Changes that affect API clients are listed here, newest first. Additions that do not change existing behaviour (a new optional field, a new endpoint) are listed too, so you can see what became available.
2026-11-21
- New endpoint
POST /api/v3/embeddings: embedding vectors for a batch of texts (modelssmallandlarge, optionaldimensions), with the new scopeayeto.embeddings. Existing API keys with all scopes (*) can use it right away; a key with selected scopes needs the new scope added. - New endpoint
POST /api/v3/data-loader/load-batch: converts several files to text in one request, in parallel, with a result or an error per file. A minimum price is applied once per batch.
2026-11-18
- The progress text of a tool call can no longer close its own block: a line of the text that
reads
***gets a zero width space (U+200B) in front, incontentand in thetool_call_outputevents of the stream (Tool calls in the answer). Until now such a line ended the block early for clients that parse the markers.
2026-11-05
- When a chat stream ends early, the results of tool calls that already finished stay in the conversation, so the next message does not run them again. Only calls that did not finish are reported to the model as interrupted.
- A model call interrupted by a disconnect is charged with the part of the prompt that the previous call of the conversation already sent at the cached price.
- Conversations with tool calls read more of their history from the prompt cache after a new message, which lowers the input cost of long conversations.
- Conversations whose older messages were automatically compacted keep using the prompt cache for their whole history, which makes their model calls much cheaper.
2026-10-19
- Workflow exports list the workspaces the steps name with their names (
workflow.workspaces); import and copy takeworkspace_mapto use other workspaces instead (Export and import). - Workflows work with workspaces: new node types
workspace.read,workspace.list,workspace.search,workspace.write,workspace.save_fileandworkspace.manage, and the triggertrigger.workspacethat starts the published workflow on changes in a workspace folder (Workspaces). Deployment triggers carryworkspace_idandfolder. assistant.runtakesworkspaces: workspaces the assistant may use in that step only.- Workflows keep values between runs: node types
state.get/state.setand newPOST /api/v3/workflow/state/list,/workflow/state/setand/workflow/state/delete(Stored values). - New run field
not_before: a queued run of a workspace trigger waits until then. - Shares carry
shared_by(user id) andshared_at(timestamp): who added the share and when. The server sets them; values sent insharing/group_sharingare ignored (workflow object). - A user with read-only access to a shared workflow sees only their own share and the groups they are in, not the other recipients.
- Users can decline a share, block users and stop accepting shares in AYETO. A recipient who does not accept your share is left out when you create or update the sharing; the returned object shows what was stored.
- New read-only field
sharing_declinedon workflows: users who declined the group sharing.
2026-10-14
- Text to speech has new models and voices:
gpt-4o-mini-tts(recommended) with 13 voices, among them the newmarinandcedar, and a speaking style set by instructions.tts/mp3takes the new optional fieldsmodel,styleandinstructions;voiceaccepts every voice of the chosen model (Text to speech). - New
POST /api/v3/tts/audio: the speech asmp3,opus,aac,flacorwav. - New
POST /api/v3/tts/voices: the models, their voices, the speaking style presets and the length limit. - The default voice is
marin(wasonyx) where the installation usesgpt-4o-mini-tts. - A request converts at most 60,000 characters (about an hour of audio); a longer
textgets422. An emptytextgets422 The text is empty(was500 No audio generated), an unknown model or voice422with the reason. - The text is spoken in parts of whole paragraphs instead of sentence by sentence, and the balance
is checked against the estimated price of the whole text before the speech is generated.
gpt-4o-mini-ttsis charged by tokens, the older models by characters.
2026-10-08
- The chat stream with
runner_version"2"repeats thetool_call_pendingstatus while the model writes the arguments of a tool call, about once a second, withdata.arguments_chars(characters written so far). Long tool calls show progress instead of a single event. - A model call interrupted because the client closed the connection is charged for the part generated until then.
2026-10-07
- New workflow steps
booster.lockandbooster.unlock: the locks of booster database records and collections that panels use, with a bounded wait and an expiry. - Workflow steps that read files accept files of booster panels the user can use (the panel's file reference or id), also in triggered runs.
- Booster panels with AYETO sign-in can call the API on behalf of the signed-in person: the
session token works as
uni-api-keywith the scopes the person agreed to (Authentication). - Booster file storage: large files can be uploaded directly to the storage service
(
upload/init,upload/complete), where the installation offers it. - New workflow trigger
trigger.booster_file: starts the published workflow when a file of a booster panel storage is stored, changed or deleted; the file isinput.file. Workflow triggers list it with the panel and the watched folder incollection. - New: booster file storage. Panels can keep files in folders with access
rules;
/booster/storage/private/*uploads (multipart or base64), lists, links, moves and deletes them with the new scopesayeto.booster.storageandayeto.booster.storage.write. - Files served by AYETO links carry
X-Content-Type-Options: nosniff; SVG, HTML and XML files are always downloaded, never shown in the browser (<img>keeps showing SVG images).
2026-10-06
- The model catalogue (
/ai_model/get_all) lists only models you can use: models switched off (is_enabledfalse) or retired (is_deprecatedtrue) are no longer returned. - The model catalogue (
/ai_model/get_all) has a new fieldis_recommended: models the administrators recommend. - The chat stream with
runner_version"2"has a new status phaseconversation_named(data.name): a new conversation is named from its first message while the first answer is generated, instead of after it. The name is stored when the event arrives. findandcountrejectfiltersandsort_keyon internal fields stored with a record (for example secrets) with422; before, such conditions were accepted.
2026-10-05
- Conversations have two new fields,
booster_panel_idandpanel_user_id: conversations of the signed-in users of your booster panels with your assistants./conversation/findreturns them with your other conversations (filter them out with["panel_user_id", "==", null]);/chatanswers403for them, only the panel user continues them. - Booster database records have a new field
owner(UUID ornull): the panel user who created the record, on panels with user accounts.findandcountacceptownerin filters and assort_key. Records created through the API haveowner: null. When a panel user account is deleted for good, the records it created are deleted too. - Collection rules have two new levels,
userandowner. Like the other levels exceptnone, they allow requests made through the API, on all records. POST /ai_model/get_allcan return models of the new typegoogle_video_gen(video generation). Videos generated in chat are listed inattachmentsof the answer, like generated images.- In chat over the API, tool actions that the AYETO app lets the user confirm in the chat (for example creating a booster panel or a workflow from the chat) are refused: the tool call ends with an error and the model continues its answer.
POST /ai_tool/get_alllists two more tools,BoosterStudioStartFunctionTool(creates a booster panel from a template) andWorkflowStudioStartFunctionTool(creates a workflow). You can give them to your assistants; they need the user's confirmation in the chat of the AYETO app, so in chat over the API they are refused as described above.- Newly documented:
POST /booster/panel/info, which returns a panel's name, its organization and the key's user, and checks access to the panel. The endpoint itself is unchanged.
2026-10-04
- Tools returned by
POST /ai_tool/get_allhave a new fieldcompanion_of_groups(array of group names): likecompanion_of_group, for a helper tool that accompanies several groups.
2026-10-03
- Tools returned by
POST /ai_tool/get_allhave two new fields:vision(the tool shows images to the model and is offered only with models that can see images) andoutbound(the tool can send data outside AYETO, for example web requests or e-mail). - Malformed list queries are rejected with
422and a cleardetailinstead of500: a condition without 3 or 4 elements, afilterselement that is neither an array nor an object, a group object with no key or several keys, anAND/ORgroup whose value is not a non-empty array, groups nested deeper than 32 levels, a negativelimit_from, and alimit_tolower thanlimit_from. 429responses of the rate limiter have the body{"detail": "Too many requests"}like every other error (thereasonkey is gone);Retry-Afteris unchanged.- A missing API key scope is answered with
401API key is invalideverywhere. .../getand.../deleteendpoints also accept the id in a JSON body{"id": "<uuid>"}; theentity_idquery parameter still works. Neither, or two different ids, gives422. A body sent to them must be JSON.- New
POST /user_credit/get_selfandPOST /version(GET still works and shares the rate limit);/versionis rate limited on the static tier (1,200 requests a minute). POST /ai_tool/get_allno longer returns hidden or deprecated tools.- New
POST /conversation/count(scopeayeto.conversation, same body as/conversation/find). - Conversation cost (
POST /usage/cost/conversation) is the sum of the credits actually charged, not a repricing at current prices; it includes the text tokens of image models and no longer fails with404for models removed since. - Chat with a model that cannot stream stores the answer and returns it (before, the answer was lost and the response returned the user's message).
- Chat attachments are checked and stored before anything else: an attachment refused
with
422(file is too large, a storage quota, invalid data) no longer leaves the user message behind. Malformed base64 returns422invalid base64 data in attachment '<filename>'(orinvalid data URI in attachment '<filename>') instead of500. Thesizefield of attachments is no longer required and is ignored. - A chat follow-up to a conversation created in an organization runs in that
organization when the request names none (charged there, membership required);
naming a different organization returns
422the conversation belongs to another organization. Another user's conversation is refused with403before any attachment is stored. - The structured chat stream (
runner_version"2") no longer sends the unusediterationfield. - Booster database: creating a record with a key that already exists returns
422a record with key '<key>' already existsinstead of500;limit_fromwithoutlimit_toreturns up to 100 records fromlimit_from(limit_frommust be lower than1000); malformedAND/ORfilters return422instead of500. - Data loader: unsupported file types return
422unsupported file type '<mime>'and are not charged. - Workflows: a file named by its id in the run input stays the user's and is no longer deleted with the run (files sent as base64 still belong to the run).
2026-10-02
- Booster database collection rules apply only to panels in public database mode; in private mode they are ignored. Users a panel is shared with read-only can now write its records and take locks (in both modes).
POST /tts/mp3requires the newayeto.ttsscope andPOST /data-loader/loadthe newayeto.data_loaderscope. Existing API keys were given both scopes; keys created by pairing a desktop client (ayeto.connectoronly) can no longer call these endpoints.POST /tts/mp3andPOST /data-loader/loadaccept an optionalorganization_id: the request runs in the organization and is charged to the user's credit there (admins and members only; guests and non-members get403).POST /tts/mp3returns422withnot enough user credit/not enough organization creditwhen the credit runs out, instead of500Failed to generate audio.- Chat refuses a model switched off by the administrators (requested directly or as the
assistant's model) with
422model is disabled, and an automatic model with no available model with422model is not available;model is deprecatedis returned before the conversation is touched. - Reading another user's conversation with
/conversation/get(refused with404) no longer updates itsaccessedfield. - The
languageheader of an API request no longer changes the user's preferred language in the AYETO app (before, every request saved it, and a request without the header reset it to English). POST /assistant/avatar/getaccepts API keys with the newayeto.assistantscope, which can now be selected when creating a key (it also covers/assistant/find).- New documentation of API v3, replacing the earlier beta notes.
- Workflows are available on the API: reading, running, streaming runs, approvals,
editing, publishing, export and import (
ayeto.workflow,ayeto.workflow.write). - Private booster database API with its own scopes (
ayeto.booster.database,ayeto.booster.database.write).