Coordinate Company Agents
Create role-based agents with individual instructions, models, fallback models, and tool access. Run work in single-agent, pipeline, or fusion modes with delegation and review.
AI Agent
NPATI AI Agent
A company-aware AI workspace that turns a request into coordinated work across specialist agents, connected models, business knowledge, research, files, and controlled tools.
Build a Company team with its own roles, instructions, models, tools, memory, and knowledge. NPATI can plan and delegate a run, gather evidence, execute permitted work, and return one result with an audit trail.
Available capabilities depend on your deployment, connected providers, permissions, and enabled services.
Company Run
Market entry analysis
1
Strategy Agent
Defines the analysis plan
2
Research Agent
Searches and cites current sources
3
Analyst Agent
Uses company knowledge and files
4
Main Agent
Reviews and returns one answer
Scoped knowledge · permitted tools · run trace
What NPATI AI Agent Can Actually Do
These are implemented capabilities in the current codebase. Provider-dependent tools appear only when their connection and policy requirements are satisfied.
Research with Sources
Search the web, fetch public URLs, and run bounded website or competitor research. Results retain normalized findings, source URLs, citations, and persisted job state.
Use Company Memory
Retrieve and update durable Company memory while keeping neutral chat, Business, Company, workspace, and user contexts isolated.
Ground Work in Knowledge
Upload documents and private files, extract their content, index Company knowledge, and retrieve relevant authorized material for a run.
Route Across AI Providers
Discover connected cloud and local models through OmniRoute, assign models per agent, and use automatic routing, fallback, health checks, and usage telemetry.
Execute in a Sandbox
Run supported JavaScript or Python work in an isolated execution environment with permission checks, time limits, secret filtering, policy validation, and approval gates.
Understand More Than Text
Process supported images, audio, speech, OCR, and documents. Generate speech when a compatible provider is connected.
Work Beyond the Web Chat
Use the shared Agent runtime through verified Telegram flows and authenticated IMAP email triggers, with configurable access, pairing, media, and session controls.
How the Agent Is Built to Work
The model is only one part of the runtime. Context, knowledge, tools, policy, and execution are resolved around every Company run.
Your Request
A goal, question, file, or company task
Active Business or Company Context
The correct conversation, permissions, memory, and knowledge scope
NPATI Orchestrator
Plans the run and selects a model or configured Company Agents
AI Models
Connected cloud or local providers with routing and fallback
Knowledge
Company memory, documents, files, and relevant retrieval
Allowed Tools
Research, MCP reads, code, email utilities, and supported actions
Policy, Approval, and Audit
Every capability is checked; sensitive actions can pause for approval
Result with Evidence
A streamed answer, sources, saved output, and an inspectable run trace
What Makes NPATI AI Agent Different?
NPATI is designed around durable company work, not a stateless question-and-answer window.
A Team, Not One Persona
A Main Agent can choose configured specialists, delegate responsibilities, run a pipeline or fusion, and review their combined work before the final response.
Company Context That Persists
Business facts, Company memory, documents, and knowledge are durable and scoped. The agent can build on authorized context instead of starting from zero each time.
Models Are Replaceable
The workspace is not locked to one model. Routing and per-agent model choices sit behind a provider layer with health, fallback, and connection-aware availability.
Tools Are Governed
An agent does not receive unrestricted access. Effective capabilities are calculated from tenant, user, agent, connection, permission, and policy state before execution.
Research Produces Evidence
Search and bounded crawling return sources and structured findings. Longer research can continue as a persisted job instead of disappearing with a browser tab.
Work Is Inspectable
Delegated runs, tool decisions, usage, failures, cancellation, approvals, and results are recorded so a user can see what happened and retain control.
How NPATI AI Agent Works
1. Define the work
Ask in natural language and attach the files or context the task needs.
2. Resolve context
NPATI selects the active Business or Company, its permissions, memory, and knowledge.
3. Plan and delegate
The orchestrator chooses an appropriate model and, for Company runs, the specialists and run mode.
4. Use allowed tools
Agents retrieve evidence or execute enabled tools. Sensitive work pauses when approval is required.
5. Review the result
Receive the answer, sources, output, status, and run trace; authorized memory can persist for later work.
Real Work You Can Give the Agent
Each scenario maps to implemented research, knowledge, orchestration, file, or execution capabilities.
Business and market research
Compare competitors, inspect selected websites, gather current evidence, summarize positioning, products, prices, topics, contacts, and SEO signals, and keep the source trail.
Company knowledge assistant
Answer questions from private company documents and memory, retrieve the relevant passages, and keep unrelated companies and users outside the context.
Multi-specialist analysis
Give a broad goal to a configured Company team so research, strategy, analysis, and review can be delegated and combined into one response.
Document and media intake
Upload a document, image, or audio input; extract text or speech where supported; analyze it with the active business context; and retain authorized knowledge.
Data and code-assisted work
Use isolated JavaScript or Python execution for calculations, transformations, and technical analysis when the Company Agent has the required permission and approval.
Channel-based company assistance
Continue persistent Agent sessions through web chat or verified Telegram flows, and route authenticated inbound email through the shared runtime when configured.
Security and Control
Autonomy is bounded by the current user, Company, connection, and tool policy. The runtime checks access before planning and again before executing sensitive work.
- Workspace, user, Business, and Company isolation
- Per-agent model, instruction, and tool configuration
- Tool allow, deny, and require-approval policies
- Human approval for high-risk or destructive actions
- Sandboxed execution with time and resource limits
- Encrypted provider and integration credentials
- Prompt-injection, credential, and PII safeguards
- API-key, OAuth, route-scope, and rate-limit boundaries
- Audit events, usage telemetry, cancellation, and recovery
- Live capability inspector for the current installation
Action decision
Allowed
Low-risk tool is enabled for this agent and context.
Approval required
A high-risk or destructive action pauses until a user approves it.
Denied
The tool, scope, connection, or permission does not allow the action.
Frequently Asked Questions
What is NPATI AI Agent?
NPATI AI Agent is a self-hostable AI workspace for business and company work. It combines model routing, specialist Company Agents, scoped memory, company knowledge, web research, files, tools, and controlled execution in one runtime.
How is NPATI AI Agent different from a chatbot?
A chatbot mainly returns text from the current conversation. NPATI AI Agent can plan a run, delegate work to configured Company Agents, retrieve authorized company knowledge, use permitted tools, execute supported actions, and preserve a trace of the work.
What can it do today?
Verified capabilities include persistent chat, Business and Company workspaces, Company Agent teams and delegated runs, scoped memory, company knowledge retrieval, web search and bounded website research with sources, file and document processing, isolated code execution, multimodal input, Telegram, and IMAP email triggers.
Which AI models and providers are supported?
NPATI uses OmniRoute for provider discovery and routing. The pinned catalog contains 352 provider definitions across API-key, OAuth, local, search, audio, and other provider types. The models actually available to a workspace depend on the providers connected and healthy at runtime. OpenAI-compatible, Anthropic-compatible, and local provider paths are implemented.
Can it research the web?
Yes. It can search, fetch a supplied public URL, and run bounded multi-page website research through Crawl4AI. The research flow normalizes findings, records source URLs, respects robots.txt, and blocks private or unsafe network destinations.
Can it use files and company knowledge?
Yes. Files and extracted documents can be stored privately and linked to a Company knowledge base. Authorized agents can retrieve relevant knowledge and Company memory without crossing workspace, user, Business, or Company boundaries.
Can it execute code or external tools?
Authorized Company Agents can use configured tools including research, files, safe MCP reads, memory, email utilities, and isolated code execution. Tool availability is evaluated at runtime, and code execution is sandboxed with permission, timeout, environment, and approval controls.
Can it act autonomously?
It can plan and complete delegated or background Company runs and use low-risk tools allowed by policy. High-risk or destructive actions can be denied or paused for explicit approval. Recurring scheduling and fully autonomous browser operation are not presented as complete capabilities.
Which communication integrations are verified?
Web chat is the primary interface. Telegram runs through the shared Agent runtime, and authenticated IMAP email can trigger work. The codebase also contains configurable messaging runtimes for Slack, Discord, and WhatsApp, but every channel still requires its own credentials, permissions, and deployment configuration.
How is access controlled?
NPATI applies tenant and context isolation, scoped permissions, encrypted secrets, API-key and OAuth boundaries, tool allow/deny rules, human approval gates, credential and PII masking, prompt-injection checks, rate limits, and audit events. A live capability inspector shows what is available in the current installation.
From Answer to Executed Work
Give Your Company an AI Team with Memory, Knowledge, and Tools
Start with a request. NPATI resolves the right context, coordinates the right agents and models, uses only the capabilities you allow, and returns work you can inspect.