Changelog

Notable changes, newest first, since the first generally available release. This page shows the latest month in full, and every earlier month has a page of its own.

Each change is tagged with what it affects:

• API the REST API. The MCP server and the CLI call the API, so they get these changes too.
• MCP the MCP server, its tools, or connecting an MCP client.
• CLI the Intuizi CLI.
• Console the Intuizi console.

September 2026

2026-09-27

• Size estimates, activation preview, and the visualization catalog in the CLI CLI

  CLI v0.1.6 adds a command for each of the four API reads that had none. intuizi audiences estimate create sizes an audience from the same flags or --file body as audiences create without creating it, and intuizi audiences estimate show <id> reads the result, both with --wait. intuizi activations preview counts the devices a frequency range would deliver and prints the histogram and the filter_hash, and intuizi reference common datastream-visualizations lists the data stream visualizations an audience can draw. audiences create and audiences estimate create take --frequency to run the dataset type’s frequency analysis, and activations create takes --freq-min, --freq-max, and --filter-hash, so a previewed range can be delivered without a --file body. See Estimate create and Preview.

• Estimate Audience Size is listed as idempotent API

  Idempotency now lists Estimate Audience Size, which already read the Idempotency-Key header, so eight endpoints honor it. A retry with the same key and body replays the estimate rather than scanning the data again.

• A build that stops at 107 Additional Info now sends the failed webhook API

  An audience or activation that Intuizi stops at 107 Additional Info, because it cannot be built as defined, now sends audience.failed or activation.failed to every webhook endpoint subscribed to it, once, with the same payload as a 4xx failure. Read data.status.id to tell them apart. The payload carries only the status id and name, and Audience Manager shows the reason for a 107. Until now a build that stopped at 107 sent no event, and only a read showed it. See Webhook event payloads.

• A cancelled Lookalike Model now ends at 400 Error API Console

  A run cancelled with Cancel Lookalike used to read 108 Modeling for good once it stopped, so a poll or a --wait on it could only time out. It still stops at its next checkpoint, and then ends at 400 Error, which is final and sends an audience.failed webhook. Audience Manager shows it as Cancelled on request. A cancel that arrives once the result is already being published is still ignored, and the run completes at 104. A run that had already stopped on a cancel before this date keeps reading 108.

• A failed cohort import now reports 5 Not Available API

  A failed import used to end at an error code outside the cohort status scale, which Get Cohort named Unknown, and it sent no webhook. It now ends at 5 Not Available and sends cohort.failed, so a cohort goes 1 to 4 on success and ends at 5 on failure. A cohort that failed before this date keeps the code it failed with, still named Unknown.

• A scheduled Origin audience moves with each cycle API

  Every cycle of a schedule restamps the audience’s datasets to that cycle’s data window, but an Origin dataset kept the weeks it was built with, so a scheduled Origin audience rebuilt the same data every cycle. An Origin dataset now gets each cycle’s window too, widened to the whole Monday-to-Sunday weeks it touches, the same widening Create Audience applies. This applies from the next cycle of every schedule, including schedules created before this date. A cross purchase block still keeps the dates it was built with.

• A schedule’s end date cannot come before its start API Console

  Create Schedule accepted a recurrence.ending.end_date earlier than the date of recurrence.start and counted the gap forward: a daily schedule starting 2030-01-15 09:00:00 with an end date of 2030-01-10 was stored with 6 runs instead of being refused. It is now rejected with 422, naming recurrence.ending.end_date. The start’s own date is still accepted and gives one run. The Scheduler in Audience Manager applies the same rule, and the CLI already refused such an --end-date before sending it.

• A monthly Custom Date schedule counts calendar months API Console

  A monthly schedule runs one calendar month apart, but Create Schedule turned its Custom Date ending into a run count using 30-day periods, so near the end date the count was off by one run or more. Monthly from 2026-10-15 06:00:00 to 2027-10-15 stored 13 runs, the last on 2027-10-15 06:00, after the end date. recurrence.cycles.available now counts the calendar-month runs up to 00:00 on end_date: 12 in that example, the last on 2027-09-15. Schedules created in Audience Manager count the same way, and daily, weekly, and bi-weekly counts are unchanged. The count is set when a schedule is created, so a monthly Custom Date schedule created before this date keeps its stored count.

• A POI submission with remove archives exactly the POIs it does not list API

  remove was applied 250 locations at a time. On a longer submission, POIs listed after the first 250 were archived and then added again with new ids, and locations the submission had just added were archived. A submission in which no location matched archived nothing. The three submission creates, and intuizi poi submissions create --remove, now apply remove once, after every location has been matched: every POI the brand had that no location matches is archived, the matched POIs keep their ids, and the locations the submission adds are kept. A remove submission whose locations carry values for its key but match no POI now archives every POI the brand had (one that holds no location at all, or in which no location has a value for its key, archives nothing), so list every location the brand should keep. If an approval is interrupted part way, its locations are still imported but remove is not applied, and sending the submission again applies it. This applies to every submission approved from this date.

• The location-id key matches on location_id API

  A location’s location_id was dropped when a submission was read, so with key set to location-id the location was added a second time, and update or remove could act on an unrelated POI of the brand. A location now matches the brand’s POI whose id is its location_id. One with no location_id, or with one that is not the id of one of the brand’s POIs, is added as a new POI. A remove submission in which no location has a location_id names no POI to keep, so it archives nothing. A location_id must be the id of one of your POIs, and a file may write it as a whole number with or without .0 (101 or 101.0). A location-id submission created before this date was read without its location_id values, so if it is still Waiting, delete it and send it again. See the matching note.

• A null key on a POI submission means gps-coordinates API

  A key sent as null (or as an empty form field) was stored as null, and the approval then matched every location to the brand’s first POI: nothing was added, update rewrote that one POI, and remove archived all the others. The three submission creates now store it as gps-coordinates, the default when key is left out, and a submission stored with a null key is matched on gps-coordinates too.

• The MCP create_poi_submission tool is marked destructive MCP

  With remove, approving the submission archives every POI of the brand that no listed location matches, so the tool now declares destructiveHint: true, and MCP clients that honor the hint ask for confirmation before calling it. Its update, remove, and key fields are now described. See the tools reference.

2026-09-26

• A Custom Date schedule runs every cycle up to its end date API

  A schedule created with the Custom Date ending stopped after its first cycle: its cycle count was stored as a negative number, so it became Fulfilled as soon as the first run finished. Create Schedule now stores the intended count in recurrence.cycles.available: one cycle for recurrence.start plus one for every full frequency period (1, 7, 14, or 30 days) in the whole days between the start and 00:00 on end_date. A weekly schedule from 2026-10-15 06:00:00 to 2027-01-01 now runs 12 cycles instead of 1. The count is set when a schedule is created, so a Custom Date schedule created before this date keeps its stored count. To run one to its end date, recreate it with a recurrence.start in the future. Never and Recurrences endings are unchanged.

• Documentation for the Intuizi CLI CLI

  A new CLI section covers the Intuizi CLI v0.1.4, which runs API v2 from a terminal, a script, or a CI job. Every command that calls the API is built on the documented endpoints and returns their responses under the same rate limits. Names are looked up in the catalogs, and flags are checked before anything is created. The section covers installing and logging in, a quickstart that reads the catalogs then builds and delivers an audience, and a command reference listing every command and flag, with the endpoint behind each one that calls the API.

• Upgrading to CLI v0.1.4 CLI

  However v0.1.3 or earlier was installed, v0.1.4 changes the following. activations create takes --datastream, without which an activation built from flags delivers nothing. --debug prints every HTTP request and response on stderr. audiences create accepts --type origin, and a WebDomain --category name now sends the category’s numeric id, which Create Audience takes, rather than its IAB code, which it refused. --wait keeps waiting while an audience draws its data stream visualizations (109) and ends as a failure on 107 Additional Info, where earlier versions failed on 109 and waited on 107 until the timeout. A wait that fails, times out, or gives up prints the last record read on stdout, so branch on the exit code. The next auth login moves a token stored in the config file to the OS credential store, where there is one. auth status names the account for a token minted by v0.1.4, and reports it as unknown for one stored by an earlier version until auth login --email mints a new one. On a 429 the CLI says on stderr each time it waits and retries, returns a 429 that asks for more than 60 seconds at once rather than retrying into a second refusal, and names the Retry-After in the error. A cohorts create --upload-reference refused by the build budget is not retried, since the refusal already used the reference. With --json, --wait prints the server’s envelope on every outcome, a timeout and a give-up included, and a give-up names the show <id> --wait that resumes it. INTUIZI_NO_KEYRING=0 or false now leaves the credential store on. The npm launcher passes a SIGTERM sent to it on to the binary. New usage errors, caught before anything is sent: a schedule --end-date before its --start date, a schedule --name over 255 characters, and a schedule --timezone outside a region (US/Eastern, GMT, or an Etc/ name). So are an audiences create --type the flags cannot build, which --help names, a --contrast-audience-id that is the seed, reference web iab-subcategories --category-ids mixing ids and codes, and --project-id with cohorts create --audience-id. A WebDomain --category given an exact IAB code or name, such as IAB1, resolves instead of listing every code that contains it. poi submissions create --list notes on stderr each body value a flag replaces, and its --update=false and --remove=false now turn off a true in the body. schedules activate notes a schedule that has made every run its ending allows. See Conventions and Login.

2026-09-25

• Build budget on worker-bound creates API Console

  Every create that starts a worker job (audiences, estimates, Lookalike Models, cohorts, and activations) counts against one rolling budget per organization, 60 an hour and 300 a day by default, counted across the console, the API, and MCP. Scheduled replays do not count. Over the budget, the create is refused with 429 and a Retry-After header, and nothing is created. Get Usage returns a new build_budget object with the limits, the live counts, and retry_after_seconds. See Limits & Quotas.

• Create Activation enforces the data-scan limit API

  Create Activation now runs the same monthly data-scan check as Create Audience. A company that has reached its enforced limit is refused with 422, and no activation is created or queued. Companies without an enforced limit are unaffected.

2026-09-24

• Every frequency analysis on Create Audience API MCP

  The analyses object now takes each frequency toggle the Audience Manager offers: frequency (POI) with its frequency_day_part sub-option, apps_frequency (Apps) and web_frequency (WebDomain). Each needs its dataset type in datasets (422 naming the key otherwise) and the same feature on your account (403); day-part needs frequency and is not available next to an Apps dataset. See Frequency analyses. The MCP create_audience and estimate_audience_size tools follow.

• Create Audience can run the Frequency analysis API MCP

  Create Audience takes an optional analyses object; {"frequency": true} runs the Visitation Frequency analysis as the audience builds, exactly as the toggle in the Audience Manager does, so an audience created through the API can be previewed and filtered by visit frequency with Preview Activation. Until now the API had no way to enable it, and every API-built audience was refused by the preview. It requires a POI dataset (422 otherwise) and the Frequency feature on your account (403 otherwise - ask your Account Manager); an unknown key under analyses is a 422 naming it. See Frequency analysis. The MCP create_audience and estimate_audience_size tools follow.

2026-09-23

• Intuizi plugin for Claude Code and Cowork MCP

  The MCP server can now be installed with /plugin marketplace add intuizi/intuizi-claude-plugin followed by /plugin install intuizi@intuizi, instead of adding the server URL by hand. The plugin carries the server declaration and a skill covering the order the tools are meant to be called in; authentication is the same one-click browser approval. It is open source at github.com/intuizi/intuizi-claude-plugin. See Getting started.

• Create Audience enforces the data-scan limit API MCP

  Create Audience now runs the same monthly data-scan check the console does before it creates anything: a company that has reached its enforced limit receives a 422 naming the limit and the reset date, and no audience is created or queued. Companies without an enforced limit are unaffected. The MCP create_audience tool follows.

• Size estimates count as usage API MCP

  An Estimate Audience Size run scans the same data the create it stands in for would scan, so it now counts toward your company’s monthly data-scan limit and is refused with the same 422 as a create once that limit is reached. Usage reports it under a new sixth operation type, estimate (label Size Estimate), appended after activation; the other five keep their positions. Estimates run since 2026-08-24 are being recorded retroactively, in the month each completed. The MCP estimate_audience_size and get_usage tools follow.

2026-09-22

• Shorter datastream errors API MCP

  The error of a failed datastream on Get Activation, List Activations and the activation webhooks no longer includes storage locations or URLs, and an error passed through from the query engine or cloud storage is cut to its first sentence, which states the reason. Processing detail that followed it is no longer returned. The MCP get_activation and list_activations tools return the same text.

2026-09-21

• Origin dataset on the API API MCP

  Create Audience and Estimate Audience Size accept type: "Origin" - people by their resolved home location, where a device originates. Its only filters are geographic: location.countries is required, and location.states, location.cities, location.dmas and location.zipcodes are optional. signal_providers is required as for the other dated types - fetch them from Get Signal Providers with dataType=Origin. Origin data is weekly with no day in it, so any dates are accepted and the window is widened to the whole Monday-to-Sunday weeks it touches. Coverage is limited to a few countries: fetch them from Get Countries with datasetType=Origin, and a country outside that set is rejected with 422. Get Dataset Types lists the type when it is enabled for your account, and the MCP create_audience and estimate_audience_size tools accept it.

2026-09-17

• Reads name what an audience or cohort was built from API

  Get Audience returns source_audience, the seed a Lookalike Model was built from, and each Cohorts entry in dataset carries cohort, the cohort it reads. Get Cohort returns source (file, audience or pixel) and source_audience, the audience a cohort was built from. Each reference is { "id", "name" }, or null when there is none or the resource was deleted or belongs to another company. The list reads, the cohort create response and the audience.* / cohort.* webhooks carry the same fields.

• Changing your password revokes every token on the account API

  Login tokens from POST /api/v2/auth/login are now revoked too, alongside API tokens, MCP tokens and OAuth connections, so log in again after a password change.

• The wildcard scope is rejected API MCP

  GET /oauth/authorize answers scope=* with 400 invalid_scope, and a refresh at POST /oauth/token whose scope is not a space-delimited string (for example a JSON array) or asks for * gets 400 invalid_scope instead of a token with a different scope. A refresh that omits scope keeps the granted scope, as before.

• A client can always revoke its own refresh token API MCP

  POST /oauth/revoke now revokes a refresh token even when its access token no longer exists.

2026-09-16

• Rejected refresh tokens return invalid_grant API MCP

  POST /oauth/token now answers a refresh token that was rotated out, revoked, expired, or issued to another client with 400 and error: invalid_grant, as RFC 6749 specifies. It previously returned 401 with invalid_request, which OAuth clients do not treat as a reason to re-authorize, so a connection that lost its refresh token kept failing instead of asking you to reconnect. See Token lifetimes.

2026-09-15

• The 90-day audience retention rule is now scoped to MAID and IP deliveries API

  It exists because MAID and IP mapping keys are retained for 90 days, so it no longer blocks an audience outright: audience reads report it under a new eligibility.notices[] array (code: audience_expired, blocks_identifiers: ["MAID", "IP"]) with allowed true, and Create Activation accepts EID / SCID / HEM pricing models on such an audience while rejecting a MAID or IP pricing model with 422. eligibility.reasons[] keeps only the rules that block every delivery (device floor, Affinity device coverage). Previously any standard audience whose data window ended more than 90 days ago could not be activated at all, whatever the identifier.

2026-09-14

• Chart opt-in on audience create is honoured API

  The optional datastreams array sent to create was validated but not applied: the audience built without its charts and the Data Streams tab stayed empty. Opted-in streams now produce their pending rows and are generated as the audience builds.

• Activation datastreams are checked against the audience API

  An enabled entry in datastreams[] that applies to none of the audience’s dataset types is now rejected with a 422 naming the stream, instead of being accepted and failing in the export. Get Datastreams items carry a new dataset_types field to filter on first; a lookalike audience counts as cohorts.

2026-09-13

• Data stream visualizations on audience create API

  An audience can now generate its charts as it builds. Send the optional datastreams array on create - the same {id, visualizing_status} shape the activation surface already uses. The rendered charts are viewed in the Intuizi console, on the audience’s Data Streams tab. No endpoint returns a visualization payload or a PDF export.

• New reference catalog: Get Datastream Visualizations API

  Get Datastream Visualizations lists the streams you may use, optionally scoped with dataset_type to the set create will accept. Distinct from Get Datastreams, which is an endpoint partner’s delivery outputs.

2026-09-09

• Cohorts from a Lookalike Model audience take a score range API MCP

  Create Cohort from a Lookalike Model audience (POST /api/v2/analyses/cohorts/create with source audience) accepts score_limit with min_score and max_score, which keeps the devices inside the model’s score bands, best scores first, capped by device_limit. A Lookalike Model audience can now back several cohorts. MCP create_cohort advertises the same fields.

2026-09-02

• Preview Cohort File works for folder sources API

  A file_uri that names a folder (no .csv, .gz or .parquet suffix, the same rule create applies) previews the header of its first data file, skipping Spark’s _SUCCESS and similar markers. Previously such a preview failed with a 422 and the identifier column had to be typed without seeing the columns. See Preview Cohort File.

• Deleting a lookalike audience that is still modelling stops the modelling job API Console

  POST /api/v2/analyses/audiences/delete-by-id (and the console’s delete) raise the same cooperative cancellation flag that cancel-lookalike sets, and the worker’s progress callback honours it even after the audience row is gone. Previously the job ran to completion for an audience nobody could see. Response shapes are unchanged.

Earlier months