Changelog

GPT-6 Astra is now supported in the OpenAI frontier recipe for Deep intelligence. The update does not switch your organization or personal recipe selection.

Review your active recipe

Owners and admins can set the Organization default. To choose a different recipe for your own work, turn on Personal recipe and select a recipe card. Your personal choice takes precedence over the organization default; turn it off to follow the organization default again.

Compare each recipe's recommendation and relative cost indicator before choosing. Adapt continues to match intelligence to the task when your chat is set to Auto.

Learn more about intelligence recipes.

Adapt still sizes its own intelligence to the work by default. Now you can override it. A picker in the composer lets you pin a chat to Deep, Balanced, Fast, or Super Fast when you know what the work needs, and Adapt holds that level for the whole conversation.

Pick a level

The picker sits in the composer toolbar, next to the attachment and integration buttons. It reads Auto until you change it.

• Auto is the default. Adapt reads each task and matches its intelligence to the work, shifting levels as the work changes character.
• Deep puts Adapt's most capable reasoning on hard, open-ended problems.
• Balanced handles multi-step reasoning.
• Fast and Super Fast keep routine, well-defined work snappy.

It sticks

Pick a level on a new chat and it applies from the first message. Once a level is pinned, Adapt won't re-route the chat on its own, so a conversation you set to Deep stays Deep for every turn.

You can change it mid-conversation from the same picker in an open chat. Switch back to Auto to unpin the chat and hand routing back to Adapt.

Which models sit behind each level depends on your organization's model recipe, chosen by owners and admins under Settings > Intelligence.

Adaptive Intelligence · Core concepts

Adapt still decides how much intelligence each task needs. Now you decide which models it reaches for. Owners and admins can pick a model recipe for the organization: a named set of models mapped across all four intelligence levels, chosen for cost, speed, or raw reasoning ceiling.

Pick a recipe

Go to Settings > Intelligence. Every recipe is laid out as a card showing exactly what you get: the model and reasoning effort behind Super Fast, Fast, Balanced, and Deep, whether the stack is Frontier or Efficient, and a relative cost indicator from $ to $$$$ so you can compare spend at a glance.

Select a recipe and it applies immediately to everyone in the organization. There's nothing to publish or roll out.

The recipes

New organizations start on the Anthropic stack.

What doesn't change

You still never pick a model to send a message. Adapt reads the work in front of it and picks the intelligence level, the same as before. The recipe only decides which model answers at each level, so switching stacks is a single decision an admin makes once instead of a choice everyone makes on every request.

The page is visible to owners and admins only. Members keep working exactly as they did.

Adaptive Intelligence · Core concepts

One person running a heavy automation shouldn't be able to spend a month of credits in an afternoon. Organizations on an active subscription can now cap how many credits each member spends per billing cycle, so costs stay predictable without anyone policing usage by hand.

Set the limit

Owners and admins set it in Settings > Billing > Member spend limit. Enter a credit limit, save, and it applies to every member in the organization for the current billing cycle. Clear the field to remove the cap.

The limit resets when your billing cycle rolls over. Annual plans get a fresh window on each monthly anniversary rather than one lump allowance for the year.

Warnings before the wall

Adapt emails each member as they approach their cap:

• 75% — a heads-up, no action needed
• 90% — close to the cap, with a note that messages pause at 100%
• 100% — the limit is reached and new messages are paused

Admins get admin-specific wording telling them they can raise the limit themselves. Each email fires once per billing cycle, and all three re-arm when the cycle rolls over or an admin changes the limit.

What members see

A member at their cap sees a banner in chat explaining that new messages are paused until the limit resets or an admin raises it, and their own credits card shows how much of the monthly limit is left and when it resets. On the usage page, admins see spend against the limit for every member, with a Blocked badge on anyone at the cap.

Usage-based pricing · Billing FAQ

Adapt can now run in Slack as your own branded app. Use your organization's name, icon, and description, powered by Adapt.

How it works

• Customize your app's name and description in Settings > Integrations > Slack.
• Create the app in Slack with a ready-made manifest, then paste your credentials back into Adapt.
• Authorize the app and invite it to your channels. Your team tags your bot by its new name — @yourcompany — and gets the full Adapt experience: answers, actions, automations, and proactive responses.

Notes

• The branded app and the Adapt app can be connected to the same Slack workspace at the same time. Conversations that start with your branded bot stay with it.
• Your branded app is a Slack app you own, created from Adapt's manifest, powered by the same Adapt agent. Capabilities are unchanged.
• To update branding later, rename the app inside Slack under App Home and Display Information; the name shown in Slack follows the Slack app, not the settings saved in Adapt.

Knowing your credit balance is useful. Knowing what consumed it is what actually lets you act. The usage page in Settings > Usage now answers both: a daily chart of what you spent, four ways to slice it, and warning emails before you hit the bottom.

Credits by day

The top of the page shows total credits for the period, split into LLM (tokens models consume reading your prompts and writing responses) and Compute (resources that run your requests end to end). Below it, a stacked bar chart plots the same split day by day, so a spike is immediately visible and immediately dated.

The date range defaults to the last 30 days and goes back up to 90.

Four ways to slice it

Group by switches the breakdown table between four views. Each row carries an activity sparkline, chat count, credits, and share of the period's total.

• Chat — every chat with its creator, source, last used time, and credits. Sort by Top usage to find the expensive ones or Last used to find the recent ones.
• User — credits per person on the team.
• Source — where the work came from: Web, Slack, GitHub, Linear, iMessage, MCP, Extension, or Automation. Scheduled tasks and webhooks both roll up as Automation, so all unattended work sits in one row.
• Intelligence Level — credits by Super Fast, Fast, Balanced, and Deep, so you can see how much of the bill comes from the heaviest reasoning.

Admins see the whole organization. Members see their own chats.

Know before you run out

The credits card on the billing page now leads with what's left in plain terms: the percentage still available and the exact figure, as in "29% credits available — 8,790,438 of 30,002,500 credits". The bar turns amber past 75% used and red past 90%.

Adapt emails owners and admins at those same two marks, once each per billing period, so a quiet month of automations doesn't become a surprise. Every email links straight to the usage page.

Usage-based pricing · Automations · Billing FAQ

Wiring an external service to an Adapt webhook used to mean piecing together the URL, the username, and the password from the webhook's detail page. Every webhook now has a Connect button that opens a dialog with the exact credential the sending service expects, ready to copy.

Two ways to connect

• URL secret — a single ready-to-paste URL with the secret already in the query string. Use it for services that only let you enter a URL, like Grain, Resend, or Typefully.
• Basic auth — the connection URL, username, and password as separate fields. Use it for callers you control, such as curl, scripts, or internal tooling that can set an Authorization header.

Secrets stay masked

The dialog shows a masked secret by default. Adapt fetches the plaintext only when you copy it, or when you reveal it with the eye toggle on the Basic auth tab, and drops it again as soon as you hide the field or close the dialog.

Lost the secret, or wiring up a new service? Regenerate secret issues a new one. The old secret stops working immediately, so update anything already pointed at the webhook.

Webhooks created before Adapt began storing a retrievable copy of the secret show "Secret is no longer readable" — regenerate to get a working one.

Learn more about event-based automations →

The MCP server is out of early access and into beta. The change that got it there: authentication is now standard OAuth, so there's no session token to dig out of your browser's cookie jar and no expiry to babysit. Add the server, approve the browser prompt, and your client handles credentials from there.

What's new in beta

• Sign in through your browser. MCP-capable clients discover Adapt's authorization server automatically and run the OAuth flow themselves.
• Credentials that maintain themselves. Access tokens are short-lived (5 minutes) and your client refreshes them silently. Your session stays valid until 30 days of inactivity.
• No org header. Your organization comes from the token, so X-Adapt-Org is no longer needed.
• Long jobs survive your client's timeout. Requests that run past 45 seconds return a chat_id and status: working instead of holding the connection open. Your client polls that chat_id until the run finishes, so a five-minute research task no longer trips the 60-second tool timeout most MCP clients ship with.

Connecting is now a single command:

claude mcp add --transport http adapt https://app.adapt.com/mcp

Reconnect existing clients

Alpha used your Adapt session token as a Bearer credential. That's no longer accepted. If you connected during early access, remove the server and add it again without the Authorization header:

claude mcp remove adapt
claude mcp add --transport http adapt https://app.adapt.com/mcp

Codex users should drop the http_headers line from the [mcp_servers.adapt] block in ~/.codex/config.toml and run codex mcp login adapt.

The endpoint and the ask_adapt tool are unchanged. See MCP server for full setup.

Automations could already deliver to your email or your Messages thread. Now any chat can.

Ask for it in the moment and Adapt sends it:

Email me this cost breakdown so I can forward it to finance.

Text me the three blockers from this thread before I get on the plane.

The answer still appears where you asked. The email or text is a copy you can act on away from your desk — with a link back to the full chat.

Route the whole chat

Set a destination once and every response in that chat follows it:

Send my responses in this chat to my email from now on.

Three things worth knowing:

• The preference is per person. In a shared chat, teammates keep their own destinations.
• The current answer stays where it started. Ask Adapt to email or text it too if you want it immediately.
• Automations you create in that chat inherit the destination. Route a chat to email, then say "every weekday at 7:30am, summarize overnight Sentry errors and open Linear bugs for the API service" — the daily summary lands in your inbox instead of the channel you happened to be standing in.

Say origin to go back to answering wherever each message came from.

Getting set up

Email needs a verified account email. Messages needs a connected phone and an existing text thread with Adapt. Adapt checks before it sends and tells you when a destination isn't ready.

One limitation: replying to a text Adapt sent you starts a new conversation rather than continuing the chat it came from. Reply threading is next.

Result delivery · Message Adapt

Automations used to answer wherever they were created. If you set one up in a Slack thread, that's where the result went — even when the thing you actually wanted was a text at 8am or an email waiting in your inbox.

Now every automation carries a result delivery setting. Pick the channel it was created in, your email, your Messages thread, or no notification at all. The setting sits in the header of the automation, and you can change it in plain English from any chat.

Your options

Email and Messages deliver to the person who created the automation. Both need to be set up first: Email needs a verified account email, and Messages needs a connected phone plus an existing text thread with Adapt. The selector shows each one's status — a masked phone number when Messages is ready, "Not connected" when it isn't — and won't let you pick a destination that can't receive.

Setting it

Open Automations, click any automation, and use the delivery selector in the header next to the status control. Only the automation's creator and org owners and admins can change it.

Or just say it:

Text me the results of my morning PR digest instead of posting here.

Email me the Monday sales summary.

Stop notifying me when the hourly inbox sweep runs.

Adapt checks which destinations you have available before it switches, and tells you if the one you asked for isn't connected yet.

Automations · Message Adapt