PORTWARE AI COWORK

Overview

PORTWARE AI COWORK is an advanced AI business assistant that lets you run PORTWARE with natural-language commands. From simple questions such as "What are my total sales this month?" to complex workflows such as "Create purchase quotations automatically for the products that are out of stock", you can do everything through a chat interface.

The module works with 7 AI providers (Anthropic Claude, OpenAI GPT, Google Gemini, DeepSeek, Mistral, LLaMA and Groq) and offers more than 120 ERP actions, automating your ERP work from CRM to accounting and from inventory to human resources.

How It Works

AI COWORK uses a three-stage agentic loop:

1. Understand (Gather): the AI analyses your request and reads the data it needs from the ERP (read-only actions run automatically).
2. Act (Action): it asks for approval before anything that writes, and runs it once approved.
3. Verify: it checks the change it made and confirms it was done correctly.
4. Report: it summarises what was done in your language.

Why AI COWORK?

Key Advantages

Module Statistics

Requirements

System Requirements

• PORTWARE, current release
• Python 3.10+ (ships with the server)
• An active internet connection (for AI API calls)
• A valid API key for at least one AI provider

Core Module Dependencies

Setting Up API Keys

Go to Settings > AI COWORK > API Keys and enter the API keys for the providers you want to use. One provider is enough.

Model Selection and Settings

You can configure the following parameters in Settings > AI COWORK > Model Settings:

AI Providers

PORTWARE AI COWORK works with 7 AI providers. Each has different strengths, speed profiles and pricing.

Auto Mode

Auto Mode lets the AI choose the best provider automatically according to the complexity of the task. You get the most efficient result by balancing cost and speed.

Smart Provider Selection Logic

Provider Comparison

Hands-free: live voice and new capabilities

Your hands stay on your work; PORTWARE works by your voice. The panel's microphone button opens live voice mode: the microphone stays open, every sentence is a turn, and the answer is both written and read aloud. The voice is the summary; the screen is the record — everything said and every answer (charts included) stays in the panel.

• Talking over it pauses the reading at once; room noise does not. "One moment / wait" pauses, "continue" resumes from the sentence that was cut, "end the conversation" closes it
• The assistant's voice goes through the echo-cancelling loop — on desktop and on the phone — so it never interrupts itself
• The conversation shrinks to the pill in the corner; listening and speaking carry on
• On the phone tapping the bar or the pill interrupts with a vibration; switching tabs or locking the screen does not end the conversation, and the lock screen has pause/stop
• Voice, speed and form of address live on the user record and follow you from device to device; the default voice is "Jarvis"
• The professional voice has a daily one-dollar cap; beyond it the same sentence is spoken with the browser's voice. API keys never reach the browser

Stand by, wake by name; it listens only to you

When you say "Cowork, stand by", it steps back with a single sentence and answers nothing until you call it by name ("Cowork, continue", "Cowork, come back"); the Wake button on the pill calls it too. The conversation belongs to the first person who speaks: a clearly different voice in the room cannot interrupt the answer, open a question or give an approval — unless it addresses the assistant by name. On the phone the voice is boosted through a limiter and is not turned down when the microphone hears something.

It knows who you are talking to

A sentence you say to someone next to you while the answer is being read does not interrupt it. Cowork decides whether something was said to it: a work request or question ("show me the overdue quotations"), being addressed as "you" or by name ("Cowork, …") counts as said to it; your own side remarks, other people in the room and greetings to others are ignored. When it is not sure, it does not cut in.

Voice commands

Commands are handled on the device, never sent to the AI, and work in the middle of an answer; "Cowork", "okay" or "please" can go in front.

As loud as a video on the phone: phones use Loud mode by default and play the voice as media. On an iPhone you interrupt an answer that is playing by tapping; on Android your voice interrupts it too. You can switch to Echo-cancelled mode from the voice menu.

It sees what you see and opens the screen for you

Every turn carries the screen you are on: filters, the rows in the list, the records you have ticked, unsaved changes. Say "show Kayra's open quotations" and the filtered list, calendar or chart opens at once — even while the conversation sits in the corner pill. Product codes are read out piece by piece, brand names with the right pronunciation.

Page Review

It reviews the open record — an empty required field, a broken e-mail, a tax or ID number or IBAN whose checksum fails, conflicting dates — and applies the fixes you choose with a single approval.

A plan is one approval

A multi-step job arrives as a single plan card; each step uses the result of the one before, and a step whose reference it cannot fill is skipped with the reason. "Book annual leave for Ali and Ayşe next week" runs on one card: type, number of days, missing allocation and approval — all or nothing.

Access doctor, Cowork Health, the learning loop

• A permission error says which record, which group and who can grant it, in plain language
• Cowork Health: one line a day — turns, answer time, failed actions, interruptions, voice clips, cost
• An undone action is a negative signal; proposed rules do not become rules until a reviewer approves them
• The morning briefing runs on its own on weekdays only; the compass button produces it on request any day

Cowork in the Office

The window in the PORTWARE Office editor is the same Cowork panel as in the back office — live voice, spoken approval and single-approval plans included. The "Go" command marks the change in yellow in Word and PowerPoint; Cowork reads, compares and restores document versions.

PORTWARE from Claude, ChatGPT and Cursor (MCP)

PORTWARE is open to every AI client that speaks the Model Context Protocol, at one address: https://<your-portware>/mcp. Claude (claude.ai, Desktop, Code), ChatGPT, Cursor and the like read and write PORTWARE with the rights of the person who connected them: the same access rights, the same record rules, nothing more. The tools are Cowork's own catalogue; a tool improved for Cowork is improved for every MCP client.

Connecting

What a client may do, and when

Resources: portware://company, portware://me, portware://me/agenda, portware://record/<model>/<id>, portware://attachment/<id> (PDF and Office text). Prompts: prepare a quotation, find stock shortages, summarise a task, my day. Protocol: MCP 2026-07-28 (and the earlier initialize revisions), OAuth 2.1 with PKCE.

Natural-Language Queries

With AI COWORK you query your ERP as if you were talking, in Turkish or any other language. You do not need to find menus or know SQL.

Data Query Examples

Web Research

Besides querying the ERP, AI COWORK can search the internet for real-time information. With the DuckDuckGo search engine and Wikipedia integration you reach up-to-date information.

Supported Search Sources

• DuckDuckGo: general web searches, company information, news
• Wikipedia TR/EN: encyclopaedic information, company history
• DuckDuckGo Infobox: basic facts such as a company's logo, founding date and sector

Weather

AI COWORK provides real-time weather information through the wttr.in API. It is useful for logistics planning, field service work and event organisation.

Keywords that trigger weather queries:

Smart Operations

For write operations that need approval, AI COWORK presents the user with a summary plan. It can plan and carry out multi-step workflows automatically.

The Approval Flow

The Morning Briefing and System Audit

Everything described so far works when you ask. The briefing is the reverse: every morning the assistant scans your CRM, activities, tasks, quotations, meetings and shipments without being asked, finds what has been forgotten and writes down what needs to be done.

How Does It Work?

• Every morning at 08:00 (Istanbul): a scheduled job runs once a day and prepares each user's own briefing.
• A star badge in the systray: when the briefing is ready the COWORK icon in the top bar carries a badge; opening the panel opens the briefing.
• No findings, no briefing: no empty notification is sent. If you see a badge, there really is something to look at.
• Personal scope: everyone sees only the records they are responsible for.
• A company view for managers: an authorised user also sees company-wide findings: unowned leads, unassigned tasks, late shipments, missing product data.
• Whenever you want, in one click: you do not have to wait for the morning; the briefing button in the panel runs the same audit on the spot.

22 Things Checked Every Morning

These are not “the AI found something”. Each finding is a named, countable query defined in code that gives the same result every time. 16 findings are personal and 6 are company-wide.

Not a Number, but What to Do

“259 records need fixing” is a report, and it moves nobody. The briefing writes a one-sentence remedy next to each finding, and that sentence is written by the server, not the model. The reason is simple: whoever wrote the query knows which field was checked; the model does not, and would make it up.

CRM Actions

Natural-language commands and actions for customer relationship management:

Sales Actions

Stock & Inventory Actions

Human Resources Actions

Accounting Actions

Project Actions

Purchase Actions

The Memory System

With its cross-session memory system, AI COWORK remembers what it learned across sessions. It records a user's preferences, frequently used query types and the operations they carry out.

How Does the Memory System Work?

• Automatic learning: user behaviour is analysed in each session and important information is saved
• Decay: old, unused information loses importance over time; current information stays in front
• Context extraction: the query types a user asks often, their preferred format and abbreviations are learned
• Cross-session: the memory survives even when the browser is closed; a week later, "do it like last time" is still understood

Workflow Engine

The State Machine Workflow Engine, inspired by LangGraph, manages complex multi-step workflows. The state of each stage is tracked, and recovery is automatic when an error occurs.

Workflow States

Features

• Automatic state transitions: allowed transitions are validated and invalid ones blocked
• Error recovery: automatic recovery with up to 3 retries
• Progress tracking: a visual progress indicator from current_step / total_steps
• Database persistence: the workflow state is stored in the database; it continues even after a page refresh
• Cross-module workflows: ready-made flows such as Quote-to-Cash, Order-to-Delivery and HR Onboarding

Subagent Orchestration

Complex tasks can be distributed in parallel to specialised subagents. Large tasks are split into smaller parallel parts, and the time shrinks.

Subagent Types

Parallel Execution

Several subagents can run at once (ThreadPoolExecutor, max. 4 workers). Tasks with dependencies start when their dependency chain completes.

The Plugin System

AI COWORK can be customised through an extensible hook system (plugin architecture). Development teams can add their own business logic and change the default behaviour.

Hook Types

Built-in Plugins

• LoggingPlugin: records every action automatically
• ValidationPlugin: validates input parameters
• MetricsPlugin: tracks performance metrics
• CachingPlugin: caches the results of repeated queries

Real-Time Token Streaming

For 5 providers, the reply streams into the interface token by token, in real time. Instead of waiting for long replies, you watch them being written.

Providers That Support Streaming

Technical Details

• Built on the core bus: the core's bus polling mechanism is used, not standard SSE (Server-Sent Events)
• Batching: sent in batches every 30 characters or 300ms (less DB commit load)
• No streaming during tool calls: only text replies stream; ERP actions run synchronously
• Live cursor: a blinking cursor shows in the interface while the reply arrives
• Fallback: if streaming fails, it falls back to non-streaming mode automatically

Multi-Model Pipeline

A cost-optimising multi-model pipeline for Turkish queries. Turkish content is translated into English by a cheap translator model, processed, then translated back into Turkish. That way economical models can be used instead of expensive language models.

Cost Comparison

Usage Metrics

See detailed metrics for every AI request in PORTWARE AI > Configuration > Analytics.

Metrics Tracked

Cost Analysis and Optimisation

Cost Reduction Strategies

• Auto mode: simple tasks go to cheap models, complex tasks to strong models
• Multi-model pipeline: 60x cost savings for Turkish processing
• Caching: repeated query results come back from the cache (zero cost)
• Tool filtering: only the tools needed are sent to the model (saves tokens)
• Prefer Groq: 500 tokens a second + low cost for simple queries

Example Use Scenarios

Scenario 1: Onboarding a New Customer

Scenario 2: Weekly Tracking of Orders Due

Scenario 3: Critical Stock Alert and Automatic Supply

Scenario 4: Monthly Financial Summary

Scenario 5: Approving an Employee's Leave

Frequently Asked Questions

Troubleshooting

Common Problems and Solutions