July 2026

The agent chat in the Interface tab now takes attachments. Click the paperclip, paste an image from your clipboard, or drop files onto the composer. Images become thumbnails you can hover to enlarge, other files become chips, and both stay on the message when you reload the page.

• Attached images now reach models that can see them. Images used to be dropped on the built-in LLM agent, so a model that could read them never got them. The five CLI coding agents fetch attachments from their sandbox instead.
• Attaching an image to an image or video model uses it as input rather than describing it: image models treat it as a reference image, and Veo and Kling use it as the first frame of the video when you have not set one.
• You can send an attachment with no text at all.

When a workflow has more than one entry point, the popup lists them so you can untick the branches you do not want this time. Each row names what it leads to, such as "Then Slack, Gmail" or "Uses Telegram". An agent entry point comes with a box for its opening message, prefilled from the node's own message, and when that agent is the only thing you run, its answers land in the chat, which opens as soon as the run starts.

You can now give an AI agent its own environment variables so it can call APIs that have no dedicated node. In the agent's Credentials panel, expand Advanced and add a set of names and values under Environment variables. Each variable is available in the agent's sandbox shell, so the agent can reference it directly, like $STRIPE_KEY in a curl command. Saved values are hidden and can be reused across agents. The agent can read a value by running a command, so scope each key to only what it needs.

Your workflows and agents can now work with email attachments everywhere email happens. When an email arrives, or when you fetch a single message, its attachments come through with their name, type, and size, and the text inside small documents is pulled out for you automatically, so an agent can read an attached invoice or contract with no extra step. You can also attach files to the emails you send from Gmail, Outlook, Resend, Mailgun, the Send Email node, and an agent's reply, using workflow files or URLs, up to 10 MB per file and 20 MB per email.

Scanned PDFs with no text layer can be read with AI text extraction. It is used only when you ask for it, costs 0.02 credits per page, and is capped at 20 pages per document. Reading the text already inside a document stays free.

Agents can now email you while you are away, and this is on by default. When an agent needs an answer, wants to hand you a link to connect an account, or hits a failure it cannot get past, it emails you at your account email. Reply to the email to answer, and the conversation picks up from there. You can turn this off for any agent in the chat Settings under Capabilities, or from the opt-out link in any of those emails.

Every agent chat now has a Files menu that lists the files the agent saved in this conversation. Open a file to preview it in place, whether it is text, markdown, code, or an image, and download any file with one click. Files the agent just wrote can take a few seconds to appear, and a refresh button reloads the list.

Agents can now propose a change to the workflow they run in. The suggestion arrives in the chat as a Proposed workflow edit card. Choose Send to builder to open the builder with the change ready to apply, or Dismiss to skip it.

Open an agent's fullscreen chat and click Share to mint a public link anyone can chat with, no account needed. The visitor sees the agent, a row of the tools it can use, and a full chat with its own thread. Every run bills to you, the workflow owner. You can pause the link or reset it to a fresh URL, and the old one stops working right away.

The WhatsApp node can now read your chats when you connect it by QR code. New operations read a chat's message history, list your contacts, mark a chat as read, and edit a message you already sent. You can reply directly to a specific message, and a chat picker lets you choose a conversation by name instead of typing a number. Agents with WhatsApp as a tool can look up chats the same way.

• When an agent run stops, the error banner now explains what went wrong and offers one-click fixes. Add credential opens a popup to connect or select credentials, and Change model opens the full model picker with search, provider filters, and favorites.
• Clicking the model name at the top of the chat opens that same model picker, the one the canvas agent node uses.
• A run that fails no longer shows leftover output from an earlier run, and a stuck run now clears with a clear error instead of spinning.

New AI agents now show in the workflow's Interface tab by default, so you can chat with one the moment it is built. After the AI builder finishes a workflow with a chat agent, a one-time walkthrough points you to the Interface tab to try it. You can still turn Show in Interface off for background agents, and existing workflows are unchanged.

When the AI builder creates a workflow with an agent that runs on a CLI coding agent, it now asks you to connect that agent's credential before calling the workflow ready, so you no longer hit a disconnected-credential error on your first run.

The CLI coding agents (Claude Code, Codex, OpenCode, OpenClaw, and Hermes) now keep a warm sandbox between turns instead of starting fresh each time, so replies come back faster and one agent can hold several conversations at once. A small indicator on the agent node shows how many sandboxes it is currently holding, with a live note while a turn is running.

CLI coding agents run in a cloud sandbox that NoClick provides, and that sandbox time is now charged at 0.05 credits per minute of uptime, billed per turn. This applies even when the agent runs on your own API key or AI subscription, since the sandbox is NoClick's compute. The built-in LLM agent runs in-process and has no sandbox charge.