November 18, 2026
- Past answers are priced. Reading the last 90 days stays free. A public answer older than that costs $0.001 the first time the workspace counts it in a number or lists it, and is the workspace’s after: free in any later number, cut or list. Answers the workspace bought are never charged. A counting call that reaches further back (
query_metrics,search_cited_pages,search_answerswithview: "diff") takeshistory:"auto", the default, buys one even sample inside the month’s budget, whose past answers are worth up to 25% of the budget in effect and at most $5, with no quote;"all"counts every answer;"free"reads the last 90 days and says what each sample would cost. The reply says what was counted and charged incontext.sampleandcontext.history, andmeta.costCreditscarries the charge. A list (search_answers,readby id,export) charges each past answer it delivers once.apply_rubric’starget.filterbuys every past answer it selects before the run, never a sample. A workspace with no credit still reads the past answers it bought. When nothing fits, the call reads the last 90 days with ahistory_unpaidentry inlimits, its counts of answers not bought said with “about”. With no key, numbers over the public window are unchanged, and a trend or a list reaches back no further than the last 90 days. share_recommendedexcludes answers whose mention of the brand still awaits judgment and returns its ownanswerscount. Recommended shares can read higher for brands whose judgments lag. The app, public pages and agent service use their own answer scopes and windows.check_brandleads with Recommended: each market’sshare_recommendedfirst, most recommended first, withshare_namedbeside it as the detail. The market it reads rivals in is the one where the brand is recommended most, and the rivals are the brands recommended most there.- MCP
tools/listis shorter: it leaves out the arguments few calls send, and a filter nested inside another argument points to the filtersearch_answerslists. Every argument still works.resolve_ids {"kinds": ["argument"], "q": ["search_answers"]}lists a tool’s other arguments, and each tool that has some says so./openapi.json,GET /v1/toolsand these pages list every argument. - A data read’s reply carries its one-sentence
data.noteunlessexplain: false. Withcompare: "prior",query_metrics’ note says whether each change is significant.confirmonfetch_live_pagesandapply_rubric, andrenderonfetch_live_pages, are no longer listed;confirmis still taken. - Parse’s public market and topic pages are computed by
query_metrics, the call an assistant makes, so a page and a reply count the same answers the same way. Logged out, numbers count the answers parse.gl’s public pages count, each product a brand of its own. They counted every public answer under the market, a company and its products as one. Signed in,filter.scope: "index"reads the same answers and the default is unchanged. query_metricsreturnsrankon a cut by brand, tied brands sharing a place as on the pages, and takesallBrands(every brand named, not the 50 named most) andtopPerGroup(the first rows within each value of the first of two dimensions).- New measures in
query_metrics:share_citing(the share of answers that cite a site or page, with both counts), and four readings of a whole filter:page_type_mix(cited pages by type),hedge_mix(how answers qualify a brand, with one sentence),own_site_citations(how often answers cite the brand’s own site, by week, and its most cited pages) anddescriptor_words(the words answers use for and against a brand). check_brandreads a brand’s shares, its rivals and the answers that leave it out over the last 90 days, the same days that decide whether Parse has enough on the brand. It read 28 days, where a brand that few answers name showed 0%. Logged out it reads the public window, as before.check_brandsays “your brand” innextfor the client of a brand-bound key. Otherwise it uses the brand the workspace marks as its own, or its only saved brand. With several saved brands and none marked, it compares the brand checked with its own rivals.- A name lookup in
resolve_idsthat finds no topic, market or question says so and gives thematch: "meaning"call that finds the nearest ones. readgiven aqst_,top_ormkt_id names the call that opens it. It readsans_answers,pg_pages, anexp_export andjob_orcf_jobs.search_cited_pageswithlistsBrand: falsekeeps the same pages whose rows readfalse: those the page judge says do not list the brand, and those it has not judged whose stored copy does not name the brand (listsBrandBysays which). It kept only the judge’s, so a brand the judge had not reached got an empty list. A call reads up to about 200 stored copies to decide; past that it says how many pages it did not check.
November 17, 2026
get_playbookis retired: 16 tools instead of 17. Work is a kind of work on a target, logged withlog_action. What only Parse holds is on the reads:check_brandreturns the brand’s facts with their states, Parse’s read of its own site and its drafts, the brand’s own sites whose robots.txt keeps an AI crawler out, and, for a brand of the workspace’s, the standing instructions and background;readof a page returns its contacts, the pitch check and, where Parse advises waiting, the person’s plan, the held follow-up and the send as three ready calls. A call toget_playbookanswers until 2027-01-31 as the read of its target, withmeta.renamed.- Each brand can carry one background text, up to 4,000 characters: positioning, voice, what is launching, which rivals matter.
get_briefreturns it asbrandContextafter the standing instructions, andcheck_brandreturns the brand’s. A person writes it in Settings; an agent writes it withupdate_settingsabout, markedfrom: "assistant"until a person saves it. - An assistant sets the brief’s webhook address with
update_settingsitself: it applies at once, checked as delivery checks it (https on port 443, a public address), and shows in Recent changes with undo.needs_confirmnow means a higher budget or resuming spending. brief.alertsnames which changes can send the brief between scheduled sends, within each channel’s limit. The default is a contradicted fact.- Track a rival page by URL to check it at the workspace’s pace. A usable read costs $0.01. A change appears in the brief and can trigger a webhook on the next delivery check.
- A question tracked from a city keeps that city as its answers’ place:
filter.placeandby: ["place"]read it after the nightly build too, and the brief reports a move in one place by name. ChatGPT takes a country only, so its answers count under no city.
November 16, 2026
- The public view’s tools also answer a GET:
GET https://agent.parse.gl/v1/tools/<tool>with text arguments as query parameters, or every argument as JSON inargs. An assistant that can only follow a link reads Parse this way, and a GET’s reply carries the address of each next step that is a public read. Tools outside the public view stay POST, and a GET reads the public view whatever credentials come with it. - The “For agents” block on public pages lists its reads as these GET addresses, and says what may be reused: figures and tables are free to republish with attribution, a link to the page beside them, a standard followed link preferred.
- A reply about one brand, market or topic carries
context.cite: the parse.gl page to credit, a ready Markdown link, the window the figures cover and the same reuse terms. Signed in or not. In a page’s Markdown form each read is an ordinary link.
November 15, 2026
- The monthly budget is the one spending limit inside a workspace: the day’s share of it is gone. A paid call runs when it fits the month’s budget and the credit, and tracking no longer waits a day. Gone with it: the quote reason
over_day_share;dayShareUsd,leftTodayUsdanddayResetsAton quotes and inget_account, whose budget also dropsspentTodayUsd(today’s spend stays underspend); atrackquote’scostliestDay,costliestDayUsdandalreadyTrackedThatDayUsd; theaskargument on paid tools; and the request kindallow_more_today, whose open requests expired.needs_confirmnow comes fromupdate_settingsalone. - A paid call offered in
nextruns as offered: it says what it costs and carries noquote: true.quote: truestill returns a price and buys nothing, and a quote that fits keeps the call innext. A quote’sneedsYesis true when the call would pass the budget, the credit does not cover it or spending is paused (needsYesWhy:over_budget,no_credit,spending_paused); the reasonsover_0.25andtrialare gone. update_settingssets anypaceat once, a faster one included: the pace spends inside the budget. A higher budget, a new webhook address and resuming spending still wait for a person.ask_ai_enginesreturns the fresh answers Parse already holds that the workspace may read, free, each markedreusedwith the day it was collected, and buys only the runs they do not fill.fresh: truebuys every run. The reply’scheckssays how many were asked, reused and bought.- A job’s size:
ask_ai_enginesandtracktake up to 1,000 rows and 100 runs a call, inside the budget. A reply lists 25 rows and names the rest. - Full texts: the full text of an answer the workspace tracks or bought, or of a page it fetched, is free. The full text of any other public answer or page is free for the first 25 a UTC day, then $0.005 each, charged when the call runs inside the month’s budget. The daily text allowances are gone.
- A sampled or refused reply says how many rows matched, what is withheld and what opens the rest, with the price.
- Every public brand, market, topic, question and comparison page has a Markdown form at its address with
.md, and both end with a “For agents” block: the page as data, a first call, the access levels and the prices. - Suggested tasks are for brands tracked daily in a workspace with paid credit on record. For any other brand
get_briefandlist_actionsreturntasksWaiting: how many tasks Parse would suggest, one example and what opens the rest. A person can hold a task as an assistant can. update_settingstakesfact: an agent writes a fact about a brand with its source. It stays unconfirmed until a person confirms it or Parse finds it on the brand’s site, andget_playbookandcheck_brandreturn each fact’s state and day.- A key can be bound to one brand: it reads and writes that brand alone, and its spend is filed under the brand.
get_accountshows spend by brand, and a brand can carry a monthly allowance inside the one budget. - An agent with no account can start a workspace:
POST https://agent.parse.gl/v1/workspacesreturns a key and a claim link. Until a person opens the link and claims it, the workspace reads the public view, keeps itslog_actioncalls and buys nothing. update_settingstakeswake: an address Parse pings with a body that carries no data when something is new. An agent sets it without a person.get_accountshows whether one is set, by its host alone.
November 14, 2026
- 16 tools instead of 31: each tool that went is now an argument of the one that does its job.
readopens any id (ans_,pg_or a URL,exp_,job_andcf_) in place ofread_answers,read_pages,read_exportandcheck_jobs.resolve_idstakesmatch: "meaning"(wasfind_similar) andkinds: ["facet"]for the dictionary (waslist_facets);query_metricstakesby: ["question_group"](wasgroup_questions);search_answerstakesrows: "mentions"or"comparisons"(wassearch_evidence);search_cited_pagestakesview: "result_lists"or"result_list_diff"(wassearch_result_lists);ask_ai_enginestakesbaseline(wasrun_baseline);tracktakesown(wasclaim_brand);get_brieftakesview: "events"withsinceanduntil(waslist_events);get_accounttakesview: "receipt"(wasget_receipt);update_settingstakestopic(wassave_topic) andcorrection(wasflag_error). resolve_ids’trackedon a brand now means this workspace watches it withtrack;savedmarks a brand in the workspace’s brand list. A client that readtrackedas “in my list” should readsaved.- Prices, permissions and limits are unchanged: a call obeys what its old tool obeyed. The old names keep working until 2027-01-31, unlisted; a reply to one carries
meta.renamedwith its canonical arguments and private values redacted. Everynextnames the new tools.
November 13, 2026
- One price: Standard checks, live pages, judgments and rows outside your own data past the free ones cost $0.01; frontier checks on Gemini and Claude cost $0.05. Reading the last 90 days of what Parse already has is free: searches, numbers, excerpts, your own answers and pages in full, and 25 full texts of public answers or pages a day, then $0.005 each. A public answer older than 90 days costs $0.001 the first time your workspace counts or lists it, and is yours after. Searches, queries, full texts (5,000 a day, 500 until a first paid credit; since 2026-11-15 a public full text past 25 a day costs $0.005) and exports of the workspace’s own data cost nothing, and
meta.costCreditson a read is 0. A workspace with credit reads in full whatever is left of its budget, so no reply carriesmeta.budgetLimitor the level reasonbudget_reached; a paid tool’s refusal names the budget itself (budget_reached,over_day_share). - One monthly budget replaces the wallet’s daily and monthly caps, the agent limit and auto top-up’s own maximum:
budget.monthlyUsdinupdate_settingsandget_account, at most a fifth of it in a day. Past the month’s budget a paid call answersbudget_reached; past the day’s share its quote saysover_day_shareand tracking waits a day (budget_paused). Until the first top-up the trial credit is the budget, with no day’s share. See Spending, tracking and stopping. - Renamed, for a client that matches on them: the error
spend_cap_reachedisbudget_reached;meta.spendCapismeta.budgetLimit; a quote’swithinAgentLimitiswithinBudgetand its reasonover_agent_daily_limitisover_budgetorover_day_share; the brief’sspend_near_capisspend_near_budget;update_settingsno longer takesautonomy,spendCapUsd,dailyCapUsd,monthlyCapUsdorautoTopUp.monthlyMaxUsdand names each when sent. New: the eventauto_topup_skipped(once a month when auto top-up would pass the budget), and atrackquote’scostliestDay,costliestDayUsdandalreadyTrackedThatDayUsd.
November 12, 2026
- Parse’s first MCP server is retired, with its 37 tools (
parse_check_brand,parse_next_tasksand the rest). Each of their pages now leads to the tool onhttps://agent.parse.gl/mcpthat does the same job. Its address,https://mcp.parse.gl/mcp, forwards tohttps://agent.parse.gl/mcp: a connection made to it reaches the same 16 tools and signs in the same way. - A page for each of the 16 tools on
https://agent.parse.gl/mcp, with its arguments, price and a call over REST and MCP, from the registry’s own definitions; every error code with its status; paging, REST and the CLI. - New guides: Your first session, Spending, tracking and stopping, Reading the numbers and Agencies and local businesses.
- Tool behaviour these pages document:
read_answersreturns only thefieldsasked for, so more answers fit a reply; a later page of a metered search is free; an assistant cannot change the wallet’s spend caps or auto top-up, in either direction; every error sayscharged: 0;list_actionstakespitched. read_exportreads an export’s file through the service, page by page, free.tracktakesplacesandlanguage;get_brieftakesbrandandpeekand marks the brief read for the key that read it;list_eventstakesbrand,fromandto;search_cited_pagestakeslistsBrand;check_jobstakeswait.- Get started goes client by client, with the sign-in and access screens. The starting credit comes with the workspace, not with its first key.
- Plans and limits leads with pay as you go, gives the agent limit’s starting values and adds team roles. Data and trust says what the agent service records and how to delete a workspace.
November 11, 2026
- Counts over May to September 2026 can differ from a reading before 11 November: about 120,000 older answers returned to the data.
context.dataChangeslists such days. - These docs teach one address,
https://agent.parse.gl/mcp: a key from Connect, andget_briefas the first call. The older server,https://mcp.parse.gl/mcp, kept working, its tools documented under its own headings.
October 30, 2026
- These docs, with every MCP tool on its own page and every page as Markdown (
.mdorAccept: text/markdown). parse_get_topicreturns the topic’s first 25 brands by default;brandLimitasks for up to 100, andbrandsTotalsays how many the topic ranks.- The MCP server, now retired, published a server card at
https://mcp.parse.gl/mcp/server-card, and parse.gl an AI catalog and a skill index under/.well-known/. - The workspace screen tools have titles.