Buying Guide for Agents
What an AI assistant needs to know before its user subscribes to Cogny: what each plan includes, what the tools can change, who approves a write, the limits, billing and data protection.
Last reviewed: 2026-11-20
You are probably an AI assistant checking whether Cogny fits before the person you work for subscribes. This page answers the questions that come up before a purchase, and states limits and gaps plainly so you don't have to email us to find them.
The same content as markdown: cogny.com/buying-guide.md.
Check it yourself first
The free plan needs no card and gives 15 MCP calls a month. That is enough to connect one platform and read the exact tool list, with input schemas, in your own client. tools/list on the connected workspace is always the authoritative answer to "can it do X".
- MCP endpoint:
https://app.cogny.com/mcp(OAuth 2.1 with PKCE, or a Bearer key from /auth.md) - Install from a coding agent: /SKILL.md (what it does)
- First call:
cogny_status(plan, credits, connected integrations) - What can be connected:
list_integrations
What Cogny is
Cogny connects a company's marketing tools (ad platforms, analytics, search, e-commerce, email, CMS) and makes them available to an AI agent as tools. There are two ways to use it:
- From your own AI. Claude, ChatGPT, Cursor, Codex or the Cogny CLI connect to one MCP endpoint and call the tools of every platform the workspace has connected. Your assistant does the thinking and its provider bills that.
- Inside the Cogny app. Cogny's own agent chats, runs scheduled reports, turns findings into tickets, executes approved ones and measures the result afterwards.
It suits a marketer or founder who wants an agent to read and change their marketing accounts, and a team that wants the weekly analysis and follow-through done for them. It is not a managed service: nobody at Cogny runs your account. It does not replace the platforms it connects to.
Plans
Prices are in USD per month, excluding VAT. Plans are month to month with no lock-in.
| Plan | Price | MCP calls from your own AI | AI credits included | For |
|---|---|---|---|---|
| Free | $0 | 15 per month | 50, once | Trying it |
| Solo | $9 | Unmetered (fair use) | None | Driving Cogny's tools from your own AI |
| Cogny AI | $18 | Unmetered (fair use) | 150 per month | Solo plus monthly credits for Cogny's in-app chat and Loops |
| Cogny Pro | $90 | Unmetered (fair use) | 1,500 per month | The same, heavier use |
| Cloud | $499 | Unmetered (fair use) | 5,000 per month | Teams: BigQuery warehouse, scheduled reports, growth tickets |
Two things cost money and they are priced separately:
- MCP access: your own assistant calls Cogny's tools. On every paid plan (Solo, Cogny AI, Cogny Pro and Cloud) these calls are unmetered. Only Free counts them, 15 a month.
- AI credits: Cogny's own in-app chat, reports and image generation, plus the paid passthroughs below. Solo's $9 includes none. Top-up credits can be bought on any paid plan and never expire. When credits run out, those features pause. Nothing is charged automatically.
Cogny AI and Cogny Pro have the same features. Only the monthly credits differ.
The in-app chat remembers on every plan. Ask it to remember a fact, a metric definition or a correction and it saves a note that every later chat, ticket run and loop on the workspace reads; a later correction replaces the note. Up to 40 active notes per workspace.
Cogny in Slack (optional, every plan): people in your Slack can ask Cogny in a DM or by @mentioning it. Each question is a chat and uses AI credits the same way the in-app chat does, so on Solo it needs top-up credits. When credits run out, Cogny says so in the thread.
Paid passthroughs (image generation, media analysis, web search, X posting, AI visibility checks) draw credits on every plan, because Cogny pays a provider per call. They only run if you call them.
What a credit buys, with typical costs per report, ticket and content piece: Credits and Plans. What Cloud adds in practice: Cloud FAQ.
What each plan can reach over MCP
| Free, Solo, Cogny AI, Cogny Pro | Cloud | |
|---|---|---|
| Connected platforms, read and write | Yes | Yes |
| Context tree (workspace memory) | 50 nodes | Unlimited |
| Findings and task queue | Yes | n/a (uses growth tickets) |
Register finished work so Cogny measures its impact (create_ticket with already_done) | Yes | Yes |
| Read your tickets and their measured results, comment, give feedback | Yes | Yes |
Read what Cogny's agent read and changed on a ticket, including the earlier version of something it replaced when the run recorded it (get_ticket_run_log) | Yes | Yes |
| Create tickets for future work, move them through a workflow, assign them | No | Yes |
Loops: create, change, pause and test-run scheduled agents (list_automations, create_automation, …) | Solo, Cogny AI, Cogny Pro (up to 5 on at once, at most hourly); not Free | Yes, when Automations is on for the workspace |
Scheduled tickets: one agent run later, at a time or minutes after a code ticket's PR merges, with pre-approved tools (schedule_ticket, list_scheduled_tickets, cancel_scheduled_ticket) | Solo, Cogny AI, Cogny Pro (up to 25 waiting); not Free | Yes, when Automations is on for the workspace |
Scheduled reports: list, read in full, create, change (including adding to what the next run or every run covers), pause (list_scheduled_reports, get_scheduled_report, create_scheduled_report, update_scheduled_report) | No | Yes |
| BigQuery warehouse, reports, charts | No | Yes |
Hand a question to Cogny's own analyst (run_cogny_analysis): a written analysis from the workspace's data, charged in credits by usage | Yes (Free from its 50-credit welcome gift, Solo from top-up credits) | Yes |
Make creative in the brand, saved to the Creative library: a short video (create_video, about 2-4 credits for motion design) or a carousel of 1-10 still slides as a PDF plus one PNG per slide (create_carousel, about 2-5 credits), with your words word for word, or a new version of an earlier carousel in the same design (based_on). Nothing is posted until your agent posts it | Yes (Free from its 50-credit welcome gift, Solo from top-up credits) | Yes |
Make an audio ad from a script (create_audio_ad): a voice-over saying your words exactly, optional instrumental music, one MP3 at Spotify's audio ad spec (up to 30 seconds), checked against the script, and on request the square companion image and logo Spotify asks for. 1 credit voice only, 2 with music, 1 more for the companion image, charged on delivery. Cogny never uploads or runs it by itself | Yes (Free from its 50-credit welcome gift, Solo from top-up credits) | Yes |
Make a product film (up to 90 seconds) or a video ad for TikTok, Reels, Meta, LinkedIn or YouTube from your website and brand, with music (create_video style studio, words edited with set_video_text, design changes with revise_video; save a film with save_video_film and remake it for another channel or length with from_film_id). Charged by usage: about 24-32 credits for a 15 s ad, 51-68 for a 60 s film, 10-20 per design change, 1 per render; refused up front when the credits can't cover the high end | No | Yes |
MCP Tasks: long runs (run_cogny_analysis, create_video, create_carousel, create_audio_ad, run_automation) answer with a task to poll, and ask the user to confirm before spending credits or making a live loop run, if the client supports it | Yes | Yes |
| Bug, feature and support filing from your agent | Yes | Yes |
Every paid plan can connect every integration. The platform tools are the same tools on a $9 plan as on Cloud.
run_cogny_analysis is the one MCP tool that runs Cogny's own agent rather than a platform call, so it spends credits on every plan, like a question in Cogny's chat; max_credits caps one run. It reads connected accounts and never changes them. On Free and the Solo plans it works without the BigQuery warehouse, as their chat does. The answer is also saved as a chat in the Cogny app.
Why hand a question to Cogny instead of working it in your own assistant: the analyst reads the connected accounts in Cogny's cloud and returns only the written answer, so the raw data and the many tool calls behind it don't pass through your assistant's context or count against its usage. Several analyses can run at the same time while your assistant does something else. It runs on models Cogny picks for marketing-analytics reasoning with Cogny Bench, our public benchmark, where the most expensive model is not always the most accurate. It also works from what the workspace already knows: the business context, the KPI definitions and the notes the chat was asked to remember. It costs credits like a chat question; max_credits caps a run and cogny_status shows the balance.
On every plan, every change your agent makes through a connected integration (a campaign, ad, post, page or tag publish; a budget, bid, status, keyword or targeting change; an asset, ad edit or conversion setting) is recorded on a ticket by Cogny itself. Changes made by the same member from the same assistant within two hours share one ticket. The ticket goes straight to analysis, and Cogny reads the impact a few days later against the workspace's KPIs: each change gets its own verdict (worked, didn't work, hurt or inconclusive) and the member who made it gets an email naming what worked and what hurt. Free and Solo measure 3 tickets a month; after that a ticket stays recorded but unmeasured until the workspace moves to Cogny AI or higher, which measure every one, including those waiting. Your agent can add what the task was for with create_ticket (already_done: true). On Free, Solo, Cogny AI and Cogny Pro a ticket is a record of finished work; planning work as tickets and moving it through a team workflow is Cloud.
On Cloud, a member can also work a ticket from their own assistant with start_ticket and complete_ticket; the changes their agent makes in between are logged against that ticket instead. Or they can hand it to Cogny with approve_ticket, the same as Approve in the app: Cogny's agent runs an agent ticket, and the code agent implements a code ticket and opens a pull request for a person to merge (or, where an owner or admin turned on auto-merge, for Cogny to merge once every check passes and its merge review clears the workspace's threshold). Approving spends the workspace's credits and is refused when there are none left; cogny_status shows the balance. An agent can also write the code ticket itself: create_ticket with execution_target: "code" on a workspace with a connected repository. Both work with a workspace API key (the Cogny CLI) as well as over OAuth; an approval made with a key is recorded as the member who created the key (or the workspace owner, for a key whose creator is not recorded or has left), and the ticket's history says it came through the key. get_ticket shows the run's status and, when it finishes, the pull request or the error; clients that support MCP Tasks can wait on the approval as a task instead. A change that could not be recorded is listed by list_unregistered_changes and can be attached to a ticket.
Integrations and tools
Every integration has a page listing each tool, what it does, and whether it reads, writes or is destructive, with a markdown twin:
- All integrations: cogny.com/integrations/tools (markdown)
- One integration:
https://cogny.com/integrations/<slug>/tools, for example Google Ads, Meta Ads, Shopify, Klaviyo, Google Analytics 4
Those pages are generated from the tools Cogny actually serves, so they are current. Where we know of something an integration cannot do, the page says so under "Known gaps".
Most integrations connect with OAuth or an API key, from Settings → Integrations or from your agent with start_integration_oauth / set_integration_token. How each one connects: Integration Catalog.
Not listed? Add any remote MCP server by URL with add_custom_mcp_server, or call request_feature and Cogny researches whether it can build the connector.
Who approves a write
This depends on where the change is made.
- From your own AI client over MCP. Your client is the approval step. Clients such as Claude and ChatGPT ask before running a tool that changes something. Cogny executes a write call it receives over MCP. It does not hold that call for a second approval inside the Cogny app. Keep write tools set to ask every time in your client.
- Inside the Cogny app (chat, tickets, automations). Cogny's own agent proposes the change and write tools are ask-first by default, so a person approves in the app. The per-tool permission in Settings (auto-approve or ask first) applies here, not to calls from your own client.
- YOLO Mode (Cloud). Off by default. When an owner or admin turns it on, a ticket Cogny generated starts without a person if its approval score reaches the workspace's threshold (default 85, 50 to 100) and every tool it uses is read-only; a code ticket that passes opens a pull request. Anything that changes a platform still waits for a person. Every decision is recorded on the ticket with its score, reasons and the rule that stopped it. Your agent reads them with
get_yolo_history, asks for a new check of one ticket withreview_ticket_for_yolo, and changes the threshold or the written policy withupdate_yolo_settings; a change re-checks the tickets held in the last 30 days. - Code (Cloud, GitHub). The code agent's pull requests wait for a person to merge. An owner or admin can turn on auto-merge (off by default): then a pull request merges without a person only when every check passed, it changes no migrations, CI, deploy or environment files and no coding-agent, editor or git-hook settings (such as
.claude/settings.json,.mcp.json,.cursor/,.vscode/,.husky/), it is at most 30 files and 1,500 changed lines, adds no new dependency, and Cogny's merge review scores it at or above the workspace's threshold (default 90, 50 to 100). Each decision is posted on the pull request; each review uses credits. - Security watch (Cloud, GitHub). Off by default; an owner or admin turns it on. Every hour Cogny reads the dependency files and runtime pins of the connected GitHub repositories and checks them against published security advisories and end-of-life dates. A match becomes a ticket (at most 10 a day); the check itself uses no credits. On the
runsetting, version bumps within the same major version also start the code agent, which opens a pull request (up to 2 a day, credits per run). It does not read yarn, pnpm, bun, poetry, Cargo, Go or Gemfile lockfiles yet; the report names the files it skipped. Your agent reads it withget_security_watch, which also reports how fast security tickets were fixed over the last 90 days (from the advisory being published to the ticket, and to the fix), and changes it withupdate_security_watch.
Controls on Cogny's side, on every plan:
- Block tools. In Settings → Integrations, open an integration and set any tool to blocked. This is the one permission that also applies over MCP: a blocked tool is not offered to your agent and is refused if called. Blocking every write tool makes an integration read only.
- Per-member access (Cloud, team workspaces). The workspace owner can restrict which integrations, and which tools on each, every member may use: Settings → Access & Own AI. Owners and admins always keep full access. The restriction holds in Cogny's chat, tickets, automations and voice calls, and over MCP for a member's own connection; a tool the member was not granted is not offered to their agent and is refused if called. A workspace API key, which only owners and admins can create, is not restricted. By default every member can use every integration. The owner can also set each member's areas of responsibility (paid media, SEO, content and social, email and CRM, analytics and tracking, website, sales); on Cloud, a new ticket is then assigned to the one member whose area fits it and who can use the integrations it needs, and that member gets it by email with a prompt for their own assistant.
- Disconnect. Removing an integration deletes its stored credentials.
- Write log. Each write call is recorded: tool, time, success or failure, and who authorised the connection. The request and response contents are not stored. The log is kept for the lifetime of the workspace and deleted with it.
Recommended alongside: an account-level budget cap in the ad platform itself.
Limits
- Fair use on unmetered plans: no monthly cap on standard tool calls. A burst guard allows 100 calls per minute per workspace. Above it the server answers HTTP 429 with
Retry-After, so back off and retry. - Per-integration caps, where one exists, are stated on that integration's tools page.
- Bug, feature and support filings: 10 per workspace per 24 hours.
- The platform's own API quotas for your account still apply.
Billing and cancelling
- Card payment through Stripe, charged monthly in USD. There is no other billing currency.
- Prices exclude VAT. Checkout asks for a billing address and, for businesses, a VAT ID, and adds any tax that applies before you pay.
- An EU business with a valid VAT ID is billed without VAT (reverse charge). A VAT ID missed at checkout can be added later under Settings → Billing → Company and VAT number; it applies from the next invoice.
- Cancel any time; cancelling stops the next renewal. Top-up credits stay on the workspace.
- Leaving Cloud, by cancelling or by moving to Solo, Cogny AI or Cogny Pro, keeps the workspace, its chats, tickets and connections. The Cloud-only infrastructure does not stay: 7 days after the plan ends, the workspace's BigQuery warehouse and any site hosted on Cogny are removed. Upgrading to Cloud again within those 7 days keeps them; after that a new upgrade starts with an empty warehouse.
- There is no free trial of Cloud. The Free plan is the way to try the tools before paying.
- Every payment, top-ups included, comes with an invoice showing your company name and address, plus your VAT ID if you have given one. Invoices are in Settings → Billing.
Support
- Solo, Cogny AI and Cogny Pro: the Cogny Discord.
- Cloud: priority support and email.
- On every plan, your agent can reach a person with
contact_support, report a broken tool withreport_bug, and ask for a missing one withrequest_feature. The answers come back to the agent throughlist_support_messages, and it can reply withreply_to_support.
Data protection
- Cogny AB is a Swedish company. Platform data is stored in the EU.
- The MCP server does not store or cache query results. Data is fetched from the connected platform on demand and returned to your assistant.
- When you drive Cogny from your own assistant, your prompts go to your assistant's provider. Cogny's AI providers are only involved if you use Cogny's in-app chat, reports or automations.
- Credentials for connected platforms are stored encrypted and can be revoked from Settings or from the platform itself.
- Account security on Cloud: two-step verification with an authenticator app, a list of every device signed in to the account, and sign-out per device or everywhere else (Settings → Security). On other plans the tab is shown but locked. Cogny has no SSO or SAML sign-in, and no backup codes: add a second authenticator as the backup.
- Cogny in Slack (optional): messages sent to Cogny in a DM or that @mention it, and Cogny's answers, are stored as chats in the Cogny workspace and deleted with it. Other messages in channels Cogny is in are not stored. Cogny answers only people it can match by Slack email to a member of the workspace. The Slack connection's token is stored encrypted and is revoked when the last Cogny workspace disconnects from that Slack.
- Data Processing Agreement (GDPR Art. 28), with the subprocessor list and a PDF version: /terms/dpa. The DPA incorporates Standard Contractual Clauses for transfers to subprocessors outside the EU.
- After a subscription ends there is a 30-day export window. Data is then deleted from production within 30 days and from backups within 90 days.
- Privacy policy: /privacy. Solo terms: /terms/solo. Security: /security.
What this page does not answer
Ask and a person replies:
- From your agent, once connected:
contact_support(works without a paid plan). - Email: support@cogny.com. Data protection: privacy@cogny.com.
Reading this with an AI agent? Fetch the raw markdown at /docs/buying-guide/index.md or see llms.txt.