or operator. The link stays valid for 7 days.
• Guardrails: the agent announces the cost before calling and respects legal calling hours (business days, 10am to 1pm and 2pm to 8pm); calls follow your approval rules.
• Tracking: recent calls and the web call links sent appear in the agent's workspace.

The Sales agent builds a duplicate-free prospecting pipeline, fed by LinkedIn, the SIRENE database, your inbound contacts and your CRM. Its workspace shows prospects added and contacted, replies and meetings, as well as send waves, which can be cancelled until they go out.

A campaign's messages are edited from the mission card: subject, follow-up emails (D0, D+3, D+7, D+14), LinkedIn invitation note, message after connecting, signature and unsubscribe line, with the variables {{firstName}}, {{company}} and {{jobTitle}}.

Three ways to send on LinkedIn

• Account connected through Unipile (paid option): immediate sends, even with your computer off.
• From your Chrome (free): the extension sends approved messages from your own LinkedIn account, at low volume: 10 invitations and 20 messages per day by default, adjustable up to 20 and 20, with a gradual ramp-up over the first three weeks. The agent decides on its own between a message (existing connection) and an invitation with a note.
• Manual: the agent drafts the message and you send it yourself.

Every LinkedIn send from Chrome goes through human approval. Routines send on business days from 8am to 7pm; a request made in a conversation goes out as soon as Chrome is open. Until a send has gone out, you can cancel it.

Enrichment (verified email, mobile number) consumes credits; company search in the SIRENE database is free.

The Ads tab of the Marketing workspace creates and tracks your Meta Ads, Google Ads and OpenAI Ads campaigns, with your own ad accounts (Settings → Integrations → Ads & creatives). Campaigns are always created paused, and launching them requires approval by default. Agents also produce ad visuals and videos (Nano Banana, Higgsfield), using credits or your own keys.

Short videos from your website

Give your website address: the agent turns it into a series of short videos.

• "Pro" videos: screenshots of your site, animated text, voice-over and subtitles, 10 to 45 seconds long, in 9:16, 1:1 or 16:9. Only the voice-over consumes usage.
• "Viral" videos: AI-generated clips (Higgsfield), 25 credits per clip, or no credits with your own Higgsfield key.

The Short videos panel lets you touch up each video, then publish it to your LinkedIn page, as an Instagram Reel or on TikTok.

By default, no critical action is executed without your approval. You define, action by action, what an agent can do autonomously and what requires explicit approval.

• Approval queue: approve, edit or reject each action, based on per-agent rules (conditions, approvers, time limit).
• Bulk sends: beyond 25 sends at once (emails, LinkedIn messages, calls, posts), approval is required; the threshold is adjustable.
• Tool Guard: Read-only, Approval first (default) or Custom presets.
• Questions: when information is missing, the agent asks you instead of guessing.
• Approve wherever you are: in the app, in Slack, in the Chrome extension or from a notification on your phone.
• Traceability: each agent lists what it sent, published or called; version history lets you restore its configuration; the audit log is timestamped and exportable.

Your documents, procedures and FAQs feed every agent reply. Search is semantic and every answer cites its sources (document and page), so it stays verifiable.

• PDF, DOCX, XLSX, CSV, TXT, Markdown and HTML ingestion, file by file or whole folders at once.
• Synced sources: Google Drive, SharePoint / OneDrive, Notion, Confluence, GitHub, Slack channels, Box, Dropbox. A link is enough, the source is recognized.
• Isolated semantic search per workspace; each knowledge base can be restricted to certain agents.
• Company knowledge: a single search queries the knowledge base, the CRM (HubSpot, Salesforce), Drive, tickets and your MCP connectors.
• Sourced, traceable answers (clickable [n]).

Agents act inside your existing tools: email (Gmail, Outlook), messaging (Slack, Teams), CRM (HubSpot, Salesforce), support (Zendesk, Intercom, Freshdesk, Front, Crisp, Jira), finance et ERP (Odoo, Pennylane, Sage, Xero, QuickBooks), databases (PostgreSQL, MySQL, SQL Server, hence the database of your ERP or business software), HR (BambooHR, Personio, Lucca, Greenhouse, Lever, Recruitee, DocuSign, Yousign), marketing and advertising (Mailchimp, Brevo, Buffer, WordPress, Meta Ads, Google Ads, OpenAI Ads), prospecting (LinkedIn, Apollo, Kaspr, SIRENE database, phone calls) and documents (Google Drive, SharePoint, OneDrive, Notion, Confluence, GitHub, Box, Dropbox).

You can also plug in your own tools via the API, webhooks or an MCP server, and act on any web page with the Chrome extension.

Packs (Marketplace → Packs) bundle skills, connectors and instructions, installable on one agent or across the workspace; for each role, the administrator decides whether they are available, installed by default or blocked.

The Chrome extension puts your agents one click away on any web page, in a side panel. It installs in one click from the Chrome Web Store and links to your workspace without copy-pasting any key.

What it does

• Side panel: chat with your agents (Manager, support, sales, finance…) without leaving the page, streamed replies, agent picker, conversation resume.
• Attach the page: in one click, the agent receives the title, URL and visible text of the page to summarize it, analyze it or extract the key points.
• Context menu: right-click on selected text ("Ask Ezyagents") or on the page ("Analyze this page with Ezyagents"); the answer arrives as a notification.
• Live approvals and questions: you allow or deny a sensitive action, and answer the agent's questions, from the panel.
• Browser piloting (optional, off by default): your agents read the page, click and fill in fields in a dedicated tab, never in the one you are looking at. In "Ask before acting" mode, every action waits for your approval; an approved site no longer asks again in the same conversation, and the "Allowed sites" list limits the domains that can be reached. Password fields are always refused.
• LinkedIn sending from your Chrome (optional, off by default): invitations and messages approved in Ezyagents go out from your own LinkedIn account, at low volume (see Prospecting & LinkedIn).

Account and plan

The extension does nothing on its own: it needs an Ezyagents workspace. On the first click on "Sign in with Ezyagents", you are invited to sign in or create an account (14-day free trial, no credit card); pairing completes automatically and your first agent is already active.

It uses exactly your plan's agents, capabilities and quotas: one business agent and 50 conversations during the trial, 300 conversations per month on Solo, all ten agents and the Manager on Pro. A conversation opened from the extension counts as a workspace conversation, and sensitive actions follow the same approval rules.

Privacy: no telemetry, page content is only read when you ask and is only sent to your own workspace's API. For organizations that lock down Chrome, the ID to allow in the ExtensionInstallAllowlist policy is edloicocglghlnikiddemiajomdpiedp.

Your agents join you in the tools your team already uses:

• Slack : « Ajouter à Slack » depuis Connecteurs, un clic et le Manager répond en message privé et dans les canaux où vous le mentionnez ; un fil correspond à une conversation, et les approbations comme les questions se règlent par boutons.
• WhatsApp : Connecteurs → WhatsApp → « Relier mon WhatsApp », scannez le QR code ou envoyez le code d'appairage, et votre Manager répond dans votre WhatsApp ; une action sensible se valide par un OUI. C'est aussi là qu'il vous écrit de lui-même quand il a une suite à proposer (réglable dans Paramètres → Général).
• Microsoft Teams: your agents reply in Teams; approvals happen in the app, in the extension or from your phone.
• Word, Excel and PowerPoint: the add-in puts your agents in the open file. In Word they fix a sentence, rewrite a passage or redo a heading without touching the rest; in Excel they write values and formulas, add a sheet, a chart or a pivot table; in PowerPoint they add slides on your theme and check overlaps, overflows and text too small. Every edit is proposed as a card to approve ("Ask before each edit", the default) or applied directly. It is installed through Microsoft 365 centralised deployment or by loading the manifest https://ezyagents.fr/office/manifest.xml, then linked to your workspace with a short code.
• Google Docs, Sheets and Slides: a sidebar puts your agents in the open file. In Docs they fix a sentence, rewrite a passage or redo a heading without touching the rest; in Sheets they write values and formulas, add a tab, a chart or a pivot table; in Slides they add slides on your theme and check overlaps, overflows and text too small. Every edit is proposed as a card to approve ("Ask before each edit", the default) or applied directly ("Apply without asking"). The sidebar is installed as a private add-on for your Google Workspace domain, on request.
• Google Docs, Sheets and Slides: with the Google Drive connector, your agents create a document, a spreadsheet or a presentation, and read and edit the ones they created or that you opened with Ezyagents, with no sidebar to open.
• Word, Excel and PowerPoint on OneDrive and SharePoint: with the SharePoint / OneDrive connector, your agents find a file, read it, write Excel ranges in place, edit a document (new version, same link) and drop the files they produce.
• Phone: add ezyagents.fr to your home screen, with push notifications for approvals.
• API and webhooks: a public API to start conversations, scoped API keys (per-agent restriction, IP allowlist, expiry), signed outgoing webhooks (HMAC-SHA256, X-Ezyagents-Signature header) and an OpenAPI specification.

AES-256 encryption at rest and TLS 1.3 in transit, secrets isolated per organization, data partitioned per tenant. GDPR compliance, DPA on request, hosting in the European Union, EU inference available (Mistral, AWS Bedrock EU).

• Sign-in: email, Google or Microsoft, with two-factor authentication (TOTP app or passkey) that can be required by role.
• SAML 2.0 or OIDC SSO with domain discovery and account creation on first sign-in, SCIM 2.0 provisioning (Enterprise plan).
• Members: email invitations, CSV import, domain verification with automatic joining, revocable active sessions, password and session policy, IP allowlist.
• Settings → General: instructions shared by every agent, brand, workspace memory, conversation retention period, temporary chats and public sharing.
• Analytics: activity and costs per workspace member.

Enterprise customers can deploy on-premise or in their own VPC. See the Security page for details.

Four plans, from trial to enterprise. 14-day free trial with no card, then Solo at €29/month (€24/month billed yearly), Pro at €99/month (€79/month billed yearly) and Enterprise on quote. No commitment, cancel anytime. Prices exclude VAT.

Included AI usage, call minutes, images, videos and credit packs are detailed in the Quotas & credits section.

In short: Trial and Solo run on Claude with automatic routing. From Pro, you choose your AI provider (Claude, ChatGPT, Mistral, Grok or Bedrock EU), your models and your fallback chain without managing any API key, with optional BYOK in both modes (your key hosted by EzyAgents, or your own gateway). Enterprise adds fine-grained per-agent routing, SSO/SAML and private deployment. See the next section.

The Enterprise plan lets you choose where the platform and model inference run, based on your sovereignty requirements. Four modes are available and can be combined (for example: on-premise + Bedrock in your VPC).

What ships for deployment

• 3-step on-premise install via Docker Compose (automated setup script: generated secrets, Postgres/Redis/API/Worker/Dashboard, TLS reverse proxy).
• Helm chart for Kubernetes, with ready-to-use profiles (on-premise and AWS Bedrock EU values).
• Terraform modules to provision AWS Bedrock in Europe (production example included).
• Anthropic-compatible LiteLLM proxy: every model request goes through a gateway you control, and no data leaves your infrastructure.
• Copy-paste-ready OpenAI and Mistral BYOK integration templates.

In the app, everything is set from Settings → Models & AI: managed provider and BYOK, hosted or self-hosted, from Pro; private deployment on Enterprise. Once the provider is configured, its models become selectable on every agent. Our team supports the rollout (DPA, architecture review, dedicated Customer Success).

Otto is the EzyAgents tool gateway for third-party agents: Claude Code, Cursor, Codex, Grok, Claude, OpenAI or xAI agents, in-house frameworks. Your agent discovers, inspects and runs the same tools as the business agents, inside your own systems (email, CRM, tickets, HR, finance, files, Slack, Teams) and on platform tools (web search, SIRENE, enrichment, voice, images, documents), with one key and one credit balance.

Create a key for your agents

• Settings → API keys → New key → "For my agents (Otto)": the key carries the tools:read and tools:run scopes, a monthly credit cap (200 by default) and, optionally, a tool allowlist or a linked agent.
• A key linked to a business agent inherits its tools, approval rules and policies. A key without a linked agent goes through Otto: every outside-facing write waits for human approval.
• During the trial, the first Otto key adds 50 welcome credits.

Four ways to connect

Examples per agent

Data & third-party APIs

Beyond the workspace connectors, the catalogue exposes third-party APIs declared once for the whole platform: about fifteen public sources with no key (Open-Meteo, OpenStreetMap, French national address base, api.gouv.fr company register, VIES, Wikipedia, Hacker News, Open Library, Frankfurter, CoinGecko, Open Food Facts, GitHub, npm, PyPI, public holidays, Jina Reader) and about twenty services on your own key, entered in Settings → Integrations → Data & third-party APIs (Exa, Tavily, Brave, SerpApi, Firecrawl, ScrapingBee, NewsAPI, GNews, Hunter, People Data Labs, Abstract, OpenWeather, YouTube, Unsplash, Pexels, DeepL, Perplexity, Yelp, Google Places, TMDB, Alpha Vantage). Apify adds thousands of ready-made scrapers (Instagram, TikTok, Google Maps, Amazon, Trustpilot…): the 4,000 most used actors are named tools of the Otto catalogue, and the whole store stays reachable through search. When EZYAGENTS holds a service's key the call is billed in credits; on your key it is included.

Price per call

A call that fails, expires or is rejected costs nothing. Rate: 120 requests per minute per key. Every call is logged (hash-chained audit, Activity feed) and irreversible actions appear in the agent action log.

Error codes

Using the Anthropic, OpenAI or Mistral API directly gives you access to a model, not a business automation platform. EzyAgents adds orchestration, guardrails, connectors and deployment choice, while staying compatible with those same providers.

In short: EzyAgents does not replace the models, it puts them to work in your processes, with the sovereignty, security and business depth the enterprise demands, and the freedom to deploy wherever you decide.