> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getcollate.io/llms.txt
> Use this file to discover all available pages before exploring further.
# Breaking Changes - Collate AI & Credits | Official Documentation
> AskCollate is renamed to Collate AI, credits become weighted and cache-aware, runs that can't finish are refused up front, and the Slack app moves to agent mode in Collate 2.0.
# Collate AI & Credits
The product rename is the visible part. Underneath it, the credit model, the desktop app's
version gate, and the Slack/Teams integration all change in ways that affect anyone forecasting
spend or relying on existing automation triggers.
## AskCollate is renamed to Collate AI
**Breaking for anything that matches on the product name.** AskCollate is renamed to **Collate AI**
everywhere: the UI, the backend, billing labels, chat prompts, and the Slack and Microsoft Teams
integrations. This is a label change across the product, not a new feature, but it breaks exact-match
automation.
Grep saved searches, dashboards, runbooks, alert rules and Slack/Teams workflow triggers for the
literal string `AskCollate` and update them to `Collate AI`. Billing exports carry the new label too,
so any spreadsheet or BI report keyed on the old one needs repointing.
## Credits become weighted, and unfinishable runs are refused up front
**Behavioural.** The same workload consumes a different number of credits in 2.0 than in 1.13, and
runs that can't complete no longer start.
1.13 charged credits from a flat token count. 2.0 introduces a weighted, cache-aware model:
| Change | Effect |
| ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Weighted by token type | Input, output, cache-read and cache-write tokens carry different cost ratios, so a cache-heavy conversation costs materially less than the same token count of fresh generation |
| Weighted by model | Credits are weighted by the model that actually served each call, so a request routed to a cheaper model costs fewer credits |
| Fixed divisor | The weighted basis is divided by a locked `tokensPerCredit` of 58,000 |
| Cache-only usage is recorded | Lump-sum recording is guarded on the weighted basis, so cache-only usage is no longer silently dropped |
| Per-conversation attribution | Credits are attributed to the conversation that spent them, not just to the tenant pool |
If you forecast spend from 1.13 token counts, re-baseline after upgrading rather than converting the
old numbers — re-run a representative week and compare.
Separately, 2.0 warns before the credit limit is reached and **refuses to start work that cannot
complete** within the remaining pool. In 1.13 a long automation could start and exhaust the pool
part-way through, leaving partial results and no credits; in 2.0 the run is rejected up front with a
clear reason. Operators get a live view at `GET /v1/status/credits`, and credits can be refunded from
the operator API.
## Billing is reorganized and admins gain read access to conversations
**Additive**, but with a policy implication worth flagging: the **Billing** area is reorganized into
Overview, Product, User Adoption, AI Credits and MCP requests, backed by new endpoints
(`GET /v1/billing/credits/conversations`, `GET /v1/billing/adoption/users`,
`GET /v1/billing/weekly-impact`). AI chat response time is now measured server-side rather than in
the browser.
Admins can now read any conversation from Billing in read-only mode, so credit spend can be
explained. This is new. If your policy forbids administrators reading user conversations, restrict
the admin role accordingly before upgrading.
Conversations also become collaborative (additive) — a conversation is no longer private to its
author, and collaborators are emailed when added. If your compliance posture assumed conversations
were single-user, review it before upgrading: content previously visible only to its author can now
be shared by that author via `PUT /v1/assistants/chatConversations/{id}/collaborators/{userId}`.
Collate's MCP integration also gains additive changes in 2.0:
| Change | Class | Detail |
| ---------------- | ----------- | ----------------------------------------------------------------------------------------------------------------- |
| Tool annotations | Additive | Collate tools are annotated with title, read-only, destructive, and open-world hints |
| Size budget | Behavioural | Collate tools route through the shared response size-budget net, with null-guarded error messages |
| Usage tracking | Additive | MCP tool-call usage is tracked and rolled up daily into telemetry, surfaced in the Billing MCP requests tab above |
See [Applications & Automation](/ai-2-0/release-2.0/breaking-changes/applications-and-automation) for the
platform-level MCP tool contract changes (new tools, changed tools, pagination and payload
contracts) that also apply to every MCP client.
## Natural-language query configuration moves
**Breaking.** NLQ provider configuration moves into the unified `llmConfiguration` block — see
[Platform & Security](/ai-2-0/release-2.0/breaking-changes/platform-and-security) for the full shape. NLQ
also gains a client abstraction with OpenAI and Google Gemini providers alongside the existing one.
## The Collate desktop app refuses to save pre-2.0 instance URLs
**Breaking.** Saving a Collate instance URL in desktop Settings probes `GET /api/v1/system/version`
and, if the version parses below **2.0.0**, reopens the dialog with an inline alert instead of
saving. It fails open on an unparseable version.
Upgrade your Collate instances before rolling out the 2.0 desktop client, or users will be unable to
point the app at them.
## The Slack app moves from assistant mode to agent mode
**Breaking.** The Collate AI Slack app moves from assistant mode to **agent mode**, and now requires
**file-attachment permissions** so it can deliver AI Analytics reports.
Reinstall the Slack app from the updated manifest and re-approve the requested scopes. Without the
file permission, report delivery to Slack fails.
Response language is also now taken from the `X-OpenMetadata-Language` header, set from the
**Language** tab of profile settings, instead of being inferred from the conversation
(**behavioural**) — a user whose profile language differs from the language they type in will get
replies in their profile language.
## Full detail
For the full narrative walkthrough of Collate AI & Credits in Collate 2.0 — including screenshots and step-by-step context — see [Release 2.0: Collate AI & Credits](/ai-2-0/release-2.0/collate-ai-and-credits).