Cogny MCP Server

What is the Cogny MCP Server?

Cogny's MCP (Model Context Protocol) server lets you connect Claude directly to your marketing data stack. Once connected, Claude can query your BigQuery warehouse, read Cogny's AI-generated reports, manage growth tickets, and access your organizational context — all through natural conversation.

Server URL: https://app.cogny.com/mcp

Protocol: JSON-RPC 2.0 over HTTPS (MCP spec 2024-11-05)

Authentication: OAuth 2.1 (authorization code flow with PKCE) — no API key needed; your MCP client handles the sign-in, and you pick which warehouse to authorize.

Tip: You can also use a warehouse-scoped URL of the form https://app.cogny.com/mcp/w/<your-warehouse-id> to skip the warehouse picker. Copy the exact URL for your workspace from Settings → MCP inside the app. Both the generic and scoped URLs work everywhere.

Quick Setup

In Claude.ai / Claude Cowork

1. Go to Settings → Connectors → Add custom connector
2. Enter the name cogny and paste the server URL: https://app.cogny.com/mcp
3. Click Add — Claude will prompt you to log in
4. In the popup, select your warehouse and click Authorize
5. Start chatting — Claude now has access to your marketing data

In ChatGPT

1. Go to Settings → Apps → Advanced settings and enable Developer mode
2. Click Create app, enter the name Cogny, and paste the server URL: https://app.cogny.com/mcp
3. Set Authentication to OAuth, check the risk acknowledgment, and click Create
4. In the popup, select your warehouse and click Authorize
5. To use it: click + in the chat box → More → Cogny

ChatGPT may prompt you to authorize twice on first use — this is normal.

In Claude Code (CLI)

The fastest route is to let the agent install Cogny itself: paste Read https://cogny.com/SKILL.md and follow the setup instructions. See Install Cogny in Your AI Agent for what that does. To add the MCP server by hand instead:

Run this in your project directory — Claude Code handles the OAuth login automatically:

claude mcp add --transport http cogny https://app.cogny.com/mcp

Or, if you already have a Cogny API token, skip OAuth with a bearer header:

claude mcp add --transport http cogny https://app.cogny.com/mcp --header "Authorization: Bearer YOUR_TOKEN"

In Codex CLI

Run this in your project directory — Codex detects OAuth and opens the authorization flow automatically:

codex mcp add cogny --url https://app.cogny.com/mcp

If the server is already added and you need to authenticate again, run codex mcp login cogny.

Via the API (Anthropic MCP Client)

Use the Anthropic API with the mcp-client-2025-04-04 beta header to connect programmatically against https://app.cogny.com/mcp. See Anthropic's MCP documentation for details.

Available Tools

All tools include MCP safety annotations (readOnlyHint, destructiveHint, etc.) so Claude understands which tools are safe to call without confirmation.

BigQuery Analytics

Query your connected data warehouse directly.

Example prompt: "What are my top 10 landing pages by conversion rate this month?"

Reports & Visualizations

Access Cogny's AI-generated marketing reports and render interactive charts.

Example prompt: "Show me this week's SEO performance report"

Growth Tickets

Manage the actionable opportunities Cogny identifies from your data.

Example prompt: "What are my open growth tickets? Assign the top-priority one to me."

Leaving feedback on a ticket

add_ticket_feedback records what an agent thinks of a ticket. Cogny reads it back when generating the next batch, so a ticket marked already_done or incorrect_data stops recurring. Always explain the verdict — the reasoning is what future generation learns from.

Feedback is appended, never overwritten, so earlier entries and human rejection reasons survive. A negative verdict leaves the ticket open on its own; pass reject: true to close it with your feedback as the recorded reason. Two constraints on rejecting, worth checking before the call:

• Agent-executed tickets can't be rejected over MCP — reject those in the web UI.
• Tickets already done can't be rejected — leave feedback without reject; it still reaches the generator.

Reading the thread

get_comments returns agent and human entries merged, newest first. Each entry carries source (agent for anything written over MCP, user for a comment typed in the Cogny UI, external_reviewer for a comment left through a review link) and type (comment, or feedback for an add_ticket_feedback entry, which also carries verdict).

Review links: comments from people outside Cogny

On workspaces with live edit, a ticket's preview can be shared with someone who has no Cogny account (a doctor checking medical claims, a lawyer, your client). They open the link, see the preview, and leave comments under their own name; the comments land on the ticket and the member who made the link gets an email for each one. Anyone holding the link can comment until it expires (7 to 90 days, 30 by default) or is revoked. While a reviewer has the preview open it runs and is billed like your own use of live edit.

Organizational Context

A hierarchical knowledge graph for company strategy, competitor intel, and team knowledge.

Example prompt: "What do we know about our competitors' ad strategies?"

Connected Platforms

Cogny federates tools from your connected ad platforms. These appear as namespaced tools (e.g., mcp__google_ads__execute_gaql) and vary based on your connected integrations.

Supported platforms include Google Ads, Meta Ads, LinkedIn Ads, Search Console, and more. Every integration's full tool list is at cogny.com/integrations/tools.

Connected a new integration mid-session? Your AI client keeps its own copy of Cogny's tool list, so a platform you connect later doesn't show up as mcp__… tools until the client reloads the list. In ChatGPT, open Settings → Apps → Cogny and choose Refresh; in other clients, reconnect or start a new session. Until then your agent can reach the new tools right away with list_integration_tools and call_integration_tool.

Cogny Analysis

Hand a question to Cogny's own analyst instead of working it through tool by tool. It works through the workspace's connected platforms, analytics, warehouse and business context, and writes up the findings with the numbers behind them and what to change. A run takes one to several minutes, and is also saved as a chat in the Cogny app.

An analysis is charged in credits by usage, like a question in Cogny's chat, on every plan: from the monthly credits on Cogny AI, Cogny Pro and Cloud, from top-up credits on Solo, and from the 50-credit welcome gift on Free. With no credits left, the tool says so and links to top-ups instead of starting.

MCP Tasks. Cogny supports the MCP Tasks extension (io.modelcontextprotocol/tasks, protocol 2026-07-28). A client that declares it gets a task handle from run_cogny_analysis and run_automation instead of a result: it polls tasks/get for the status and latest step and receives the answer there, can stop it with tasks/cancel, and doesn't hold a connection open while it runs. If the client can show forms (elicitation), the task first asks the user to confirm the credit spend for an analysis, or a live loop run with write access, and nothing starts until they say yes (tasks/update). A cancelled analysis stops at its next step and is charged for what it used until then. A client without Tasks gets a task_id, polls get_cogny_analysis and can stop the run with cancel_cogny_analysis; run_automation then answers as before. Without Tasks there is no confirmation form, so the agent asks the user before it starts an analysis.

Cogny Video

Hand Cogny the lines a short video should say and get back a finished MP4 in the Creative library, in one of two styles:

• Motion design (style: "motion_design", the default): animated text with the workspace's logo and brand colours, the exact words given and nothing else. Cogny writes the video, checks its frames for layout problems and with a visual review, fixes what they flag (up to three previews) and renders it, in one to three minutes. Every line stays on screen long enough to read (about a second plus 0.3 s a word), and emphasis picks the hero line: the punchline or the price, set as the biggest type and held longest.

• Influencer (style: "influencer"): a UGC-style clip of an AI-generated creator saying your lines to a phone camera, made by Cogny's AI video generator in one to five minutes. 3 to 15 seconds, vertical (the default) or landscape; a creator says about 2.5 words a second, so a 10-second clip takes about 25 spoken words. Numbers, prices and web addresses are said in full and count as the words they are said in: "$280 a month" is said "two hundred and eighty dollars a month", "cogny.com" "cogny dot com". The creator is never a real, named or lookalike person. The clip has no text in the frame, so use the lines as written as the post's caption.

  The speech is checked. When the clip is delivered, Cogny transcribes its audio and compares it with the script: every number must be right, and the other words may differ by at most one in five (a brand name spelled differently by the transcription is fine). If it doesn't match, Cogny makes the clip once more, asking for a slower and clearer delivery. Never more than once. If the second take doesn't match either, you still get the better of the two, with a warning that quotes what the creator actually says. The result says whether the speech was checked and what it found (speech_check in the structured result: matched, transcript, wer, takes).

Brand fonts. Motion design videos and carousels are set in the workspace's brand fonts (Settings → Brand & Context, heading and body) when they are among the free fonts Cogny bundles: Inter, Space Grotesk, Open Sans, Manrope, Poppins, Outfit, Quicksand, Montserrat, Roboto, Lato, DM Sans, Plus Jakarta Sans, Work Sans, Nunito, Nunito Sans, Raleway, Figtree, Rubik, Source Sans 3, IBM Plex Sans, Playfair Display, Lora, Merriweather, Source Serif 4 and JetBrains Mono. A font that isn't (a commercial or custom typeface) is set in the closest bundled one where there is a close match (SF Pro and Helvetica in Inter, Georgia in Merriweather, Gotham and Proxima Nova in Montserrat), otherwise in Inter, and the result says so in one line (fonts in the structured result: the families used and the one they replaced).

A motion design video costs about 2 to 4 credits: each frame preview and the render is about 1 credit. It needs 4 credits available to start. An influencer video costs about 0.9 credits a second (8 seconds is about 8 credits, 15 seconds about 14), charged when the clip is delivered; it needs one clip's credits available to start. If the speech has to be redone, the second clip is charged too, so the most an influencer video costs is twice that (22 credits for 8 seconds, 26 for 10). The redo only runs when the workspace still has the credits for it; otherwise the first clip is delivered with the warning. Both clips stay in the Creative library. The speech check itself costs a small fraction of a credit. With too few credits to start, the tool says so and links to top-ups. Like an analysis, a video runs as an MCP task for a client that declares Tasks (with a confirmation form first when the client can show one) and with get_video / cancel_video for any other client. Members with a view-only role can't start one.

The tools were called create_cogny_video, get_cogny_video and cancel_cogny_video when they launched; those names still work but are no longer listed.

Cogny Carousel

Hand Cogny the slides of a carousel and get back a finished set in the Creative library: one PDF, which LinkedIn shows as a swipeable document post, and one PNG per slide for a multi-image post on LinkedIn, X or Instagram. Each slide shows only the words you give it, word for word, with the workspace's logo, brand colours and brand fonts (as for videos, above) and, by default, its number ("2 / 7"). Cogny designs the slides, then checks every one before rendering: text inside the frame and not overlapping, type large enough to read on a phone, the hero text the largest on its slide, one design across all slides (same type, colours, margins, and the logo and number in the same place) and the right words on the right slide. It fixes what the checks and a visual review flag, up to three previews, and renders the set in one to three minutes. If the slides still fail after three tries, nothing is rendered and the result says why.

A carousel costs about 2 to 5 credits: each slide preview (all slides at once) and the final render is 1 credit, and writing the slides is a fraction of a credit each time. It needs 6 credits available to start. Like a video, it runs as an MCP task for a client that declares Tasks (with a confirmation form first when the client can show one) and with get_carousel / cancel_carousel for any other client. Members with a view-only role can't start one. Cogny never posts it by itself.

Posting it. With LinkedIn Organic connected, post the PDF as a document carousel with create_post: text (the post), document_url (the PDF URL from the result) and document_title (required by LinkedIn and shown on the document); schedule_post takes the same fields plus scheduled_for. For a multi-image post pass the PNG URLs as image_urls instead. On X, call upload_media with file_url for each PNG, then create_post with the media_ids; X takes at most four images a post. Post only after the user has seen the slides and said yes.

Example prompt: "Ask Cogny why paid search conversions dropped in the last 14 days and what we should change. Keep it under 50 credits."

Automations

Scheduled agents that act through write tools you greenlight. Every run is a ticket with a full transcript. Automations are included with Cloud. If these tools are not listed, automations aren't switched on for this workspace; contact support.

A scheduled run uses credits like any automation run. More in Scheduled tickets.

To point an automation at your own app, see Connect your own app to Cogny automations.

Workspace Status

Bugs, Feature Requests and Support

Your agent can reach the Cogny team directly, with the details it already has, instead of you retyping them somewhere else. Each filing returns a reference. The answer comes back to your agent through list_support_messages and to you in the Cogny app (the life-buoy icon in the header), with an email if nobody has read it within a day. A request for a platform Cogny doesn't connect to yet is researched on the spot instead, and the verdict usually comes back in the same call.

A workspace can file up to 10 of these per 24 hours. Replies don't count toward that.

Your agent and the app's support inbox show the same thread. Threads are per person: teammates don't see each other's.

Authentication

Cogny uses OAuth 2.1 with authorization code flow and PKCE. When you add Cogny as a connector, your MCP client handles the OAuth flow automatically — you just sign in with your Cogny account.

Scopes: Access is scoped to your warehouse. You can only query data, reports, and tickets belonging to workspaces you're a member of.

Token refresh: Access tokens are short-lived and automatically refreshed by your MCP client.

Subscription Tiers

The tools available depend on your Cogny plan:

Every plan reaches the tools of the platforms you connect, with the same read and write tools on a $9 Solo plan as on Cloud. Cloud adds the BigQuery warehouse, reports and growth tickets. On Free, tool calls count against the 15 monthly MCP calls; on Solo, Cogny AI, Cogny Pro and Cloud they are unmetered under fair use (a burst guard of 100 calls per minute per workspace, answered with HTTP 429 and Retry-After).

When your own AI client calls a write tool over MCP, that client is the approval step: it asks you before running the tool, and Cogny executes the call it receives. To take a tool away from your agent altogether, set it to blocked under Settings → Integrations.

Evaluating Cogny for someone? The buying guide for agents covers plans, approvals, limits, billing and data protection, and Integrations and tools lists every tool of every integration with what it does and whether it reads or writes.

Troubleshooting

"Unknown tool" error

Your subscription tier may not include that tool. Use the cogny_status tool to check your workspace status and available features.

Authentication fails

1. Ensure you're signing in with the email associated with your Cogny workspace
2. Check that your workspace has an active subscription
3. Try removing and re-adding the connector

No data returned from BigQuery

1. Verify your BigQuery warehouse is connected in the Cogny dashboard
2. Use list_datasets first to confirm available datasets
3. Check that your SQL query uses the correct dataset and table names

Connected platform tools missing

Connected platform tools only appear after you've linked the platform in your Cogny dashboard. Go to Settings > Integrations to connect Google Ads, Meta Ads, etc.

An analysis reads "working" for a long time

Most analyses finish in one to a few minutes. One that stops making progress (its job ended) is reported as failed on a later poll, within ten minutes, with nothing charged. Start it again.

Timeouts on long queries

BigQuery queries are capped at 500 rows. If your query is timing out, add LIMIT clauses or narrow your date range.

Privacy & Security

• All data stays within your Cogny workspace — Cogny does not share data between workspaces
• BigQuery queries are read-only (SELECT/WITH only) — your data cannot be modified
• OAuth tokens are scoped per-user and per-workspace
• See our Privacy Policy for full details

Support

• Email: support@cogny.com
• Documentation: cogny.com/docs