The new workflow debugger lets you run any workflow live, right from the builder, and watch every step as it happens.

What's new

• Run it live: Hit Save & run and your workflow starts in a real messenger inside the builder. No publishing, no separate tab.

• Watch each step: The canvas highlights the block that's running, so you can follow the exact path your workflow takes.

• See what's happening: A live event stream shows which branch a condition took, what an API action returned, what the AI decided, and more. Expand any step to inspect the details.

• Test as anyone: Run as a fresh guest, or as one of your real contacts to check targeting, conditions and anything that depends on customer data.

How it works

1. Open a workflow and click Test.

2. Pick Guest or a specific contact.

3. Click Save & run and interact just like a customer would.

4. Tweak a step, hit Save & restart, and try again.

Great for

• Catching broken conditions and API calls before customers ever see them

• Testing how a workflow behaves for different types of contacts

• Debugging bigger, multi-step workflows without the guesswork

Every test run lands in your inbox with a Test badge, so it's easy to find and never gets mixed up with real conversations.

Give it a try on your next workflow.

Busy conversations get messy fast. A customer fires off three questions in a row, a thread jumps between topics, and suddenly it's not obvious which message you're actually answering. Now you don't have to leave anyone guessing. You can reply directly to a specific message, right inside the inbox.

What's new

• Reply to any message: Hover over a message and click Reply to quote it in your response. Works for customer messages, internal notes, Kai's replies, and emails.

• A clear "Replying to" banner: Before you send, a banner above the composer shows exactly which message you're responding to, with one click to cancel if you change your mind.

• Quoted context in the thread: Your reply carries a compact preview of the original message, so everyone can see what it's about at a glance.

• Jump to the original: Click any quote to jump straight to the message it references. A real lifesaver in long threads.

• Works with everything: Plain text, rich text, and attachments all get clean previews.

How it works

Open any conversation in your inbox, hover over the message you want to answer, and hit Reply. Type your response as usual and send. The quote comes along automatically. That's it.

No busy threads left behind. Give it a try and let us know what you think!

Your knowledge base isn't one-size-fits-all — now your AI agents aren't either. With knowledge scoping, you decide precisely which content each agent can use to answer, with simple tags.

What's new

• Tag any content — help center articles, FAQ snippets, uploaded files, and crawled website pages — right where you manage it.

• Scope any AI agent to those tags: let it use only content with certain tags, or everything except certain tags.

• Keep internal knowledge internal — content tagged internal stays out of customer-facing answers by default.

• Manage all content tags in one place under Project settings → Tags → Content tags.

• Scope Kai too — set a project-wide default, or scope a single "Let Kai answer" step in your workflows.

Great for

• Internal vs. external: give your team's copilot access to internal notes, while customers only ever see public content.

• Multiple products or brands: point a Billing agent at billing content and an Onboarding agent at onboarding content — no cross-talk.

• VIP & segments: route VIP users to an agent scoped to VIP-only knowledge.

How it works
Add tags to your content, open your AI agent's Knowledge tool, and pick All knowledge, Only these tags, or All except these tags. That's it — the agent instantly answers from just its slice of your knowledge base.

Available now.

Your team’s emails just got a more professional, personal touch. You can now create email signatures in Gleap that are automatically filled in with each agent’s own details on every reply.

What’s new:

• Reusable signatures – build one signature for your whole team under Settings → Email → Signatures, starting from a ready-made template or from scratch.

• Automatically personalized – name, role, photo, phone, calendar link and more are pulled from each agent’s profile, so every reply is signed correctly.

• Live preview & multi-language – see exactly how it looks as you edit, with a separate version per language.

• Human replies only – signatures are added when an agent replies, never on AI/bot messages.

Add a signature to your reply email template and you’re done — every human reply gets a polished sign-off, automatically.

We just added GLM 5.2 to Gleap, and honestly, this one's a big deal.

Z.ai (formerly Zhipu) released GLM-5.2 on June 13, 2026, and it's their most capable model yet. Here's why we're excited to bring it into the platform:

A genuinely usable 1M context window. The window jumps from 200K in GLM-5.1 to 1 million tokens, with outputs capped at 131,072 tokens per response. For Kai working across whole repos, long ticket histories, and sprawling conversation threads, that's exactly the kind of headroom that matters.

Built for real engineering work, not just chat. The GLM series emphasizes coding, tool usage, multi-step reasoning, repository analysis, and long-running agent workflows — which lines up perfectly with how Kai actually operates.

The benchmarks back it up. On SWE-bench Pro, GLM-5.2 scored 62.1, beating GPT-5.5 (58.6) and its own predecessor GLM-5.1 (58.4). On the long-horizon FrontierSWE test it hit 74.4%, surpassing GPT-5.5 (72.6%) and finishing in a near-tie with Claude Opus 4.8 (75.1%). It also took first place on the Design Arena benchmark with an ELO of 1360 — not bad for an open model.

Open weights, no borders. Z.ai released the weights under an MIT license, with their docs explicitly noting no regional limits and technical access without borders. Great news for anyone who cares about model sovereignty.

Running on our own EU GPU cluster. We're provisioning GLM 5.2 ourselves on our own GPU cluster in the EU — no third-party API, no data leaving the region. Your prompts, your code, your customer data stay inside our infrastructure. That's the whole point of an open-weights model: we get frontier-class capability and full control over where it runs.

And it's seriously cost-effective. It beats GPT-5.5 on multiple long-horizon coding benchmarks at roughly 1/6th the cost.

GLM 5.2 is available in Gleap now — pick it from the model selector and give it a spin.

We just added two of the most exciting models on the market to Gleap. One is the new frontier ceiling. The other is the best price/performance deal we've seen in a long time. Both are live for Kai today.

Here's the short version: pick Fable 5 when you want the absolute best, pick M3 when you want fast, cheap, and still really good.

Claude Fable 5 — frontier, no compromises

Fable 5 is Anthropic's first publicly available Mythos-class model, and it's the most capable model they've ever shipped to the general public. It's state-of-the-art on basically every benchmark that matters for the work Kai does — software engineering, knowledge work, vision, long-running tasks. The longer and messier the job, the bigger its lead.

In practice this means Kai can take on the gnarly stuff: deep multi-step reasoning, complex tickets that used to need a human, and long autonomous runs without losing the thread.

The catch: it's expensive. Fable 5 is the premium tier and priced like it ($10 / $50 per million tokens at the source). So we'd recommend it for your highest-value workflows where quality wins over cost — not for blasting through thousands of routine tickets. For anything in regulated high-risk territory (security, bio, chemistry), Anthropic automatically falls back to Opus 4.8, so you don't lose coverage there either.

Use it for: your hardest agentic tasks, complex resolutions, anything where "good enough" isn't good enough.

MiniMax M3 — the balance pick

M3 is the one we're most excited about for everyday use. It's an open-weight frontier model with a 1M-token context window, native multimodal input, and genuinely strong coding and agentic performance — it beats GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro and gets close to the top Claude models.

And it does all that at a fraction of the cost. We're talking a different league on price — which means you can run M3 across your whole inbox without watching the meter.

The 1M context window is the part that quietly matters most for Kai: it can hold huge amounts of conversation history, docs, and code context at once, which makes resolutions sharper and handoffs cleaner.

Use it for: your default. High volume, fast responses, long context, great quality, low cost. This is the workhorse.

How to switch

Head to your Kai settings and pick the model per agent. You can mix and match — run M3 as your default and route the heavy hitters to Fable 5. No migration, no downtime, just flip the switch.

As always, more power on autopilot. Questions? Ping us — we read everything.

The Gleap team

You've been asking for it — and it's finally here. You can now bring YouTube videos directly into Kai's knowledge base, so your AI can answer questions using the same product walkthroughs, tutorials, and webinars you've already published.

Drop in a single video, a full playlist and Kai takes care of the rest. We analyze each video frame-by-frame and end-to-end — no transcripts to upload, no manual chapter tagging — so Kai understands not just what's said, but what's actually shown on screen.

What's included:

• Single videos, playlists. Paste one URL or batch-import hundreds at once.

• Smart chapters. Each video is automatically broken down into chapters with titles, timestamps, summaries, key facts, and standout quotes.

• Visual understanding. Kai picks up on what's demoed on screen — UI elements, diagrams, code — not just spoken audio.

Why this matters: If you've ever recorded a walkthrough for a feature that users keep asking about, you can now turn that 12-minute video into instant, cited answers — without rewriting a single help article.

Getting started:

1. Head to AI agents → Content → Add content → YouTube

2. Paste any YouTube video, playlist

3. Watch the videos light up as Kai ingests them

A quick note: Kai can only analyze Public videos. Unlisted and Private videos aren't accessible through YouTube's API — flip them to Public in YouTube Studio and re-import.

Let us know what you point it at first

Writing documentation takes time you'd rather spend building. Our new AI-powered Chrome Recorder extension changes that — just record your screen, and Kai turns it into a polished help center article or changelog entry, ready to publish.

How it works:

• Install the Gleap Chrome Recorder extension from the Chrome Web Store

• Choose whether you want to create a help center article or a changelog entry

• Hit record and walk through the feature, fix, or update you want to document

• Kai automatically detects the key steps and picks the right moments for screenshots — most of the time, there's nothing to adjust

• Want to fine-tune it? You can reorder or edit steps, add annotations, or turn any step into an animated GIF before submitting

• Kai then writes a structured draft, which you can review and publish directly from your Gleap dashboard

Because Kai actually understands what's happening on screen, the drafts aren't just transcripts. You get proper headings, clear instructions, and screenshots pulled from the moments that matter.

Read more about creating help center articles or creating release notes with the Chrome Recorder.

We're excited to announce Telegram as a new support channel in Gleap! You can now connect a Telegram bot and manage all customer conversations directly from your Gleap inbox - just like WhatsApp, Instagram, and Facebook Messenger.

What's included:

• Receive and reply to Telegram messages from your Gleap inbox

• Full media support - images, documents, videos, voice notes, and stickers

• AI bot and workflow support with interactive inline buttons

• Typing indicators shown to customers while agents compose replies

Getting started:
Create a Telegram bot via @BotFather, paste the token in Settings → Channels → Telegram, and you're live in under 2 minutes.

You can now power Kai with GPT-5.4 - OpenAI's latest and most capable model yet.

Expect sharper, more accurate responses with significantly fewer hallucinations, improved agent skill execution, and faster response times across the board. A meaningful upgrade for any team looking to get the most out of their AI support experience.

Head to your Kai settings to switch models and give it a try (AI agent → Settings → Models).