Skip to main content
These workflows require an enrolled Buyer account or the integrated own-supply sandbox capability on a Media Company. Call get_status first. An integrated Media Company stays in the same account and never uses switch_account for campaign work.

Inspect account activity

In the standard buyer workspace, open Activity from the Account group in the left navigation. Activity is available before you select an advertiser and shows the account’s Calls and Changes. Advertiser and campaign filters narrow the view; they do not control whether Activity is available. The integrated buyer campaign workspace keeps its campaign-focused navigation. In the Agents workspace, open API calls to inspect the same account-level activity without leaving that workspace. When you request proposals through Murph, its runtime continues the same request for up to eight dispatches within 30 seconds. It preserves the campaign revision and idempotency key. An unfinished response is provisional, including when no products have arrived; it does not mean there are no matches. If the wait limit is reached or accumulated result pages exceed its response-size limit, Murph retains the request and reports incomplete results. It does not automatically start a replacement round or attach provisional products. Completed results still follow the normal product-selection and attachment checks. Reading full targeting capability pages preserves the original request and any confirmed attachments. If a first request is rejected before execution starts, you can correct it and retry in the same conversation. Errors during an existing execution or with an uncertain outcome keep that execution protected from replacement.

Buyer workflow at a glance

  1. Confirm buyer operator readiness.
  2. Create or select an advertiser.
  3. Create a campaign with its brief, flight, and budget.
  4. Request proposals from ready sellers.
  5. Accept a quoted proposal or stage returned products.
  6. Inspect the staged media buys and resolve any draft issues.
  7. Launch the campaign explicitly.
  8. Query bounded campaign delivery.

0. Confirm buyer operator readiness

Call get_status before creating buying work. When operatorIdentity.usableForBuying is false, new discovery and buying calls are blocked with BUYER_SETUP_REQUIRED. An account administrator must call save_buyer_operator with the buyer’s real non-platform domain and choose whether this account represents the whole_operator or a specific_unit with a stable AdCP operator_unit ({ "id": "east-coast" }).
When the domain is usable, scopeStatus is unclassified, and locked is false, reuse that domain and choose its scope before new AdCP 3.2 provisioning. Existing buying remains available during that scope migration. If the identity is locked, follow the support action from get_status instead. Confirming the commercial operator does not add users, change membership, or change the login organization. See Buyer setup and go-live for scope selection and identity-locking rules.

1. Create an advertiser

Creating requires name and brand. primaryCurrency defaults to 'USD' when omitted; pass any ISO 4217 code to override, or update it later while the advertiser is unlocked. When name or brand is missing the tool returns a neutral needs_input result naming the missing fields. A needs_input result is a question, not a failure; nothing is saved until the tool is called again with the answer. The currency stays editable until the first campaign or seller binding locks it. To change it, send advertiserId with the new primaryCurrency as an ordinary update; you never need to archive and recreate the advertiser. sandbox is immutable after creation.
Retain the returned advertiserId. To update, send it with only the fields to change. Do not send sandbox on an update.

2. Create a campaign

Creation requires advertiserId, name, and an idempotencyKey. flight and budget are optional for a draft. Add them when they are known, or save the draft first and complete its terms later. Write the brief from the buyer’s stated goal, audience, and what is being promoted.
Creation does not launch. Retain the returned campaignId and revision. In the Campaigns Page, New campaign creates an untitled draft in one click when the Page is scoped to an advertiser. An unscoped Page first asks you to choose an advertiser. The campaign opens immediately, where you can edit the name, budget in the advertiser’s currency, and flight in the advertiser’s reporting time zone. Tracked campaigns remain read-only. In chat, ask for a campaign and the agent creates the draft from your request, then opens the campaign.

Cap exposure within each seller media buy

Set frequencyCap with level: "mediaBuy" before requesting proposals when the buyer wants one AdCP 3.2 counter across the participating packages bought from each seller. Discovery limits the result to sellers and products that declare compatible support, and launch checks that support again. Each seller has an independent counter. This field does not cap exposure across different sellers. See Media-buy frequency caps for the request shape, supported reach units, and current level boundaries.

Target geographic areas

Update save_campaign targeting

save_campaign now accepts the AdCP targetingOverlay field. The requirement-shaped targeting field is deprecated, with behavior unchanged. Do not send both fields in one call. Removal is a separately announced change (AI-10440) with at least 14 days’ notice. Send canonical codes in the new field. Before:
After:
Campaign targeting is the default for proposal discovery and for a product that omits targetingOverlay. An explicit targetingOverlay on a product supplies that buy’s targeting, and campaign targeting still narrows it: overlapping values are kept, and a product value outside the campaign returns CONFLICT. Media-buy readback returns each package under AdCP Package names (package_id, budget as a number in the buy’s currency, bid_price, performance_standards), with the effective targeting_overlay and its targeting_source (campaign or buy). The source is recorded when the draft is staged, so later campaign changes never rewrite the buy’s history. On save_media_buy, products[].targetingOverlay accepts the AdCP package targeting overlay except signal_targeting. Typed signal_targeting_groups remains supported. signal_targeting is a priced, product-scoped selection and will come through the signals path. The overlay includes geographic fields, frequency caps, property and collection lists, placement and collection selection, audiences, browser, language, device, age restriction, dayparts, store catchments, proximity, and keyword targeting. Sellers enforce the fields their returned product supports. demographics remains discovery and readback only while seller transport is on AdCP 3.1. save_media_buy rejects it before contacting a seller. Do not add targetingOverlay when updating an existing draft; it is create-only. Countries accept ISO 3166-1 alpha-2 codes or names and regions accept ISO 3166-2 codes or names; both resolve to canonical codes before dispatch. Postal areas use the country-aware AdCP shape. Each geographic field must contain at least one value. Nielsen Designated Market Areas (DMAs) remain available while saving a campaign or staging a media buy. On save_campaign, use targeting.geoMetros. On save_media_buy, use products[].targetingOverlay.geo_metros to include DMAs or geo_metros_exclude to exclude them. Campaign targeting can still use its V3 requirement form. On a media-buy overlay: Deprecated. Numeric codes replace names. Removal will be announced in release notes at least 14 days ahead; planned for 2 November 2026 (AI-10440). Codes must be in the Nielsen DMA dictionary; unknown codes return nearby code candidates.
save_media_buy returns an actionable validation error when it cannot resolve a DMA name. Do not guess a code.

3. Request proposals from eligible sellers

get_status reports how many destinations are currently ready and includes a bounded readyDestinations sample for explanation. Ready means the seller will be considered for contact when you make a fresh request_proposals call: it has an active sales agent and meets the buyer-specific canBuy checks. The request rechecks the complete marketplace server-side before it contacts that same group. Campaign constraints and the fresh recheck can still exclude a seller from contact.

Seller cohort rules

The server determines the final cohort: When sellerIds is supplied, any requested seller outside the campaign list, outside an authorized seller scope, or not currently eligible returns CAMPAIGN_SELLERS_INELIGIBLE with canonical seller IDs and contacts nobody. Existing calls that supply sellerIds now narrow the round and never reach sellers outside that scope. Without sellerIds, a campaign list instead contacts its eligible subset and reports unavailable campaign sellers. The market requirement below is checked before any seller cohort is contacted. To broadcast, omit sellerIds and use a campaign with no seller list, then set confirmBroadcast: true. Existing callers that supply sellerIds now narrow the round; they do not broadcast. The response’s appliedCohort shows whether the round used the requested seller IDs, the campaign seller list, the authorized seller scope, or a confirmed broadcast, along with the bounded seller IDs contacted. Channel groups narrow the round. When the campaign has channel groups, a seller whose declared channels match none of them is not contacted, and the response names it in skippedSellers with the reason channel_mismatch. An audio-only campaign therefore does not contact display-only sellers. Returned products that fit none of the groups are dropped and counted on the seller’s outcome as channelExcludedProductCount. A proposal round needs a market. Set at least one country on the campaign with save_campaign targetingOverlay.geo_countries (for example ["US"]) before calling request_proposals. A region in targetingOverlay.geo_regions, a Nielsen DMA in targetingOverlay.geo_metros, or a postal area also counts: the round reaches the sellers that cover that country. Exclusions alone name no market. Scope3 does not read geography out of the campaign’s outcome text, so “US only” written there is not targeting until it is set on the campaign. request_proposals enforces this requirement: it refuses a campaign whose targeting names no country and contacts no seller. The VALIDATION_ERROR names targetingOverlay.geo_countries; set it (or targetingOverlay.geo_regions or targetingOverlay.geo_metros) with save_campaign, then retry with the new campaign revision. Every solicited seller receives the same brief: the campaign’s stated outcome, flight, geography, channels, language, device constraints, required creative formats, audience requirements by reference, and the campaign’s ranked optimisation goals with their targets. A seller reads a goal such as “clicks, target cost 3.00 or less per click in the buy currency, counted by the seller’s delivery metrics” and can price against it. Event goals disclose the event type and never your event source id or tracker configuration. The campaign budget, pacing schedule, other sellers being considered, and internal configuration are withheld; the categories withheld from each seller are recorded on the brief artifact.
Which call a seller receives. Sellers are contacted through AdCP get_products in brief mode and answer with proposals inside that response. A seller whose capabilities declare proposal support, either as media_buy.supports_proposals: true or by listing request_proposals in media_buy.lifecycle_tools, and that declares the exact AdCP 3.2 release (3.2 in adcp.supported_versions) instead receives the canonical request_proposals call with the same brief and its criteria (the flight, geography and channel filters, and any targeting overlay), and its canonical proposals are stored with their commercial terms, so the commitment recorded for each proposal is read from the seller’s own per-purchase terms. Sellers without that declaration keep receiving get_products. A seller that declines to propose is recorded as that seller’s failure with its stated reason, never as an empty result. The canonical call can also carry the primary cost goal as a structured outcome target with no volume, to sellers that declare they can plan one (see What sellers see). The campaign budget is never sent.
Before a round, state which sellers will be contacted. If the buyer has not asked to contact all eligible sellers, get an explicit yes before calling with neither list. After the round, say who was contacted. Before a seller-visible brief is recorded or sent, each seller must also pass the campaign’s structured market and channel constraints. For an active sponsored buyer using a sandbox advertiser, the server instead confines the cohort to that buyer’s sponsoring storefront and applies the sandbox transaction path. The sponsoring storefront does not need to be open to the public marketplace for this no-spend workflow.
expectedSellerId is a Storefront ID, not an internal customer ID. It is an optional fail-closed precondition for automation that must remain confined to one seller. A fresh round fails if the buyer’s current server-side authority does not resolve exclusively to that seller. The value can narrow an already-authorized scope; it cannot authorize a seller or reduce a normal marketplace buyer’s complete eligible cohort. Omit it when broad marketplace discovery is intentional. For a fresh round, the call durably schedules complete eligibility enumeration and returns running; the frozen seller count may therefore be zero on the first response while discovery is pending. Each background seller attempt has a 30-second bound. Retry the exact same idempotency key until the result becomes complete, partial, or failed. The execution stores the resolved seller cohort, so retries never silently add, remove, or duplicate sellers. Each seller may return:
  • quoted with qualified Proposal IDs;
  • products with a productQueryId; or
  • failed with a bounded error.
Repeating the same idempotency key returns the same proposal round. Use a new key only when intentionally asking sellers for a fresh round. Evaluation instructions are recorded but are not yet applied to ranking; review the returned results yourself. Only one proposal-request execution may run for a buyer at a time, across all campaigns. Poll the active execution to terminal before starting another. After the execution is terminal, follow every page.nextCursor. perSeller contains at most 50 outcomes on the current page, while product-heavy outcomes may continue for the same seller on the next cursor. Product details are bounded for transport; detailsTruncated: true marks a bounded field projection. The productId remains the selection key. Meanwhile, summary.sellersRequested always counts the full frozen cohort.

4. Stage a media buy

Read what a seller committed to

A proposal’s allocations, and each row of its products include, use AdCP ProductAllocation names: product_id, allocation_percentage, pricing_option_id, and rationale. Per-seller errors from request_proposals are AdCP errors (code, message); a seller’s own explanation, when it gave one, is in seller_explanation, fenced as untrusted text. A product’s resolvedAgeTargeting is an AdCP age range (min, max, include_unknown); max is left out for an open range such as 65+. Every quoted proposal carries goalAnswers, one entry per allocation, and a proposal-level commitment. They are read off the seller’s terms, never its prose: the pricing option it chose for the allocation, the product’s delivery type and measurement terms, and any optimisation goals the seller declared when it asked to optimise the budget itself. Each answer names the pricing model and fixed price, the delivery type, the goal you asked for, the target the seller’s terms actually commit or aim at (answeredTarget), and a commitment kind: commitment.kind is the weakest kind across allocations, because a proposal is only as committed as its least committed part; commitment.answeredTarget is the worst answered target across allocations (the highest cost per unit, the lowest rate, or the lowest return); commitment.meetsAskedTarget is true only when every allocation answered your target and met it by its kind (a cost at or under your ask, a rate or return at or above it), false when any allocation missed it or did not answer with the same kind, and null when your goal carries no target or a maximise target. Proposals recorded before this field existed are read the same way from their stored terms, without the asked goal. Each goal answer also carries goalCoverage when the proposal kept the product’s AdCP optimization declaration: coversPrimaryGoal, and uncovered, listing each goal the product cannot optimize to by priority (1 is primary) with a reason code. Proposal-less products returned by request_proposals carry the same goalCoverage. In text, a proposal read and the request_proposals result name each such product with its uncovered goals, for example priority 1 (metric_not_declared). A product that declares no optimization capability covers no goal. You can still book it; it reports against the goal rather than optimizing to it. When your primary goal is a views, viewable_rate or completed_views goal with a target and the seller’s forecast carries a range for it, goalCoverage.expected adds what the product is expected to deliver: the rate’s basis, its low, mid and high (0 to 1), the rate required by your target (for a cost target, at the allocation’s CPM), and a verdict of likely, possible or unlikely. Text names it in the same line, for example expected viewable rate 78%-88% vs 70% needed (likely), and the request_proposals result shows it as primaryGoal: .... Expected performance against your goal explains each basis. Goal-seeking campaigns explains the goal statement these answers respond to and who carries the risk under each commitment kind.

Accept a quoted proposal

The proposal version must still be current and belong to the campaign. The result is a draft media buy; accepting a proposal does not launch it. The current schema requires idempotencyKey on every call, although proposal acceptance derives retry safety from the qualified proposal version and does not consume the supplied key.

Stage returned products

For a seller that returned products without a Proposal, preserve every returned identity field and use that seller’s productQueryId as the idempotency key:
Do not reconstruct qualified IDs. When returned, inventorySourceId, salesAgentId, and pricingOptionId distinguish the exact Product route and price selected from the returned catalog. When a returned Product advertises signal_targeting_options, select an eligible Signal through that Product’s targetingOverlay. Preserve its signal_ref, value type, bounds or values, activation handle, and pricing identity exactly as returned; the Seller validates eligibility at launch. save_media_buy accepts flight when staging from returned products or accepting a proposal. It accepts top-level budget for one returned product or a one-allocation proposal. For multiple products or allocations, keep the returned allocation or set products[].budget for each selected product. Accepting a proposal creates one media buy, so every allocation must use the same settlement currency and seller route. If a proposal spans currencies or seller routes, save_media_buy rejects it before claiming the proposal or creating a draft; request separate proposals by currency or seller route.

5. Inspect staged work

List media buys under the campaign:
Or list every current buy for an advertiser across its campaigns, paged (limit up to 200; pass back nextCursor for the next page):
Each row reports its campaignId (when the buy belongs to a campaign), phase, pause state, flight, the seller (as sellerId plus sellerName), and the buy’s gross budget. sellerId is always a Storefront ID, never an internal customer ID. It does not retain Proposal evidence after the acceptance response, so preserve the proposalSource fields returned by save_media_buy when that audit link matters. When both filters are present, campaignId wins. An advertiser that is not in your account returns NOT_FOUND, not an empty list. Read one buy’s execution tree with get:
The base object carries campaignId, sellerId, sellerName, budget, flight, and the why-visibility fields (pendingReason, errorCode, forwardedAt, buyerReference). When a submitted change is not live yet, it also carries pendingChange; see A change that is not live yet. Includes add: pendingChange is also accepted as an include. The base object already carries it whenever a change is waiting, so asking for it changes nothing. Any other include is echoed in unavailableIncludes with the reason. The content[0].text of every read mirrors these facts — seller names, budgets, line items, packages, and format labels — so an agent reading only text sees the same buy a structured-first host does. Proposal reads and request_proposals outcomes name their sellers the same way. Archive is a visibility flag, not a lifecycle phase. A read or search(kind: "media_buy", filter: { isArchived: true }) reports an archived buy with isArchived: true and its preserved phase. This lets you distinguish an archived draft from a completed or canceled buy. Older archived buys may report phase: "completed" because the previous archive lifecycle overwrote their status with ARCHIVED and did not retain the earlier phase. completed is a compatibility stand-in for those rows, not a record of their pre-archive phase. At this step, inspect each draft, apply any supported correction with save_media_buy({ mediaBuyId: ... }), and continue only when the staged set is the one you intend to launch. Supported corrections on a draft:
  • flight — move the start or end date.
  • budget — a new total, on a draft with exactly one line item.
  • products[] — per line item, change budget, bidPrice, or pricingOptionId, or set remove: true to drop the line item. A product that is not already on the buy cannot be added; stage a new media buy for it.
  • isArchived: true — remove an unwanted draft, failed, canceled, or rejected buy from the buyer’s list. The campaign is unchanged. Failed, canceled, and rejected buys are removed locally without another seller cancellation. A live or pending buy is cancelled through the campaign or the v2 update contract.
Anything the draft cannot absorb (a new product, a different inventory source or seller route, changed Signal targeting, restoring an archived buy) is refused with NOT_IMPLEMENTED naming the field, never silently ignored.

Change a live media buy

Once a media buy has gone to the seller, save_media_buy can change its flight and its budget:
  • Before the buy starts, send a new startAt, endAt, or both.
  • After the buy starts, only the end date can change. flight needs both fields, so send the buy’s current startAt unchanged with the new endAt. A different start on a running buy is refused.
  • products[].budget sets a line item’s new total budget, including what it has already spent. A top-level budget does the same on a buy with one line item. The seller spends against the line item’s packages, so the change goes to the ones still running:
    • a package whose flight or pacing period has ended keeps its budget;
    • the open packages share the rest in their current proportion, and none is set below what it has already spent.
  • The lowest total you can set is what the ended periods hold (their budget, or what they spent if that is more) plus what the open periods have already spent, including fees; a lower total is refused with that amount. A budget change to a line item whose flight and pacing periods have all ended is refused.
  • When the same product is on the buy as more than one line item, name the one to change with products[].lineItemRef; otherwise the change is refused with the line items listed.
  • A live buy’s bid cannot change here yet.
  • Retrying a change with the same idempotencyKey and the same request returns the first result instead of applying it again. The retry is still checked against the buy as it is now, so if the buy changed in between (for example it was archived, or its waiting change was rejected), the retry gets that refusal instead of the first result; it never applies the change twice. The same key with a different request is refused, so use a new key for each new change.
To end a running buy on 15 December, read its flight and send the start back:
The call waits for the seller to answer:
  • action: "updated" means the seller confirmed the change and it is live.
  • An error with code SERVICE_UNAVAILABLE means the seller did not confirm. Read the buy before trying again; its pendingChange shows whether a change is still waiting.
  • action: "pending_approval" means the call succeeded, but the buy still has a change that is not live on it yet. The returned mediaBuy shows the live values and carries pendingChange, described below. Do not send the change again or tell the buyer it is done.

A change that is not live yet

A media buy can have a submitted change that has not reached the delivering buy yet. A single-buy read, get({ kind: "media_buy", id }), and the mediaBuy returned by save_media_buy show it as pendingChange. Every other field still describes what is live, and the buy keeps delivering on those terms until the change takes effect:
  • status is the status of the change itself, typically PENDING_APPROVAL. The buy’s own phase is unchanged.
  • pendingAt says whether the change is waiting on the storefront operator (storefront), the seller’s sales agent (salesagent), or neither can be told (unknown). It can be absent.
  • differences compares the flight end (endTime), the total budget, and the attached creatives with their live values, and lists each one that differs as live and proposed. A change to anything else, such as a bid, can leave it empty.
  • reason and proposalId appear when the change recorded them.
Once the change takes effect, the new values show as live and pendingChange is gone. Do not tell the buyer a change has taken effect while pendingChange is present. Search rows and the campaign mediaBuys include do not carry it. This mediaBuy.pendingChange is not the top-level pendingChange on a pause or resume response. That one’s status is the status the buy moves to once the seller confirms. An abbreviated read:

Pause, resume, or cancel one media buy

To pause one running media buy without pausing its campaign or sibling media buys, call save_media_buy with its mediaBuyId, isPaused: true, and an idempotencyKey. To resume that same buy, call it again with isPaused: false. Send isPaused on its own: do not combine it with a flight, budget, product, or archive change. Only an ACTIVE buy can be paused and only a PAUSED buy can be resumed. The response’s action is paused or resumed once the seller has applied the change. Some sellers accept a pause or resume and apply it later; the action is then pause_requested or resume_requested:
  • isPaused, previousStatus, and newStatus keep the buy’s live values. After pause_requested the buy is still ACTIVE and may keep delivering. After resume_requested it is still PAUSED.
  • pendingChange.status is the status the buy moves to once the seller confirms, and a pause also carries pendingChange.pauseRequestedAt.
  • Read the buy again to see the change take effect. Do not tell the buyer the pause or resume is done until it has.
To cancel one media buy without changing its campaign or sibling buys, call save_media_buy with mediaBuyId, isCanceled: true, and an idempotencyKey. Send isCanceled on its own. A canceled buy returns unchanged when called again; isCanceled: false is refused because cancellation cannot be reversed. Some sellers or operator-controlled buys must confirm cancellation, so the response can say that the request is not immediate and report a non-terminal new status until confirmation arrives.

5a. Review the draft before going live

When the host renders MCP Apps, open_campaign_receipt opens Review & go live for one draft campaign: its budget and flight, the staged media buys with their budget split, why each buy is not live yet, and the readiness blockers still standing between the draft and launch (a creative that is not ready, or no media buys staged). The tool returns the shared MCP App directive, a compact text summary of the same facts, and the projected receipt in structuredContent.receipt. It reads only: going live remains the explicit save_campaign step below.
The receipt covers draft campaigns only. For a campaign that has already gone live, open_campaigns_page with the same campaignId opens its record instead. The headless equivalent is get({kind: "campaign", id, include: ["mediaBuys"]}), whose mode: "review" workspace carries the same readiness.blockers.

6. Launch explicitly

Launch is a two-call update to an existing campaign, not part of campaign creation. First preview — this call launches nothing:
The response is action: "pending_confirmation" with campaign.revision and, under launch, the media buys that would go live and their combined budget. Show it to the buyer. Only after they say yes, confirm with the previewed revision:
See Launch a campaign with explicit confirmation for the full preview shape. Do not combine launch with pause or archive in the same call. A launch can partially write downstream execution state even when no media buy activates; read the structured error, fix the cause, re-read the campaign revision, and retry deliberately.

7. Query campaign delivery

Use get_delivery with report: "campaign_delivery". Supply either an explicit UTC date range of at most 90 inclusive days or range: { "lifetime": true } for everything from the campaign’s earliest reported delivery to today, and choose only the metrics and dimensions needed by the caller:
For the whole life of a campaign, grouped by the seller it ran with:
Metric IDs that AdCP defines use AdCP’s names, such as completed_views, conversion_value, and completion_rate, both in metrics and as the keys of each row’s and the totals’ metrics. The response echoes the resolved window as period and always carries totals: the same rollup as the rows, over every matched row rather than the page (totals.rowsIncluded equals page.total). When matched rows span more than one currency, totals.currency is null and money metrics are unavailable with reason: "currency_unavailable"; counts and unitless rates still total. Nothing is FX-converted. seller is the storefront the buy was placed with and sales_agent the AdCP sales agent behind it, both resolved from the products on the buy. A row whose seller cannot be attributed keeps seller: null, groups with its peers, and still counts in totals; it is never folded into another seller or dropped. Filter by advertiserId, campaignId, channelGroupId, mediaBuyId, or packageId. channelGroupId is applied before the report is bounded, so rows from other Campaign groups cannot displace matching rows. packageId requires a bounded date range (startDate and endDate, at most 90 days) and cannot be combined with range: { "lifetime": true } — the reporting operation fetches all campaign data before filtering by package, so lifetime package filtering is rejected before dispatch. An integrated Media Company must name at least an advertiser, campaign, or media buy; the server re-proves that scope against its sandbox advertiser and exact own Storefront before querying. Advertiser-wide integrated queries also fail closed if any current buy under that advertiser has wider supply. Rows preserve the Buyer reporting denomination: spend, ecpm, cpc, and cpa are gross and fee-inclusive where the underlying buy has pinned terms. The response names its currency and reports numeric zero as available. A null derived rate remains unavailable when its denominator or conversion signal is absent. This projection is seller-reported delivery viewed through the Buyer hierarchy; it is not Buyer measurement. The V2 compatibility source does not expose ordered revision evidence, so V3 does not infer SNAPSHOT or OFFICIAL finality or billing eligibility, and finality.revision is null. When a revision is present, it is an AdCP ReportingRevision with one added field, received_at: when Apostra received that revision. Follow nextCursor without changing the query when page.truncated is true.

Read the connected provider live

When the current provider response matters more than stored aggregation, use the same tool with report: "live_campaign_delivery", one exact campaignId, and a bounded range or range: { "lifetime": true }. Do not pass metrics, dimensions, other filters, or a cursor. The existing delivery field retains the provider’s full response for current callers. SDK-defined fields are validated and the whole response is bounded by the tool’s response limit, but provider extensions remain opaque. Use the new deliverySummary field for a stable typed result: identity, currency, reporting period, status, finality, aggregate metrics, media-buy totals, package summaries, and paging and truncation signals. If adding the summary would exceed the response limit, deliverySummary is omitted and the existing delivery response remains available. The summary keeps the AdCP get_media_buy_delivery response’s own field names and places: aggregate totals, media-buy totals, and package summaries carry their delivery metrics (impressions, spend, and so on) directly on each object. The summary includes up to 250 metric aggregates, 100 media buys, and 100 packages per media buy. Its count and truncation fields (metric_aggregates_truncated, by_package_truncated, media_buy_deliveries_truncated and their _total_count fields) are the only fields it adds to AdCP’s objects; they identify locally omitted items separately from provider pagination. Package summaries keep AdCP’s breakdown completeness fields where provided: by_<kind>_truncated, by_<kind>_suppressed, and by_<kind>_pagination. Package finality retains its measurement window and superseded window; reach retains its unit and window, and rates retain their pricing model. Detailed geography, creative, device, window subseries, and other breakdown rows stay outside deliverySummary. The summary also omits provider credentials, free-text provider messages, opaque extensions, and raw reporting rows; those provider fields may still appear in the legacy delivery field and its text rendering. Qualified metric aggregates retain their scope, metric ID, numeric value, and compact qualifier; vendor identity is limited to domain and brand ID. It marks its authority as live_provider and does not claim reporting-pipeline finality or billing eligibility.

8. Save uploaded assets as a creative

Select the owning advertiser before preparing the upload. Only a finalized JPEG or PNG scope3-asset:// reference bound to that advertiser can be passed verbatim as sourceAssetRef to the existing save_creative tool, together with the matching advertiserId or campaignId. If no advertiser was selected before preparation, keep using the reference through the existing account-scoped delivery flow. For an eligible reference, use the existing creative noun; there is no separate adoption tool:
The upload and campaign must resolve to the same customer and advertiser, and the authenticated principal must own the upload. Prepared, expired, cross-customer, and cross-advertiser references are refused. sourceAssetRef cannot be combined with assets; its JPEG or PNG media supplies the canonical image format identity. The private source is temporary. Saving copies its verified bytes into the governed creative store, so the Creative remains usable after the upload source expires. The stored manifest contains the governed asset and a one-way fingerprint only—not the source reference, private path, or a signed URL. When you include campaignId, the server creates the advertiser-scoped Creative first and then applies the existing idempotent campaign membership. Retry with the same reference and name if a response is interrupted; the retry returns the same Creative and finishes a missing attachment. Use advertiserId instead to save it without campaign membership. For an advertiser Creative with multiple uploaded assets or a promoted MP4, first search creative_format with the advertiser and selected product. Pass the returned opaque format id as creativeFormatId, then bind each durable asset assetId to its declared slot in sourceAssets. This keeps the Creative in the advertiser Library without requiring a campaign or contacting a provider. Ordinary assets are unbound and cannot satisfy those format slots:
The signed format is account- and advertiser-bound. The server revalidates its seller, product route, option, and declaration before reading an asset or writing the Creative. If formatKind or formatOptionRef is also supplied, it must match the selected format. The MP4 must have reached promoted, and all sources must resolve to the same customer, principal, and advertiser. A successful save confirms the canonical advertiser Creative. For a new Creative that should be attached immediately, use campaignId, formatKind, and the campaign product’s exact formatOptionRef instead of advertiserId and creativeFormatId. The save returns providerContacted: false; video destination delivery remains deferred. Do not claim delivery until later sync and exact provider readback prove it.

9. Author a social creative

A social creative is authored copy, not an uploaded file. save_creative accepts a social block and stores each field in the standard AdCP text and URL slots, so the creative round-trips through save, get, and search without a preview or transcode step.
Platforms name the same authored field differently (TikTok uses display_name and ad_text, Meta uses primary_text). When a creative is pinned to a platform format, the format declaration decides which slot ids are valid: the save is refused with the list of declared slots if a slot is not declared, and per-slot length limits apply. Use components to write those platform-native slots. Page or profile identity (a Facebook Page, an Instagram account) is not a save_creative field; a format that declares it as a slot takes it through components. Social copy is content but carries no format identity, so a new creative still needs formatKind (for example image or video_hosted) or a media asset. On update (creativeId + campaignId), each slot named in social replaces the existing value on that slot and unnamed slots are left unchanged.

Read a creative in full

get with kind: "creative", sourceId (the campaign ID), and id returns the complete record. Every field save_creative also accepts comes back under the same name and nesting, so a read can be edited and saved back: save_creative takes tags only as labels.tags, the same place the read returns them; see Dimensions for label values. Format identity is canonical only: a creative read never carries a legacy agent_url. A creative without a canonical formatKind reports requiresUpgrade: true and cannot be assigned to a new media buy until it is upgraded through the v2 API. search with kind: "creative" scopes by filter.campaignId or filter.advertiserId. An advertiser-scoped search narrows with formatKind, assetType (media kind, for example IMAGE or VIDEO), role (evergreen or reference), source (uploaded, generated, connected), and promoted. A campaign-scoped search supports query only, and the narrowing filters are refused under campaign scope rather than silently ignored. Each row reports formatKind, mediaKind, assetCount, and requiresUpgrade; the text block repeats them for text-only hosts. Pagination is the opaque cursor from the previous page. Filters by readiness state, archive state, or media-buy assignment are not available: archived creatives are not listable, and an isArchived: true filter is refused rather than returning the active list.

Archive a creative

save_creative with isArchived: true, creativeId, and advertiserId removes the creative from every campaign and frees its name, so a later create under the same name is a new creative rather than a dedupe hit. Archiving is permanent: isArchived: false is refused, an archived creative reads as NOT_FOUND, and there is no allowArchived read for creatives.

Interactive buyer Pages

Three buyer Pages have fixed v3 owners. Each owner binds one MCP App resource in its tool descriptor, so a host that renders MCP Apps opens the same Page from Murph, Claude, or ChatGPT; a host that does not render them receives the text summary. The Pages self-fetch their data, so none of these launchers puts the list into model context — use search and get for text answers. The compatibility open_page enum never lists these Pages; the owner tool is their portable contract. Task pages: Open Advertisers, Open Campaigns, and Open Review & go live.

Lifecycle operations

  • isPaused: true pauses an active campaign; false reactivates it.
  • isArchived: true archives it from default lists without changing its phase, pause state, media buys, or budget commitment. It is refused while any executable or unsettled media buy remains; cancel or settle each named buy first, then archive the campaign.
  • isArchived: false restores an archived campaign with the same phase and pause state it had when archived. Send it alone, then re-read the campaign before making further changes. Legacy rows whose status was overwritten to ARCHIVED have no recoverable earlier phase: they read as archived with phase: "completed", including in an archived completed-phase search, and cannot be restored; create a new campaign instead.
  • desiredPhase: "canceled" requests cancellation of the campaign and every nonterminal media buy. Send it alone with the current expectedRevision. Cancellation is asynchronous: a requested cancellation is canceled, cancel_pending, or cancel_failed; a buy that was already terminal keeps its observed terminal phase. The campaign stays ending until every buy is terminal. Retry with the latest revision to re-drive pending or failed buys without re-ending confirmed ones. This remains separate from archive and does not archive the campaign.
  • save_creative with isArchived: true archives a creative permanently and frees its name; creative unarchive does not exist, so isArchived: false is refused (see Author a social creative).
  • A tracked campaign is read-only until it is adopted or duplicated through the existing v2 workflow.
  • autonomy fields are accepted for forward compatibility but are not persisted yet.
See Preview limitations before replacing a v2 buyer integration.