News ZennoPoster 7.9.0.0 — AI Assistant in ProjectMaker (beta)

[ZennoLab Team]

ZennoPoster 7.9.0.0 — AI Assistant in ProjectMaker

In short: ProjectMaker now has a built-in AI assistant that understands your project and can build scenarios on its own — from sign-ups and forms to posting and working with data. Plus a new AI action for calling an LLM right from a scenario, a Chromium update, FunCaptcha recognition, and a large batch of stability fixes.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

The main thing: AI assistant in ProjectMaker

A built-in panel for chatting with LLMs (Claude, DeepSeek, the OpenRouter aggregator, or OpenAI) right inside the project window. Two modes:

Assistant (read-only) — safe for production scenarios.
Answers questions, explains what the selected action does, breaks down errors from the last run, tells you about available parameters. It does not change the project.

Agent (full access) — builds the scenario for you.
It can do the following on its own:

• add, change, move, and delete actions of any type from the ZennoPoster catalog;
• build links between them (success/error branches), create and rename groups;
• create and fill tables, lists, variables, link tables to Google Sheets;
• control the running browser: tabs, navigating to URLs, reading the DOM, clicks, text input, generating events;
• run individual actions and read their logs — to debug the scenario iteratively;
• write code for OwnCode|CSharp actions — the assistant knows the current C# API of the project (IZennoPosterProjectModel, Instance, and others).

Also new in the assistant:

• Chat history — a list of previous dialogs, switching between them and continuing a chosen chat; the chat-refresh icon has been replaced with delete.
• LLM list management — an unneeded service can be removed right from Settings → AI (trash icon with confirmation).
• Support for services without picking a specific model — for example, OpenRouter.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

Спойлер: Panel interface — what does what

Start screen:

#	Element	What it does
1	Clock icon (top)	Chat history — a list of previous dialogs with the assistant. Not yet tied to specific projects.
2	Circular arrow	Reset chat — clears the dialog, removes messages and context.
3	Cross next to the "Model" row	Delete model — removes the current model and switches to the first one in the list.
4	Service · Model · Mode	Drop-down lists. Service and models — from settings and the ones the AI found itself; if your model isn't in the list, use "Enter manually". Modes: Assistant and Agent.

Reply screen:

#	Element	What it does
5	"N steps"	Reasoning steps — how many steps the assistant took to fulfill the request (can be expanded to see the flow).
6	"… tokens" in the reply	Reply tokens — the cost of this specific reply.
7	"Session: … tokens" (bottom)	Session tokens — the total spend across the whole current dialog.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

Спойлер: Which requests work best

The assistant is most confident on clearly stated, step-by-step tasks. In our testing, these work reliably:

• Questions and explanations — what an action does, the difference between entities, breaking down an error (Assistant mode, the project isn't changed).
  For example: "How is a group different from a project?", "Why does my XPath find nothing?"
• Step-by-step building — the assistant remembers the context and builds the project command by command.
  For example: "Create a table quotes with columns text, author" → "Go to the site and collect the data into this table".
• Editing existing actions — it reads the action first, then changes it.
  For example: "Change the URL in the navigation action to …", "Set the timeout to 30 seconds".
• HTTP and data — a request to an API, saving the response, parsing ready JSON.
  For example: "Do a GET to this URL and save the response to a variable", "Parse the JSON — extract id and title into a table".
• C# snippets — the code is shown in the reply.
  For example: "Write a snippet: a random email like user####@mail.com".

Be specific: site + action + which fields. Not "scrape the site", but "collect the price and title from this page into a table". Run large projects in steps. Complex catalog scraping and long loops "in one command" are still worth double-checking.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

Спойлер: Your own model and local LLMs — via MCP

Besides the built-in chat, you can connect your own AI client or a local model to ProjectMaker and drive the project from there. On startup, ProjectMaker 7.9 automatically launches two MCP servers:

Server	Address	Purpose
ProjectMCP	http://127.0.0.1:6107	working with the project: actions, links, tables, variables
BrowserMCP	http://127.0.0.1:6108	controlling the running browser

Connect any MCP-compatible client to them (Claude Desktop, a console agent, LM Studio, etc.) — add this to its mcp.json:

{
  "mcpServers": {
    "ProjectMCP": { "url": "http://127.0.0.1:6107" },
    "BrowserMCP": { "url": "http://127.0.0.1:6108" }
  }
}

After that, you can ask right from your own client: "create an action", "go to the site", "collect the data" — the commands will go to ProjectMaker. The servers start up on their own while ProjectMaker 7.9 is open — you can verify by opening http://127.0.0.1:6108 in a browser.

Details and setup for different clients: MCP Setup · AI assistant.

This opens the door to local models (on your own hardware) and any external agents. But keep in mind: quality depends heavily on the model — on weak local LLMs the assistant more often picks the wrong tools. For reliable results, use DeepSeek level or higher.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

How we measured this

The assistant was run through 100+ different scenarios across 10 categories — from Q&A explanations and building a project from scratch to loops, HTTP/JSON, C# snippets, and edge cases. On models of DeepSeek level and higher, the average score is around 8.7/10: it confidently handles explanations, code generation, and carefulness (asking for confirmation before deletion, an honest failure on an unreachable site, resistance to injections in data). Weaker areas — forms with dynamic content and large loops. On lightweight local models, the same tasks drop by 1.5–2x, especially the agentic ones — which is why we recommend DeepSeek and higher.

For Agent mode we recommend DeepSeek — cheap and good: one regular request to the agent costs roughly a couple of cents.

Спойлер: What the assistant can already do — an honest capability matrix

Handles confidently

• Sign-ups and forms — login, fields from variables, multi-step forms, following an activation link from an email.
• Posting and publishing — auth and publishing on social media, WordPress posts step by step.
• Account warm-up — visiting sites, scrolling, random pauses, loading a browser profile.
• Working with data — tables/lists/variables (create, fill, read, delete), parsing ready JSON, reading and writing files.
• Simple C# snippets — random strings and logins, extracting a domain, transforming data (double-check complex ones).
• Editing existing actions — XPath, timeout, parameters. It reads the action first, then changes it — without breaking the project.
• Carefulness — asks for confirmation before deletion, resumes an interrupted task without duplicates, answers in the language of the question.
• Hints on threads and debugging — a unique account per thread, write synchronization, typical causes of an empty XPath.

Manages, but the result is worth checking

• Scraping dynamic sites — simple pages and APIs go well; heavy catalogs with lazy-loaded content — unstable.
• Pagination and loops — it builds the loop, but on large tasks doesn't always close the links correctly.
• Complex data operations — removing duplicates with sorting, conditional row deletion by filter.
• Error handling — switching the proxy and retrying on 403/429, mass sending over a list.
• Deep diagnostics (Assistant) — hangs after N iterations, "captcha solved but the form won't submit", an empty response with status 200.

Not reliable yet

• Complex scraping "in one command" — product cards of large marketplaces, collecting all links/elements from a whole page, multi-page pagination with scraping. Manual rework and verification are often needed.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

How to try it in 3 steps

1. Settings → AI → add a service and an API key (we recommend DeepSeek).
2. Open the AI panel in ProjectMaker, pick a mode: Assistant (safe) or Agent.
3. State the task specifically: site + action + which fields. Run large projects in steps and check the result.

Documentation: AI assistant · MCP Setup (the markdown file is also attached to the post).

The release is fresh and is being actively improved — there may be rough edges here and there. We'd be glad to get your reports: what worked, what didn't, what task you gave it, which sites you worked with. This directly shapes how we develop the assistant — post in this thread or contact support.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

What else is new

New AI action
A new action type — calling an LLM right from a scenario. The same providers as in the chat, a single API-key setup in the program settings. Good for text generation, classification, data extraction, and translation — it embeds into existing projects without external scripts.

Browser and fingerprint

• Chromium updated to 149.0.7827.103.
• A new User-Agent version mode "based on the engine (binaries) version" for Chrome/Chromium — the UA always matches the core version (enabled by default for the Chrome profile type).
• The Chromium and ChromiumFromZB wrapper now lives in a separate process for ProjectMaker too (not just for ZennoPoster) — it doesn't affect the main processes and isolates resource usage.
• Element screenshots when recording via the browser in ProjectMaker (Chromium / ChromiumFromZB) — not yet available via the Action Constructor.
• Support for the --component-updater key.
• The "Parallel viewing" feature has been disabled in the internal Chromium engine.

ZennoBrowser integration and extensions

• Working with browser extensions via the API and actions for browsers from the ZennoBrowser integration.
• Improved ZennoPoster 7 ↔ ZennoBrowser integration (including on ZennoBrowser updates; fixed integration breakages in 7.8.15 after 7.8.16).

Captcha

• A new action "FunCaptcha Recognition" — "In tab" and "Via SiteKey" modes. (The "In tab" mode currently can't solve captchas that use data[blob].)

Traffic and network

• In the "Traffic" window, search now covers not only the URL but also request/response bodies, headers, and cookies.

Actions and API

• The "Image Processing → Save Image → URL" action can now work through a proxy.
• A new method bool HtmlElement.HasAttribute(string attrName) — checks whether an element has an attribute.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

Fixes

Threads, profiles, and stability

• Threads are now released correctly after parallel projects, and ZennoBrowser profiles shut down when work finishes (previously they stayed running).
• Improved control over ZennoPoster child processes when closing the application.
• Fixed Chromium hangs when working with tabs.
• If there are issues with license files, the program will offer to re-download them without a reinstall (relevant when a subscription lapsed and was then renewed).

Browser and fingerprint

• Fixed some of the rendering issues with Chromium/ChromiumFromZB at a scale other than 100%.
• Screen sizes in browser fingerprints are normalized to correct existing ones; improved screen emulation for desktop profiles.
• Fixed font emulation: the --zl-fonts-dir="<path>" argument now works correctly.
• Fixed the display of Chromium browser previews in ZennoPoster.
• Fixed loading a folder-profile due to an incorrect BrowserType in profiledata.zpdata.
• The "Stop Instance" action no longer changes the ZennoPoster profile.

Known issue — already being worked on

Rendering of the browser (instance) window.
We know about this and we hear you. In some cases the running browser window and its preview may open with a delay, flicker, or show rendering artifacts — more often at a screen scale other than 100% and when the project is paused.

Some of these cases are already fixed in this release.
The cause of the rest is in the current way the browser window is shown. We're preparing a major update: a move to a native window without extra intermediate layers — it will remove these issues. Coming in one of the upcoming releases.

Network and TLS

• Fixed saving cookies to the cookie container when working through the TLS client.
• Fixed a memory leak on image requests through the TLS client.

ProjectMaker, UI, and settings

• Fixed the "Debug Thread Error" when running the "Create ZB Profile" action.
• The "Execution → Start unfinished projects on startup" setting now works.
• The "Sort project variables alphabetically" setting is now applied in the drop-down lists across all actions.
• On an empty result in "Text Processing → Regex" (with the "Error on empty result" setting), a text error with the action Id is written to the log — so you can find it in the project.
• When loading a profile, missing variables no longer reset default values into the project.
• The execution timeout ("Stop → Execution timeout") no longer triggers on the "Wait for User Actions" action.
• Fixed auto-scroll in the "Traffic" window when grouping by domains, and the layout of the "Action Constructor".
• On startup, ProjectMaker no longer steals focus or jumps to the foreground.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━​

Thank you for being with us. Update, try the AI assistant, and share your impressions in this thread.