Veqen: AI cost monitoring and protection

action support (Veqen support's requests to view the dashboard), approve, deny or revoke (grant_id; owners), apps (AI assistants the person connected with OAuth) or disconnect (family_id).

One action on an incident from list_incidents: shut_off (its key, as shut_off_key; confirmed as there; for an agent spend incident, freezes its card), resolve, expected ('this was us', quiets the key's rules for a week), rotated (the key was replaced everywhere, proof for a refund claim) or claim (start a refund claim, for amount_usd or the incident's amount). Confirm first.

Records a provider's monthly invoice total (period YYYY-MM), for providers whose billing Veqen can't read. Month-end books it. Needs an owner, admin or finance member.

Records credits from an AI provider. kind is prepaid (paid for, an asset) or promotional (free). Dates are YYYY-MM-DD. Needs an owner, admin or finance member.

A coding tool's invoice for a month, as lines: {kind: seats|usage|credit|other, amount_usd, quantity (seats billed)}. Credits are given as positive amounts. Checked at once.

action get (Veqen's answer as it stands, the approval and the matched transaction) or cancel (the agent didn't go through with it: it stops counting toward the month and a pending approval closes).

action list (state pending, approved, denied or expired) or approve and deny (approval_id, with a note): a person's answer to an agent that asked to spend past its mandate. Confirm with the user before answering.

action list (every card, whose agent it is, status, limit, freezes waiting), assign (card_id to agent; '-' for a person's card; '' to read it from the card), freeze or unfreeze (card_id; incident_id to record it there; needs freeze access; with two-person shut-off on, a freeze waits for a second person), confirm (freeze_id, the second person) or limit (card_id, limit_usd a month, set at the platform). Confirm with the user before changing a card.

view overview (every agent's spend this month, how much an action explained, incidents and approvals waiting, the platforms connected), agent (one agent: mandate, cards, latest transactions, actions and approvals), transactions (agent and match filters), actions (agent and state filters) or report (period YYYY-MM: the month's agent spend controls report for the close and the auditor).

action list, create (name, expires_days 30, 90 or 365, allowed_ips: a send-only key for an app), rotate (key_id; the old key works a day) or revoke (key_id). Full-access keys aren't made or rotated here, so an assistant can't keep access after it's disconnected. A new key is in key, shown once: tell the user to store it as a secret, never in code.

Freezes a finished month's chargeback (YYYY-MM); the close books its work and coding agents. Confirm first.

Approves and freezes a month's journal entries (period YYYY-MM). Confirm with the user first. Needs an owner, admin or finance member.

action status (plans, subscription, this period's spend, the next renewal), checkout (work_agents and/or product_inference, yearly: a Stripe link for the person paying to open, since card details only go to Stripe), portal (Stripe's page for the card, invoices and cancelling), set_interval (interval month or year), drop (product work_agents or product_inference), redeem_offer (partner_code on a running subscription; checkout takes partner_code too), jvzoo (purchases bought with your email) or activate_jvzoo (receipt). Needs an owner or admin. Confirm before changing a subscription.

action list (the platforms Veqen reads for agents: Stripe Issuing, Ramp, Brex, Mercury, what each needs, and which are connected), connect (platform, fields: a read-only credential, checked by listing cards), disconnect, freeze_access (platform, fields: the separate credential that may freeze cards and set limits) or remove_freeze_access.

action role (member_id from list_members, role), remove (member_id) or cancel_invitation (invitation_id). Owners alone add or remove owners. Confirm first.

action set_access (provider and key: a credential allowed to change key status there; it's never shown again), remove_access (provider), set_policy (rule and policy: alert, preview or disable; only dormant and owner_gone can shut keys off) or set_options (two_person_shutoff, share_patterns). Needs an owner, admin or security member. Confirm first.

A finished month's work and coding agent cost by cost center, team, person and tool, matched to each tool's invoice, with its journal entries. Work and coding agents Enterprise.

Call before an AI agent spends money. Answers allow, approval (a person decides; call again with the same action_id, or agent_action, to see their answer) or deny, from the agent's mandate, with the reasons and what it has left this month. The action is recorded either way and matched to the card or account transaction when it appears.

Each person's work and coding agent cost for a month (YYYY-MM, this month by default), tool by tool at the company's contracts, optionally for one team. People see what they'd see in the dashboard.

Each coding tool for a month (YYYY-MM): connected or not, its cost at the company's contracts, its plan, and its invoice checked against what Veqen expects.

Cost of the same call on several models, per call and per month: tokens for text, embedding and realtime models, units (images, seconds, characters, queries or pages) for the rest.

The second person's confirmation of a waiting shut-off (an id from waiting_shut_offs). Confirm first.

Connects a provider's billing or a coding tool's usage with the fields list_connections names (an admin key, usually); provider stripe with a key field connects revenue. Brings in 30 days to check, up to a minute. Confirm first: it stores a credential.

Provider credits with a balance that expire within the given number of days.

Deletes a budget by its id from list_budgets. Needs an owner or admin.

Deletes a bill, credit grant, price, saved view, vendor agreement or savings baseline by its id from the matching list. Confirm first.

Disconnects it and deletes the billing Veqen brought in from it. Confirm first.

Re-prices the organization's real token volumes on another model (standard rates).

Compares two periods and names the models, features and customers that moved most. The baseline defaults to the same number of days just before.

Month-end and quarter-end spend forecasts with an 80% range (spend so far plus each remaining day at its weekday's recent average), for the organization or one customer, feature or project.

action get (the latest audit: running, or its report), run (reads 30 days from every connection; ready within the hour), share or unshare (run_id: a link anyone signed in to the organization can open for 30 days).

Freezes how the organization paid the providers and seat tools in covers during base_start to base_end (YYYY-MM-DD), before new terms start on changed_on, so each later month shows what the new rate saved apart from usage changes. Do it before the new terms start. It can't be edited afterwards, only deleted. Confirm first.

The connected organization, its plans (work and coding agents and the product-inference band) and this billing period's usage.

Where alerts go (email, Slack, webhook) and the latest alerts sent.

What one customer cost: total, by feature and by model, and margin when Stripe revenue is connected.

One incident by its id from list_incidents: its timeline, each hour it covers with the reasons, and its claims.

Revenue from Stripe against AI cost per customer, least profitable first.

Rates for one model of any kind. Token rates are USD per million tokens: input (in), output (out), cache read (cr), cache write (cw, cw1h) and audio (ain, aout), with batch, priority and long-context tiers. Models billed per image, second, character, query or page have a unit rate (unit), and may have size, quality or resolution variants (v:...).

One agent run, conversation or job, by its run id (get_runs lists the costliest): its total, how long it took, and each AI request in order (up to 200) with its model, tokens and cost.

Agent runs, conversations or jobs (calls tagged with the same run id, or one OpenTelemetry trace) that ended in a date range: how many, their total, average, median and p90 cost, and the most expensive.

Ways to spend less, from the last 30 days: prompt caching, the batch API, the same open model on a cheaper host, and cheaper models from the same provider, each with an estimated saving a month. Test a cheaper model's answers before switching.

AI spend for a date range: totals, an optional breakdown by one dimension, filters on the others, and this month's run rate and forecast.

One spending spike, by the id list_anomalies gives: what it cost against a usual day, and the models, features, customers and users that moved most against the same weekday a week before.

kind tool_usage (a coding tool's usage export; tool), teams (an HR export: email, team, cost center, valid from), people (tool, name or id, email), bills (a cloud billing export summed by month; provider, and match to keep only rows whose description has it) or rate_card (your prices). Pass the file's text as csv.

What a provider's refund process asks for: the key and its changes, then hour by hour (or day by day for an overcharge) what the provider billed against what the company's own code made, as columns and rows.

Emails an invitation to join the organization with a role (admin, member, finance or viewer). Confirm the address with the user first. Needs an owner or admin.

action ask (for the signed-in person: amount_usd a month and a note), approve (request_id at amount_usd a month, which becomes the person's limit in every tool) or deny (request_id).

Engineers' and teams' monthly work and coding agent limits the person may see, and the requests for a higher limit they answer or made. Set a limit with set_budget, scope engineer or team.

Spend spikes in the last days: a day when the organization, or one customer, user, feature, project or model, cost at least twice the median of the same weekday over the four weeks before.

Monthly provider bills entered or imported, next to what Veqen recorded for the same month.

Every budget, what's been spent against it this period, and who's closest to the limit.

Refund claims on incidents, newest first, with what was claimed and credited.

Each AI provider and coding tool whose billing or usage Veqen can read, the fields it needs and how to get them, and whether it's connected and working.

Credits held with AI providers: how much is used, what's left, and when each expires.

Incidents in one state (open by default): spend nobody on the team made, runaway or possibly stolen keys, and overcharges, with the key, amount and the reasons each was flagged. act_on_incident shuts the key off or settles it.

Everyone in the organization with their role, and pending invitations (for owners and admins).

Current models in the price book, cheapest first, for one provider or one kind: text, embedding, realtime, image, video, speech, transcription, rerank or ocr.

Every key Veqen sees at the organization's AI providers, with its owner, status and the flags that make it risky (unused for 30 days, its owner left, open to any API). Needs an owner, admin, finance or security member.

Discounts and prices the organization set, used instead of published prices.

Matches a name or id a coding tool reports (from teams_and_people, show matching) to a person's email.

A month's journal entries for AI costs (period YYYY-MM): approved and frozen, or what they'd be if approved now, and whether every connected provider has reported. Needs an owner, admin or finance member.

action policy (how costs are booked; send policy, in the shape it comes back, to change parts of it), reopen or true_up (period YYYY-MM: post a correction to the invoice). Needs an owner, admin or finance member.

action get (name, plans, your role, the Get started checklist), rename (name), set_track (track: coding, audit or sdk), audit_log (who changed what; paid plans), leave (you leave; your full keys stop working), delete_code (emails the owner a code) or delete (owners: confirm set to the exact name, and the challenge and the code the owner reads you from that email; can't be undone). Confirm before leave or delete.

One person's work and coding agent month (an email, or tool:id for someone not matched yet): each tool at the contract, day by day, their limit and their requests for more.

Like simulate_workload, but at the organization's own prices: its discounts and the prices it set.

Sets a price per million tokens for that provider's models from a whole month's bill (bill_id from list_bills) and what they served.

Which providers Veqen has shut-off access for, what each rule does when it fires, and the two-person shut-off and pooled-patterns options.

Tracked spend against what OpenAI and Anthropic billed: capture rate, untracked spend and the models that differ most.

Actions agents took with money, told after the fact (each {agent, amount, currency, payee, purpose, rail, kind, card, run, action_id, at}), up to 1,000. Matched to card and account transactions. Idempotent by action_id.

action list, or save (name, group_by, start, end, filters, and schedule to email it) a breakdown for the signed-in person.

Coding-tool seats nobody used for 30 days, and seats people who left still hold, with what each costs a month.

What an agent may spend without a person (agent '*': the default for agents without their own): per_action_usd (refused above), monthly_usd and approval_over_usd (a person approves above), merchants (payees it may pay on its own; empty: any), blocked (never), paused, auto_freeze (freeze its cards when an incident needs action; the protection role). Only the fields given change; a limit of 0 removes it. remove true deletes it. Confirm with the user first.

Records an AI vendor agreement, or replaces one's terms (agreement_id from vendors view agreements). Only the vendor and dates are needed; commitment, tiers, notice and overage are optional. Reminders go out before the notice date. Confirm first.

channel slack, protection_slack (incidents and misuse go there instead) or webhook (a new signing secret comes back once), with url; an empty url turns it off. channel test sends a test alert everywhere.

Creates a budget on AI spend. scope is org, customer, feature, user or project; value '' gives every customer (or feature, user, project) its own limit. mode 'alert' notifies at 50/80/100%; 'block' also makes the SDKs refuse matching requests at the limit. scope engineer (value an email, '' or 'team:') or team sets a work and coding agent limit, which 'block' pushes into each tool that takes one. Needs an owner or admin.

What the company pays a coding tool from a day (YYYY-MM-DD): plan, seat_usd, seats, unit_usd, discount_pct, included_usd. Costs are priced at it.

Sets a provider key's owner (an email or a service's name) and the coding agent it's for ('' reads it from the key's name, '-' none).

Sets the organization's own price for a model (per million tokens, or per unit for image, video, audio and document models), or with model '*' and discount_pct, a discount on a provider. New AI requests use it. Needs an owner, admin or finance member.

Whether Claude Code's telemetry names people by their email (true) or by account id, matched by hand (false).

Puts a person (email) in a team and cost center from a day (YYYY-MM-DD); their earlier days keep their team.

Sets the teams a member (user_id from list_members) manages: they see those engineers' cost and hear about spikes.

Turns a provider key (an id from list_provider_keys) off at its provider, or back on (action restore). Needs shut-off access for that provider. Where the provider can only delete keys, it fails until confirmed is true: say so to the user, since that can't be undone. With two-person shut-off on, the result is 'waiting' for a second owner, admin or security member (confirm_shut_off). Confirm with the user first.

A month of calls across any mix of models, at list prices: each line's cost per call and per month, the total, a projection with monthly growth, where each line ranks among every current model of its kind, the same model on other hosts, and cheaper models from the same provider. Usage is per call.

action get, link (domains: your company's email domains), portal (intent sso, domain_verification or dsync: a WorkOS link for whoever administers the identity provider), refresh (read domain verification) or save (enforced, default_role, groups: directory group to role). Changes are for owners. Confirm before requiring SSO.

show teams (who is in which team and cost center, and who manages which), matching (tool names not matched to an email, with likely emails), contracts (what the company pays each tool) or privacy (how people are named).

Marks a claim submitted, credited (with credited_usd, which then counts in provider credits and the close) or denied.

view rates (what each AI provider and seat tool really costs as a share of list price, whether the agreed rate is on the bill, and seats nobody used; start and end YYYY-MM-DD, the last 30 days by default; use before saying a vendor overcharged or whether a deal is good), agreements (every vendor agreement: notice dates, commitment pacing and tiers; state filters), agreement (one, by agreement_id, with spending day by day), baselines (proven savings: each frozen baseline's newest month) or savings (baseline_id and month YYYY-MM: what the new rate saved apart from usage changes, and whether the agreed rate is on the bill).

Shut-offs waiting for a second owner, admin or security member, with two-person shut-off on.