TECHNOLOGY / THE CORE DOSSIER
Simple to use.
Engineered in depth.
Your AI is the entry point. Behind it, an architecture separates access, acquisition, understanding, memory and execution. Here you will find the technical choices, the problems they solve and the actual flow of information.
01 / ACCESSIBILITY AS A DESIGN PRINCIPLE
A familiar entry point.
The Core manages the connections.
An initial setup brings the model and sources together in your OmniAgents environment. You connect the accounts you use, authorise their services and enter your website’s URL. The Core then reuses valid connections in your daily work.
YOUR INTERFACEChatGPT
Claude
Compatible client
Claude
Compatible client
Connection to the model→MCP
Authorised context and tools
Authorised context and tools
OMNIAGENTSAI×9Memory · Roles · Policy
Tools · Receipts
Tools · Receipts
OAuth / connector→Accounts and scopes
Protected tokens
Protected tokens
YOUR SOURCESGmail · Contacts · Calendar
Drive · Google Photos
Meta · Instagram DM
iCloud Mail
Your website · Web research (Business)
Drive · Google Photos
Meta · Instagram DM
iCloud Mail
Your website · Web research (Business)
You configure your environment.
For Google, a single consent covers the services listed in Google’s window. The connection uses PKCE, can include permissions already granted and reuses the connection for the same account. Credentials remain encrypted on your computer
Connections remain available.
You can disable a service or revoke the account connection. The Core keeps connection status, permissions and source progress separate. It asks for another step when needed.
The two authorisation boundaries, in detail+
MCP → Core checks the client, resource, user, organisation, session, expiry and permitted tools. Core → provider checks the account, scopes and individual capability. MCP HTTP access with OAuth has a read scope of archive.read; write operations follow the runtime’s enabled paths. Integrated Chat with a direct API is an additional entry point in the Business experience.
mcp-oauth-core-auth · universal-connector-oauth · universal-connector-service · native-connector-oauthYOUR ENVIRONMENT, ALSO FROM YOUR PHONE
The Core on your computer, Mac or Windows.
Your AI, also in your pocket.
Install OmniAgents on your computer, Mac or Windows, and finish setting up the model and sources. With the remote connection active (omniagents.link), you can also use your AI from your smartphone: ChatGPT on a Pro or Business plan, or Claude if your plan allows remote connectors. The Core on your computer provides memory, context and tools; sends and changes are confirmed in the app on your computer. The OmniAgents mobile client is coming soon.
Computer on and connected
through the desktop
The dashboard can remain closed. To continue from your phone, the computer running OmniAgents must remain on, online and reachable; when it disconnects, mobile access resumes when it comes back online.
02 / FROM SOURCE TO KNOWLEDGE
Importing is only
the first step.
Content must be identified, read and understood without losing its origin, revisions and permissions. The system keeps these responsibilities separate.
- 01DetectionSource · account · revision
- 02AcquisitionOriginal · attachments · occurrences
- 03ExtractionNative text · OCR · structure
- 04ControlCoverage · fingerprint · hash
- 05UnderstandingWork units · model · roles
- 06ConsolidationKnowledge · memory · provenance
OCR with a memory of the work already done.
Digital PDFs, scans and hybrid documents follow a process that preserves pages, extractions and coverage. Reusable results also depend on the parser version and options: a technical change may require a new derivation.
Deduplication without losing the history.
The same attachment received by email and found in a folder can share the extraction work. The two occurrences retain their own source, context and authorisation. The content cache does not replace access control.
ORIGINAL CONTENTDocumentHash + origin + revision
NATIVE TEXT
Readable content and position references.
OCR
Text extracted from images of the pages.
STRUCTURE AND COVERAGE
Processed units, missing data and visual requirements.
Traceable
derivation
derivation
Understanding remains a subsequent step.
Formats, limits and resumption+
Units follow the format: PDF pages, slides, sheets or parts of a document. The process applies explicit limits to size, pages/sheets, archive expansion and duration. The cache is isolated by tenant and organisation; locks coordinate identical extractions. Sources have separate checkpoints, and unchanged revisions can avoid further model work.
source-delta-supervisor · connector-source-ingestion · canonical-document-learning · canonical-source-learning03 / CONTINUITY WITHOUT MIXING THE LAYERS
The chat has a memory.
The Core builds knowledge.
Journal
Messages with author, session, encryption, provenance and request identity. Content cannot assign itself a different author.
Promotion
Useful information is selected and linked to sources. An AI-generated statement does not automatically become a fact.
Knowledge & Deep Memory
Facts, relationships and persistent context with origin, revision and status. Retrieval respects role, validity and permissions.
Knowledge of the application.
Product/App Knowledge gives agents context about OmniAgents functions and workflows. The model and Eric also become a guide within the app: you can ask how to start, where to find a function and how to proceed.
Knowledge of your environment.
Documents, facts, relationships and chat memory, by contrast, concern your information assets. Context composition distinguishes the application manual, the agent’s role and the user’s authorised data.
ObservedCanonical factInferenceHypothesis→Active / superseded / conflict
The data remains yours.
Conversation memory, documents and the information assets of your environment remain separate from ownership of the software.
04 / AI BECOMES AN OPERATIONAL ASSISTANT
Research. Prepare.
Act, with your control.
The model can propose work; the Core takes it towards tools and services through a persistent process. Sources, permissions and results remain linked.
- 01RequestPerson · event · schedule
- 02PreparationContext · roles · versioned draft
- 03AdmissionTool · scope · policy · confirmation
- 04ExecutionLease · intent · adapter
- 05VerificationReceipt · read-back · commit
- 06ReturnConversation · status · references
Find and compare sources.
Research through an adapter or the model’s capabilities, followed by reading the selected pages. Research, opening URLs and building Knowledge each have their own steps.
Research → sources → reading → dossier with referencesFrom content to a post.
In preparation: publishing images on Instagram requires an additional Meta permission that is not currently active. When it becomes available, nothing will be published without your confirmation.
In preparation → additional Meta permission → confirmationIt starts with your URL.
Enter your website’s address to import its available public content. The Core links it to sources and memory, so your AI can find and use information about your work.
Website URL → import → content → contextMessages have context too.
Search and read the available conversations of the authorised Instagram account. Prepare, send or reply with text messages through the connector’s permissions.
Account → conversation → context → authorised replyPermission confirmedContinues in the authorised tool
Your approval is neededStops at the content to be confirmed
Uncertain external outcomeReconciles before repeating the action
Why a persistent queue matters+
A task has identity, correlation, checkpoints and execution ownership. The scheduler assigns work; the worker maintains the lease; the gateway admits external effects. The task’s validity is checked again in the commit transaction. A late result cannot become canonical state after execution authority has been lost. Effects that have already occurred at the provider are reconciled: revocation does not undo them retroactively.
Delegated work requires an active runtime and model transport. Task persistence provides continuity independent of the dashboard, without attributing to a chat window a process that is not running.
persistent-operational-execution · persistent-execution-workflow · worker-scheduler · control-plane/registryResearch, reading and publishing have distinct capabilities+
The connector’s scope determines the actual operations. Instagram requires a publishing scope separate from those for reading and messages. Opening a public page does not enable writing to its CMS; WhatsApp receives only messages that arrive after connection. These boundaries guide tool activation and consent requests.
05 / REQUIREMENTS, TESTS AND INTEGRITY
Behaviour is described.
Then put to the test.
Checks also cover negative cases: a changed source, a revoked permission, an expired approval, a lost lease, a timeout and an unverifiable result.
Tracked atomic identities
Each entry preserves references and status. The register distinguishes requirements, evidence and deferred items.
Reconciliation of 4 October 2026Backend and UI contracts
Backend and UI checks on the four changes in the technical batch.
Receipt of 4 October 2026Memory, gateway and scheduler
Ten suites include tests of approvals, model results and concurrency.
Execution of 19 September 2026What is checked beyond the positive result+
Memory: source and revision, conflicts, invalidation and the distinction between facts and hypotheses.
Approvals: exact content, affected resources, expiry and payload changes.
Model Gateway: parsing, outcomes, credentials, lease loss and takeover.
Scheduler: concurrency, resumption and fairness between queues and sources.
Integrity: manifests and hashes define the changes. The 4 October batch checked 1.449 files: four changed, 1.445 unchanged and zero unexpected differences.
The register counts requirements; the suites count tests within their respective scopes and dates. They are different measurements, retained with their evidence.
06 / QUESTIONS TO ASK AN AI CORE
The technical questions.
Concrete answers.
How does the initial setup work?
You configure a single OmniAgents environment: connect the model, then the accounts and sources you want to use, such as Google, Meta, WhatsApp and your website. Google brings its services together under one consent; Meta and WhatsApp retain their own connections. The Core stores and reuses valid connections during work; you return to settings to add a source or renew consent.
How does the model connect to my digital world?
MCP is the connection that lets a compatible model communicate with OmniAgents. It is part of the AI connection. OAuth instead handles authorisation for external accounts: source credentials remain protected in your environment.
How do I continue from my phone?
On your phone, you use your AI connected to OmniAgents through the remote connection (ChatGPT Pro or Business, or Claude with a remote connector). The environment stays on your computer, which must remain on, online and running OmniAgents; the dashboard can stay closed. Sends and changes must be confirmed in the app on your computer. The mobile client is coming soon.
What happens when an authorisation expires?
The Core rechecks the connection’s scopes, status and revision. Google can renew the token through a refresh token; when required by the provider or grant, reconnection is requested. Disabling a service and revoking the entire account have different effects.
How does it handle a PDF with text and scanned pages?
The extraction process distinguishes native text from OCR. In hybrid PDFs, it preserves contributions from the same page and records coverage, hashes and the extractor revision. Text coverage and understanding of visual elements remain separate measurements.
Is the same attachment processed more than once?
The cache reuses a derivation compatible with the content, tenant, organisation, classification and parser/options version. Different occurrences of the file each retain their origin and authorisation. A lock coordinates concurrent identical work.
Is a newly acquired source already Knowledge?
Acquisition makes the content and its provenance persistent. Extraction and OCR produce a technical derivation. Learning builds Knowledge only through the model process and its coverage and persistence checks. These are separate, traceable stages.
Can the model guide me in using OmniAgents?
Product/App Knowledge describes the application’s sections, functions and workflows. It provides application context for agents: the model and Eric can explain functions and guide you through your work. This guidance remains separate from the private facts of your environment and does not grant permissions.
Does memory also keep the AI’s mistakes?
The journal preserves the conversation; promotion into knowledge is selective. An assistant’s statement does not automatically become a fact. Memories distinguish observations, canonical facts, inferences and hypotheses, with validity and conflict states.
Can a slow response end up in the wrong chat?
Request, conversation, organisation and session are correlated. Before saving and returning the result, the channel and execution ownership are checked again. A late result cannot ignore cancellation or lease loss in the paths protected by the Core.
How are parallel activities managed?
The scheduler, persistent jobs and leases separate work selection, execution and the right to write. Configurable limits contain concurrency and resource use. Checkpoints let incomplete steps resume without repeating completed ones.
What prevents a document from giving orders to agents?
Acquired content remains untrusted external data. Permissions come from identity, policies and registered tools. The text of a page, PDF or model response cannot grant privileges.
Can the AI really publish or work on the web?
The Core includes distinct paths for web research, reading pages, drafts and actions through adapters. Active features include Instagram direct messages and importing your website via its URL. Instagram publishing is in preparation. Each action uses authorised accounts, scopes and tools, with confirmation when required by the policy.
Is a timed-out publication repeated?
An ambiguous outcome enters the reconciliation process. The provider’s receipt, reading the result back and the local commit are separate steps. For a non-idempotent action, the Core avoids treating silence as a failure to retry blindly.
What remains if I change models or discontinue Business?
The Core preserves the memory, sources and knowledge of your environment. Changing to a compatible model does not transfer the provider’s private history. The end of a Business licence is separate from export, retention and deletion: the data remains the customer’s.
FOR THOSE WHO WANT TO BUILD THE NEXT STEP
Assess the technology.
Talk about the project.
Designed and developed by Giampaolo Rao, independent developer.
OmniAgents is seeking partners for development, partnerships and funding. The discussion starts with architectural choices, use cases and the product’s direction.