Summary
Summary
Kaeso 0.7.0 adds decision models in three API formats, a rerank endpoint, guardrails, webhooks, SCIM 2.0 provisioning, key pools, batches and a Helm chart, new built-in providers from Amazon Bedrock and Azure to TypeSafe, and brings the console, documentation and UI library into the Kaeso Vanguard design. Updates from the console or the installer now check the install's settings before switching anything and stop before changing files when the new gateway would refuse one. Three settings that 0.6.0 accepted are refused at startup, so check kpmr.interval, mcp_servers.sync_interval and smtp.user before you upgrade.
- 01
- 02
- 03
Added62
Added
- 01#909
Asking Kaeso about usage in a client that supports MCP Apps opens an interactive view with the period's totals, a chart and the top models, providers or keys, through the new
show_usage_statstool; clients without MCP Apps get a one-line summary plus the same data asstructuredContent. - 02#910
MCP servers that ship MCP Apps views show them when a client reaches them through a project's MCP portal: the portal answers
resources/list,resources/templates/listandresources/read, and reading a view needs the same right as calling a tool that names it. - 03#1244
An identity provider can add and remove an organization's members through SCIM 2.0 at
/platform/v1/orgs/{org}/scim/v2, with its own bearer token per organization created under Security › Provisioning; removing a member at the identity provider removes them as an admin would, and only accounts at the organization's verified domains are touched. - 04
- 05
- 06#1331
Projects see when their models shut down, who still calls them and what replaces them: replies from models with announced dates carry
Deprecation,SunsetandLinkheaders, a request naming a retired model fails with 410, and aliases can move to the replacement in one click or on the shutdown day, both off by default. - 07#1353
A project can add model lines such as
claude-opus-latestthat follow the provider's newest stable model with a policy offollow(the default),ask,delayorpin; a move never happens when the project has no key for the provider or when a caller that may use the line may not call the new target. - 08#1347
Usage, a member's stats, a project's budgets and its summary can be shared through read-only links that people without a Kaeso account can open; each link is live or a snapshot, can hide costs and member, key and project names, expires after 1, 7, 30 or 90 days or never, and is checked against its maker's current rights each time it is opened, while request logs are never shareable; the rights
usage.shareandusage.share.managegovern them. - 09
- 10#1349
Kaeso emails 7 days and 1 day before an organization's platform key expires, to members who hold
platform_keys.createand to the key's creator while they are still a member. - 11#1350
Test key on Providers asks the provider whether it takes a key by reading its model listing, without billing a request;
POST /orgs/{org}/credentials/{credential}/checkanswersok,refused,unreachableorunknown. - 12
- 13#1233#1329#1341
Analytics breakdown tables sort by any number and download as CSV, and a project's rows open the requests they count; project and account API keys can be searched and sorted, project members searched and filtered by role, and an organization's projects sorted by their requests or cost in the last 30 days.
- 14
- 15#1317
A member's or key's limit can be copied to someone else with Copy to…, and limits sort by how much is used.
- 16
- 17
- 18
- 19
- 20
- 21
Gateway
22–49- 22#969
Apps make decisions through Kaeso at
/v1/decisions,/v1/systemoneand/v1beta/decisionson OpenAI, Google, OpenRouter, Vercel, Baseten, Fireworks and Perplexity models, with Kaeso's keys, limits, budgets, logs, costs, alias routes and response cache; a new settingbody_limits.decisions(default 64 MiB) bounds their bodies, and chat, token counts and embeddings on a decision-only model are refused with a message that names the decision endpoints. - 23#970#974
TypeSafe is a built-in provider: add its key and call Jev from any decision format, with images and box questions, which are refused before sending on Jev 1.13, and Cloudflare's Clef decision models (
@cf/cloudflare/clef,clef-flashandclef-omni) answer decisions in any decision format through the Cloudflare provider with the same key it uses for chat. - 24#972
Custom providers serve decisions in System One's or OpenAI's format once an org admin turns them on, with the path and, for System One, the image placement their server takes (
assets,state,rootornone); detection never turns decisions on, and a decision model can be added only while its decisions are on. - 25#959#966
Standard Compute and Perplexity are built-in providers: Standard Compute calls its smart routing or a pinned model from any app format, with routers listed first and marked Router in Add models, and Perplexity calls its Sonar models, with search fields such as
search_modepassing for Chat Completions apps only. - 26
- 27#1108#1116#1363
Volcengine Ark, StepFun and Baidu Qianfan are built-in providers, with a new provider-level API root so Ark's
/api/v3and Qianfan's/v2work like every other provider, Qianfan's embeddings, and StepFun's China service, Qianfan's international service and BytePlus ModelArk as base URL overrides. - 28#1174#1246#1247
Amazon Bedrock is a built-in provider: a Bedrock API key with its AWS region calls Claude models through the Messages route and other models through Chat Completions, requests can be signed with AWS Signature Version 4 from an access key or the gateway's own AWS role via
aws.ambient_credentials(instance admins only), and models that answer only Converse, such as Amazon Nova and Meta Llama, are reached through it. - 29#1187#1254
Azure is a built-in provider: add the key of an Azure OpenAI or Microsoft Foundry resource with its endpoint and call the resource's deployments from any app format, a key can instead be a Microsoft Entra ID service principal for resources that turn key access off, and a per-model Provider model ID sets the id Kaeso sends to the provider.
- 30
- 31
- 32#1356#1357
fal and Replicate are built-in providers for media: add a key (a token for Replicate) and make images with FLUX and Seedream and videos with Veo 3.1 and Kling from OpenAI's image and video formats, with fal also mapping Wan 3.0 and Replicate Nano Banana Pro; only the models they map by hand are accepted, and other models are refused with the list.
- 33#1361#1362#1366
Deepgram, Cartesia and AssemblyAI are built-in providers for speech from OpenAI's audio format: Deepgram transcribes with Nova-3 and Whisper and speaks with Aura voices, Cartesia speaks with Sonic and transcribes with Ink Whisper, and AssemblyAI transcribes with Universal-3.6 Pro through its Sync API; fields a provider has no place for are refused before sending.
- 34#961
The docs, the UI library's site and the console link a legal notice and a privacy policy set with
legal.imprint_urlandlegal.privacy_url, each anhttpsaddress; the gateway refuses to start with anything else, and an install that sets neither shows no legal links. - 35#963
Kaeso reads request bodies sent with
Content-Encoding: gzip,x-gziporzstd, as TypeSafe's SDKs send them, on/v1,/v1betaand/v2; other encodings such asbrordeflate, and stacked ones, get 415 in the app's format, a body that decodes past the route's body limit gets 413, and a corrupt one 400. - 36#1188#1189#1318
Model API replies carry the provider's request id in
kaeso-upstream-request-id, non-streamed replies carry what they cost inkaeso-costwhen the key's owner holdscosts.view, and every answer, streams included, carrieskaeso-limits, an RFC 9651 list of what is left of the key's rate limits and hard budgets, when the key's owner may see the project's limits. - 37#1215#1218#1230
A request can shorten how long Kaeso waits for each provider with
kaeso-timeout(1 to 3600 seconds, never raising a wait), keep its content out of the request log withkaeso-log-content: off, and tag itself with up to 10kaeso-tagsthat the request log filters by; browser apps may send all three, and another value answers 400. - 38#1220
cors.allowed_headerslists header names that preflights from listed origins also allow, so a browser app can send a header an alias route splits on, andserver-timingis exposed to browser apps;*,cookie,forwarded,x-forwarded-…and names that aren't header names stop the gateway at startup. - 39#1308
The response cache honours
Cache-Control: max-age, taking a stored reply only when it is at most that many seconds old, andonly-if-cached, which answers from the cache or with 504 without calling the provider. - 40#1207#1208#1209#1210
/metricsgainskaeso_update_available, the KPMR fetch and attempt timestamps includingkaeso_kpmr_last_attempt_failed,kaeso_connections_open,kaeso_connections_max,kaeso_model_requests_in_flight, andkaeso_tool_calls_totalwithkaeso_tool_call_duration_secondsfor MCP portal tool calls. - 41#1235
With
otel.endpointset, each model request is exported as an OpenTelemetry trace over OTLP/HTTP JSON: akaeso.requestserver span that continues the app'straceparent, a client span per route step that called a provider, and MCP portal tool calls asexecute_toolspans;otel.contentadds the prompt and reply where the request log keeps content, and spans never delay or fail a request. - 42#1248
The request log stores each long message an agent conversation repeats once per project (ClickHouse migration
0017_content_parts), so a session's stored content grows with its turns instead of their square, and every reader gets the body back byte for byte. - 43#1236
A route step tries its model again after a dropped connection, a 5xx or an overloaded provider, never after a timeout, a rate limit or a refused request; steps take
retriesfrom 0 to 3 and projects a Retries setting, the request log lists each try made again as retried, andkaeso-routeaddstries=N. - 44#1237
Organizations send operational notices, such as budget alerts, limits reached, expiring keys and MCP server changes, to up to 10 webhooks signed as Standard Webhooks specifies, as Kaeso's JSON event or a Slack line; deliveries are at least once and tried again after 1, 5, 15 and 60 minutes, a receiver that fails 20 times in a row is turned off, and the rights
webhooks.viewandwebhooks.managegovern them. - 45#1238
Organizations register guardrails, HTTP services of their own that Kaeso asks before a request goes to a model and, if chosen, before the reply goes back; a guardrail can allow, block or replace messages, requests are signed like Standard Webhooks, a block answers
content_filteredwith status 400 in the app's format, and only chat requests are checked. - 46#1240
Built-in guardrail rules find secrets and personal data inside Kaeso, such as provider keys, tokens, email addresses, IBANs and card numbers, plus up to 20 custom patterns, and mask them, block the request or only count them; Mask in the request log keeps the values out of the log, and Restore in the reply puts masked values back in the model's answer.
- 47#1241
A project can spread its requests over several keys of one provider: assigning a key adds it to the project's pool with a weight from 1 to 100, a caller stays on one key, and a key gives way to the next when the failure is the key's own, such as a refused key, billing or a rate limit; video jobs, image generations, speech, realtime and ElevenLabs still use the primary key.
- 48#1242
Responses apps can store and continue responses at OpenAI where the organization and project allow it;
store: trueandprevious_response_idare refused everywhere else with a message that says why,conversationandbackground: truestay refused, andGETandDELETE /v1/responses/{response_id}work for stored responses. - 49#1243
Apps can run OpenAI-format batches through Kaeso at providers whose batch API takes them as they are, today OpenAI;
/v1/filesand/v1/batchescheck each line against the key and model access before anything reaches the provider, a finished batch is charged at the batch price, 50 percent for OpenAI, with one usage and request log row per line.
KPP
50- 50
Console
51–58- 51
- 52#1222#1224#1225#1231
The request log filters by alias, fallback, endpoint, minimum cost, the model or alias name the app sent (
requested_model), outcome, response cache, duration and error message;GET …/logs/errorslists the most common failures, which the console shows above the list when Failed is on, and an hour under Slowest hours opens the slow requests. - 53
- 54#1312
The request log's
cost_basissaysnonefor requests no provider billed andestimatedwhen Kaeso estimated the tokens; entries logged before still sayreported. - 55
- 56#1227#1228#1229
The log inspector links to the model, alias, key and sender it names and filters the list by each; a logged request in the Anthropic, Chat Completions, Responses or Gemini format opens in the Playground with its settings and last message, a catalogue model opens there from its sheet, and key, member, service account and model sheets open their requests and audit history.
- 57#1006
The Kaeso plugin asks for the install's MCP address when it is installed in Claude Code, and for Claude on the web, desktop and mobile and for Cowork, Account › Agents has an Add to Claude button, and the plugin's skill a link, that opens Claude's custom connector form with the install's MCP address filled in.
- 58#1004
Decision answers line their bars up under each question without repeating the answer, and the request log lists each question with what its answers mean.
UI library
59- 59#976
The UI library gains
ProbabilityBar,DecisionResponseandQuestionEditor.
Self-hosting
60–62- 60#1232
Kaeso installs on Kubernetes with a Helm chart that each release publishes; it runs the gateway and console hardened as Compose does, migrates once in a Helm hook, and offers an Ingress or HTTPRoute, a PodDisruptionBudget, an optional ServiceMonitor and bundled single stores for trying it out, and the new setting
updates.installed_withmakes the console's Updates page show the Helm update command on such an install. - 61#1205
kaeso healthcheck --readyasks/readyzfrom inside the gateway's container, prints its answer and exits 0 only on a 200, for platforms whose readiness check runs a command; plainhealthcheckstays on/healthz. - 62
Changed12
Changed
- 01#1217
An install takes less memory, CPU and disk: ClickHouse writes only the diagnostics an operator reads, request bodies are compressed in 16–32 MiB blocks so agent sessions are stored several times smaller, and the gateway allocates with mimalloc without transparent huge pages.
- 02#1252
Kaeso Vanguard brings the console, docs, UI library and KPMR into the design of the redesigned kaeso.app, with cut plates, hairline structure and flat surfaces; no prop, export or variant key was removed or renamed.
- 03
- 04
Gateway
05–08- 05#1120#1121
The gateway stops at startup when
kpmr.intervalis below 5 minutes, whenmcp_servers.sync_intervalis below a minute, or whensmtp.useris set withoutsmtp.password, where 0.6.0 accepted the first two and sent email without authentication; an empty password counts as unset, and an emptyKAESO_CONFIGnow counts as unset instead of stopping the gateway. - 06#1216
A request after a pause no longer waits for the database or an account renewal: caches stay warm while they are used, ChatGPT and Grok accounts are renewed ahead of time by a new
account_renewaljob, bodies are checked and decoded in fewer passes, and a fallback'soverhead_usnow leaves out the failed steps' wait for their providers. - 07#930
@modelcontextprotocol/ext-appsis updated from 1.7.5 to 2.0.4 for the usage view the gateway serves. - 08#1393
A fresh install starts with KPMR's model data as of 20 November 2026, 08:44 UTC: 2,934 models from the 50 providers Kaeso calls, with no model removed or newly deprecated.
KPP
09- 09#1249
Large OpenAI Responses requests, such as a coding agent's resent conversation, decode in about three quarters of the time; a 3 MB body decoded in 10.2–10.7 ms against 13.2–14.2 ms before.
UI library
10- 10#991
astrois updated to 7.4.7.
Self-hosting
11–12- 11
- 12#1407
An update from the console or the update window that changes
compose.yaml, as every update from 0.6.0 does, first checks that the new gateway takes the install's settings; when it would refuse one at startup, the update stops before switching anything, the running release keeps serving,compose.yamlis put back and the update's detail names the setting, and the installer's--updatedoes the same check.
Fixed64
Fixed
- 01#931
Pages that render Kaeso UI on a server hydrate without an error when a table shows dates or relative times, a chart highlights a day, or a list starts long or with rows selected.
- 02#960
UI library formatters follow each page's language on a server through new
useFormatters()andcreateFormatters;StatusCodeshows a focus ring, wrappedCodeBlocklines break only at spaces, and dark-theme provider marks stay legible. - 03#978
TypeSafe's image refusals (wrong format, too small, too large) are one sentence each, so the Playground shows their reason.
- 04#1010
A lazily loaded component asks for its code only at a fresh address after a download fails in Safari or Firefox.
- 05
- 06#1282#1283#1289#1296
Emails work in more cases: a failed run of the monthly cost report or weekly summary is retried a minute later instead of skipping the month or week, with several gateways every emailed setup link works until setup is done, budgets with the period
totalsend their alert emails, and unsubscribe links in emails sent before a master-key rotation keep working while the old key is configured. - 07#1278
Reading one of your own keys needs only
account.keys.view, as listing them does. - 08#1343
SSO group rules match Microsoft Entra ID directory roles for users whose groups are read from Microsoft Graph.
- 09#1421
A finished video or image generation shows the tags, client app, alias, model line and paying key of the request that started it in the request log and usage, so filtering by them counts its cost.
Gateway
10–31- 10#971
Gemini apps reach Kaeso under
/v1as well as/v1beta, so google-genai 3 works with its defaults, andGET /v1/models,GET /v1/models/{model}andPOST /v1/decisionsanswer in Gemini's format when the request carriesx-goog-api-keyor?key=and noAuthorizationheader. - 11#1113
Browser apps can send
Cache-Controlto the response cache; CORS preflights allow it. - 12#1206
Web searches that Anthropic and OpenAI bill per call are part of a request's cost from a
searchunit price in the catalogue; searches without a price log the cost as partly unknown. - 13
- 14#1251
A budget counter rebuilt while a request finishes counts that request exactly once.
- 15
- 16#1258#1383#1385
Thinking and sampling settings match what models accept: DeepSeek models turn thinking off when an app turns it off, Claude Opus 4.7 and later leave out
top_pandtop_kas they do a temperature, and a thinking budget reaches Claude models that refuse budgets as adaptive thinking, each change named inkaeso-adaptedorkaeso-left-out. - 17
- 18
- 19
- 20#1267
Request logs are written again after a ClickHouse migration instead of failing until the gateway restarts, and
/readyzreports degraded while writes fail. - 21
- 22
- 23#1274
Audit log, member, user and usage filters with a NUL character answer 400 naming the filter instead of 500 or 503.
- 24
- 25#1280
A ChatGPT account at its plan's usage limit keeps every model of the account out of routes for at least 5 minutes, or as long as the provider says, up to an hour.
- 26#1288
The usage timeline counts a finished video or ElevenLabs image once and without a 0 ms request.
- 27#1291
Groq json and text transcriptions are priced by Groq's own length, and audio sent as a link shows an unknown cost instead of $0.
- 28#1387#1409
A request body over the size limit gets Kaeso's 413 in the app's format behind a proxy such as Traefik instead of the proxy's 502, and so does an answer Kaeso sends before it has read a large body, such as for an invalid key, a wrong method or a body over the MCP limit; HTTP/2 is left alone, and a client that sends the body without waiting for
100 Continuecan still get the proxy's 502. - 29#1000
The request log shows logged JSON with its keys in the order the app sent them, so a decision's answers sit under their own questions and options.
- 30#1122
The README names every built-in provider, and a test keeps the list in step.
- 31#1257
kaeso instance admin-ssoandkaeso sso verify-domainsend the email the same change sends from the console.
KPP
32–37- 32#994
Gemini no longer bills cached image and audio tokens again as uncached input.
- 33#1112
A Gemini token count with a field another format lacks is refused under the name the app sent.
- 34#1123
OpenAI Videos' checked decoder refuses the request kinds its plain decoder refuses.
- 35#1298
A DashScope or Gemini Interactions video job with an out-of-range time reads it as no time instead of panicking.
- 36#1265
An image's
detailreaches the other OpenAI format: Chat'simage_url.detailbecomes Responses'detailand back. - 37#1287
System One refuses a score level of null as it comes in instead of accepting a request it can't send, and OpenAI's Decisions API confidence follows TypeSafe's choice and score formulas.
Console
38–54- 38#1221
The request log's format filters list every format Kaeso logs, from the list they accept.
- 39#1114
The Playground keeps models whose id starts with
~or@, such as OpenRouter's~…-latestand Cloudflare's@cf/…models. - 40#1256
Clearing a project's response cache shows its success message instead of reporting an empty answer.
- 41#1271
The month forecast takes the last seven days' pace in the month's first week too.
- 42#1299
Analytics' links to the request log open the UTC days Analytics counted.
- 43#1297
Updating to a release that is no longer the newest shows the newer one on Instance › Updates instead of asking for a check the console can't do.
- 44#1259
The key expiry picker offers no day whose start in UTC has passed, and a refused date names
expires_on. - 45#1367
Guardrails has its own icon in the settings navigation instead of Security's.
- 46#1368
Organization pages and settings fit their content at every width and language: the SCIM address shows once, the agent preset fits a phone, webhook URLs and MCP handles end in an ellipsis, and role descriptions get a text area.
- 47#1413#1415#1417
Owners and admins who reach a project through their organization role, not as members, can choose which models their keys may call from a key's sheet, project pages name people from the organization who aren't members of the project instead of showing their user id, and a tool's sheet no longer sets its MCP server's name in a code font.
- 48
- 49#1401
Analytics fits phones and tablets: charts answer a tap, grouping is a menu on a phone, Requests stays beside the cost, and the custom period's dates show whole.
- 50
- 51
- 52
- 53#1410
In the request logs, Tab always reaches the filters after the Models and Tools switch, and switching with the arrow keys keeps the focus on the switch.
- 54
UI library
55–58- 55#1406
Stat tiles that load their numbers show them at full size on phones instead of the smallest size.
- 56#1408
Tables whose rows have a right-click menu take one Tab stop instead of one per row, and screen readers no longer meet focus stops between the rows.
- 57#1412
On touch screens, the model and provider pickers answer a finger-sized tap around them, like every select.
- 58#1416
Long tables and lists such as the request logs stay where you scrolled when the next page of rows loads, instead of jumping back near the top.
Documentation
59–62- 59#1115#1119
The Platform API page and the tool-call route no longer say the Playground runs tools, the refusal for a caller without a person says who may call tools, and the configuration page says
updates.checkis read on every check and that background jobs can hold up to 11 connections outside the pool. - 60#1392
The documentation describes what is new in 0.7.0 and what to check before upgrading to it, and its tables fit a phone.
- 61#1338
The KPP docs record that OpenAI's Decisions API confidence follows TypeSafe's choice and score formulas.
- 62
Self-hosting
63–64- 63#1266
With
--domain :80the installer writesKAESO_PUBLIC_URLandKAESO_SMTP_FROMfor a port, from the new--public-urlor, without it, a local address with the port, so the install it writes starts, and it repairs an install it wrote before. - 64#1420
An installer update that the new release refuses leaves
compose.yaml.previousas it was, instead of a copy of the runningcompose.yaml.
Security2
Security
- 01#920
A key with platform access can no longer widen its own model list, Tools list or limits, or leave another key looser than itself; such changes are refused with 403
permission_deniedand need a console session.
Gateway
02- 02#1346
Email goes out through lettre 0.11.24, which fixes six SMTP advisories, RUSTSEC-2026-0560 to RUSTSEC-2026-0565, among them replies a man-in-the-middle could inject after STARTTLS.































