Model fallback and load sharing. A Model can name other Models in fallback and extra servers in
additionalEndpoints. When a server answers 5xx or 429, the next server or Model answers the request. Several
servers of one Model share the requests. Access rules, budgets and the audit record follow the Model that
answered. See Fallback and several servers.
Smaller MCP tool results. An admin can set resultCompression on a Server (off, lossless or shorten).
Smaller tool results mean fewer input tokens in every later turn. Measured on 2026-11-02 with opencode 2.1.4 and
Codex 0.159.3 (model openai/gpt-4.1-mini, 3 runs): shorten used 49% fewer input tokens per session with opencode
and 31% fewer with Codex; lossless used 16% and 3% fewer. In 20 test questions, no answer was lost with
lossless, and shorten answered as well as no compression. It is off by default. See
Compressing tool results.
Savings page (Evidence → Savings, admins only). It shows what compression and model routing saved in the last
30 days, in estimated tokens and dollars, by day, group, server and router. It needs evidence.enabled. See
Savings.
Servers from an image, with each person's own sign-in. A Server that runs an image over HTTP can now have a
connection: each person signs in to the vendor with their own account, and Ambit passes that person's token to
the container through ToolHive's proxy. A Server that runs an image can also set env (values or Secret keys) and
args for its container.
Masking handles large tool results.ambit-guard accepts tool results up to 32 MiB (before: 4 MiB) and no
longer runs out of memory at about 12 MiB.
Two privacy fixes for Presidio users. (1) Setting Presidio no longer turns off the built-in masks (Korean
resident registration numbers, US Social Security numbers, card numbers, emails); the built-in masks run first
and Presidio adds to them (issue #189). (2) While Presidio is set, JSON tool results are not compressed, because
compressed JSON hid most names from Presidio: the share of names found fell from 93.2% to 10.0% (issue #207).
Fixes: the Zoom preset no longer fails the nightly contract check by mistake, and Codex no longer logs
failed to decode models response (the gateway answers its /v1/models request in the shape Codex reads).
Upgrade both charts, ambit-crds first. The CRDs of Model, Server and ServerRequest have new optional
fields: fallback and additionalEndpoints on a Model; resultCompression, env and args on a Server and in
the server part of a ServerRequest.
An older controller does not know them. Existing objects need no change.
The ambit-guard memory limit is now 512 MiB (before: 128 MiB), set by the new value
protection.memoryLimitMiB. The memory request stays 32 MiB, so scheduling does not change. A large tool result can
make the guard use up to 512 MiB on its node.
New chart values (all optional, with defaults): protection.memoryLimitMiB (512),
protection.presidio.timeout (10s), evidence.summaryRetention (26280h, three years),
console.savings.referenceInputPricePerMillion (empty). protection.presidio.anonymizerURL is no longer used; an
existing value is accepted.
With an external ClickHouse, the Savings page needs more rights for the ClickHouse user: CREATE VIEW,
ALTER (modify TTL) and SELECT on system.tables, in addition to what it had.
A Model with baseURL (a server you host) that lists additionalEndpoints must give each of those entries its
own baseURL. Ambit refuses the Model otherwise.
The Savings page counts from the day the summaries are created after the upgrade. There is no backfill of older
days. All its numbers are estimates, and routing savings do not subtract the classifier's cost (issue #216).
Compression changes the text of a tool result. Codex sends the model structuredContent, not the text, so for a
tool that returns both, compression does not help with Codex. opencode 2.x in its default mode sends neither.
ambit-guard refuses tool results larger than 32 MiB.
A Server that runs a stdio image cannot have a connection.
The nightly contract check of the DocuSign preset fails, because DocuSign's server does not ask an unauthenticated
request to sign in (issue #222). The preset itself passed the live check for this release.
The console no longer offers the Databricks and PagerDuty presets: their test accounts stopped working, so they
could not pass the live check before this release (issue #244). A Server you already made from either preset is
not changed.
Changelog entries are one file per pull request in changelog/unreleased/.
CI: the image is built in its own job, and e2e runs in more parallel parts; a full run takes about 17 minutes instead of 26.
Docs: Compressing tool results now has measured savings with opencode and Codex, and says which part of a tool result each client sends to the model: Codex sends structuredContent and not content[].text, so compression of the text does not help for a tool that returns both.
A Model's fallback answers a request for the Model's name only for people who may use both Models. A person who may use the fallback but not the Model gets 403 for the Model's name, whether the Model works or fails.
When a Model has a baseURL (a server you host), every entry in additionalEndpoints needs its own baseURL. Ambit refuses the Model otherwise, so no prompt can reach the provider's public server by mistake.
Guard: with evidence.enabled, ambit-guard writes one record to the audit trail for each tool result text it compresses (Server, type, bytes before and after, the caller's groups; no email, no content). The coming Savings page counts them. A new AmbitGuardRecordsDropped alert (with alerts.enabled) shows when records were dropped. No call is changed or slowed by this.
Savings page (Evidence → Savings, admins only): what tool result compression and model routing saved in the last 30 days, in estimated tokens and dollars, for each day, group, server and router. Every figure is labelled as an estimate. Compression dollars use each group's own average input price, else console.savings.referenceInputPricePerMillion, else tokens only. One control, "Later turns that carry each result" (default 0), shows what a result that is sent again would add; Ambit does not measure that number. Routing is measured against the most expensive priced model each router can choose, and does not subtract the classifier's cost. It counts from the day the summaries were created; there is no backfill. See Savings in the docs.
Evidence: with evidence.enabled, the console creates two daily summary tables in ClickHouse (ambit_llm_daily_v1 and ambit_compress_daily_v1) and a read-only admin API GET /api/savings that reads them. The summaries hold counts only (no email, no content) and are kept for the new value evidence.summaryRetention (default 26280h, three years). They count from the day they are created. With an external ClickHouse, the user also needs CREATE VIEW, ALTER (modify TTL) and SELECT on system.tables. The new value console.savings.referenceInputPricePerMillion sets the price used for groups without priced LLM requests. The page that shows this comes in the next release note.
Tool result compression: an admin can set resultCompression (mode: off, lossless or shorten) on a Server to make its tool results smaller before the client stores them, so each later turn costs fewer tokens. It works on JSON, CSV and tables, YAML, logs and test output; other text is not changed. lossless removes only layout. shorten also leaves out repeated log lines, passing tests, YAML comments and managedFields, marks each cut, and tells the model to call the tool again with a narrower query. It is off by default, runs before masking, and a fault in a compressor never fails a call. See Compressing tool results in Concepts.
Guard: ambit-guard serves Prometheus metrics (ambit_guard_compress_*) on port 8081, the chart adds a PodMonitor for them and an AmbitGuardCompressionFailing alert when alerts.enabled, and a panic in the guard now fails one call instead of ending the process (which made every MCP call fail until the pod restarted).
Masking: ambit-guard now accepts tool results larger than 4 MiB (the gRPC limit is 32 MiB). With the default memory limit, results up to 10 MiB were masked in a test; larger ones can make the guard run out of memory (see Masking in Concepts).
Masking with Presidio: a tool result is sent to the analyzer in a few large calls instead of two calls per string, and Ambit replaces the findings itself. The anonymizer is no longer called. The new value protection.presidio.timeout (default 10 s) limits all analyzer calls of one result.
Masking with Presidio: the built-in masks (Korean resident registration numbers, US Social Security numbers, card numbers, emails) now run first, and Presidio adds to them. Before, setting Presidio turned the built-in masks off.
A Server can now have a connection (each person's own sign-in) and an image, when the image talks HTTP (transport: streamable-http or sse). Before, connection needed a url. A stdio image with a connection is still refused, with a clearer message. The person's token reaches the container through ToolHive's proxy. The console's "request an app" form accepts the same combination.
A Server that runs an image can now set env (inline values, or a key of a Secret) and args for its container. Names that ToolHive sets are refused.
A Server that runs an image and has a connection can now have its tools listed: bind the Connection (serverURL) to the address of the server's ToolHive proxy (http://mcp-<name>-proxy.<namespace>.svc:8080/mcp). Before, such a server was refused by a bound Connection and its tools could not be read. The console is now allowed to reach the proxy of a server for this.
A Model can list other Models in fallback and extra servers in additionalEndpoints: when a server returns a 5xx or 429 response, the request is answered by the next server or Model, and several servers of one Model share the requests. Access rules, budgets and the audit record follow the Model that answered. See Model providers.
Masking with Presidio: while protection.presidio.analyzerURL is set, the guard no longer compresses JSON tool results (issue #207). Removing the spaces in JSON lowered the share of names Presidio finds from 93.2% to 10.0%, so names reached the client unmasked. JSON now keeps its original layout; logs, tables, YAML and test output are still compressed. The new outcome skipped_presidio of ambit_guard_compress_results_total counts these results.
Masking: ambit-guard now uses less memory for a large tool result, and the new value protection.memoryLimitMiB (default 512) sets its memory limit. Before, a result of about 12 MiB made the guard run out of memory and restart, and every MCP call failed while it restarted. Results up to the 32 MiB message limit are now masked (see Masking in Concepts).
Preset check: the nightly contract check accepts resource metadata that the server names on another host (Zoom does this), if the metadata's resource is the server's URL or a parent of it and the address is https. This fixes the false failure for the Zoom preset.
Codex: the gateway answers Codex's request for the model list (GET /v1/models?client_version=...) with {"models":[]}, the shape Codex reads. Codex no longer logs failed to decode models response. Other clients get the same OpenAI list as before.
Savings page: the group chart now shows each group's share of the total instead of dollars that looked larger than the total, and says that people in several groups count in each group. "Counted since" now shows the oldest day the summaries hold, so it never conflicts with the chart.
Usage page: reads the daily LLM summary table for the days it holds completely, instead of scanning every audit record. The numbers are the same; the page is much faster with a large audit trail. Days before the summary began still come from the audit records.
The console no longer offers the Databricks and PagerDuty presets in its app list. Their test accounts stopped working, so they could not pass the live check before this release. A Server you already made from either preset is not changed. They come back when they pass the live check again.