Integrations

Four ways Attensira connects — and what each one may do

Attensira plugs into a stack in four documented directions: a first-party GitHub App that opens pull requests and never merges them, an MCP server your assistant calls, plugins and your own MCP server that the agent calls, and edge middleware that counts AI crawler fetches. Each one has a published boundary rather than an implied one.

This page lists only connections that exist. Where there is no integration — a CMS publishing connector, for instance — it says so instead of promising one. Full behaviour for each is on docs.attensira.com.

Karl-Gustav Kallasmaa, Founder & CEOLast updated 3 September 2026

The four connections, and their direction

Direction is the thing that gets confused, so it is worth stating first. Two of these let Attensira reach outward; one lets your tools reach in; one lets the agent reach tools you own.

Attensira integrations by direction and boundary
Connection	Direction	The boundary
GitHub App	Attensira → your repositories	Opens pull requests. Never merges, never pushes to your default branch, never pushes twice, never touches CI config.
MCP server	Your assistant → Attensira	18 tools. Ten read-only, three spend credits, two destructive. No tool takes a workspace id.
REST API	Your code → Attensira	The same 18 capabilities as plain HTTP. Key only; the key names the workspace.
Plugins and your own MCP server	The agent → your tools	A granted plugin is the only way a tool enters the agent's toolset. Grants are per workspace and revocable.

A fifth path is not a connector at all but is worth listing beside them: edge middleware, which counts AI crawler fetches at your own edge and sends them to Attensira.

GitHub: a first-party App, not an OAuth connection

Attensira finds things worth fixing on your site. GitHub is how a fix leaves the product and arrives somewhere you can review it — the agent opens a pull request against the repository the page actually lives in. It is a first-party GitHub App you install into your own organisation, not an OAuth app and not brokered through a third-party connector. GitHub is explicitly not listed among connected applications for that reason.

1
Install the AppFrom the project's GitHub settings in Attensira, press Connect GitHub. GitHub takes over and asks which account or organisation to install into.
2
Choose the repositoriesGrant the App only selected repositories, and pick the ones whose content you want changed. GitHub, not Attensira, holds this list — changing it later is done on GitHub's installation page.
3
Attach them to the projectBack in Attensira, choose which of the installed repositories this project may open pull requests against. A project can have several, so each pull request lands in the repo the page actually lives in.

Exactly three permissions

Permissions the Attensira GitHub App requests
Permission	Access	What it is for
Contents	Read and write	Read the files behind a page, and push our branch
Pull requests	Read and write	Open the pull request, and read its state afterwards
Metadata	Read	Mandatory for every GitHub App; lists the repositories you granted

No Actions permission, no Administration permission, no access to secrets. If GitHub's consent screen shows you something not on that list, you are not installing the Attensira App.

A red build or a merge conflict is surfaced to you and left there. The no-second-push rule has no exception for fixing our own work, so nobody should sit waiting for a follow-up commit that is not coming.

A project can attach several repositories, which is the normal case rather than an edge case: a marketing site, a docs site and the app itself are often three repos, and a finding on /pricing belongs in a different one from a finding on /docs/quickstart. Every pull request we open carries the attensira label, which is a stable hook for a rule on your side, and draft mode makes every one a GitHub draft so most merge automation leaves it alone. Disconnecting ends access immediately and leaves pull requests we already opened exactly as they are. Full behaviour is in the GitHub documentation.

MCP clients: Attensira inside the assistant you use

The MCP server is one endpoint — https://mcp.attensira.com/mcp over Streamable HTTP — and every client below reaches it. What changes is where the URL goes and which credential you use.

MCP clients documented as connecting to Attensira
Client	Connects directly	Credential
Claude Code, Claude Desktop, Claude.ai	Yes	OAuth (native remote HTTP)
Cursor, VS Code + Copilot	Yes	OAuth — browser prompt on first connect
ChatGPT	Yes	OAuth — connector configured in the UI
Windsurf, Cline	Yes	OAuth varies by build; a static key always works
Codex CLI	Yes	Static key, kept out of config via --bearer-token-env-var
n8n	Yes	Static key — headless, nobody to approve a prompt
Zed and older stdio-only builds	Via the mcp-remote bridge	OAuth through the bridge

claude mcp add --transport http attensira https://mcp.attensira.com/mcp

Prefer OAuth wherever a browser can open: there is no key to paste into a committed config file, and revoking one client leaves the others working. Use a static atn_live_ key where nothing can complete a browser sign-in — a workflow firing at 3am is exactly the case OAuth cannot serve. See the MCP server page for the tool set and the credit split.

Plugins and your own MCP server: tools the agent may call

Plugins run the other way. A plugin is a third-party tool the agent may call on your behalf, and the rule that governs it is short: a granted plugin is the only way a tool enters the agent's toolset. Until you grant one, the agent cannot see its tools at all. Grants are per workspace and revocable at any time.

If you run your own MCP server, register it with a name and a URL and the agent can call it like a built-in. Attensira calls tools/list and registers what it finds, namespaced so collisions are impossible:

mcp__<slug>__<tool>

# a "search" tool on a server registered as "warehouse"
mcp__warehouse__search

Two auth modes. none for an open server, or apiKey for a key sent on each request.
OAuth is refused. Attensira has no token store, so registering an OAuth server returns an explicit error rather than accepting the configuration and failing later with a 401 nobody can diagnose. Use apiKey, or put an authenticating proxy in front.
An unreachable server does not vanish. It becomes a single tool named mcp__<slug>__unavailable that explains the problem when called — because a tool that disappears looks to the agent like a capability that never existed.
Discovery is bounded. A discovery timeout, a call timeout, a tools-per-server cap and a whole-session discovery budget mean one slow server is skipped for that session rather than delaying every answer.

Built-in tools are assembled first, so a registered server can never shadow one. Plugins also cannot be force-refreshed — there is no endpoint behind such a control, and the product omits the button rather than showing one that spins and does nothing.

Edge middleware: the only place a crawler fetch exists

AI crawler measurement is available on Growth and Business, and it has to happen in edge middleware. That is a hard technical constraint, not a design preference: AI crawlers do not execute JavaScript, so a client-side beacon has never seen one and never will — by the time the script runs, the only visitors left are humans with browsers.

Attensira generates the install instructions for your workspace in Settings, carrying the ingest endpoint and your workspace's own values. Use the generated copy rather than an example found elsewhere; a snippet without those values sends data nowhere. Over MCP or REST, get_ai_traffic_install and GET /v1/ai-traffic/install return the same thing, including a copy-pasteable edge worker.

There is no verified-bot claim here. User agents are self-reported and trivially spoofed, so treat the data as directional evidence about coverage rather than an audited log of who visited — and never reconcile crawler counts with citation counts. They are separate systems.

What is not an integration today

Listing an integration that does not exist is the textbook doorway-page failure, so these are named rather than implied.

No CMS publishing connector. Publishing goes through GitHub pull requests. If your site is not in a repository, drive the writing from your own side with the REST API or MCP.
GitHub is not an OAuth connection. It does not appear under connected applications, because it is a GitHub App installed on GitHub's side.
A registered MCP server cannot use OAuth. This is refused explicitly at registration rather than half-working.
Connecting anything does not change measurement. Mention rate, citation rate and source share are untouched by every integration on this page. Sampling stays at most once per prompt per model per country per workspace-local calendar day.

Revoking, and what survives it

GitHub. Disconnect in the project's settings or uninstall the App on GitHub — either ends access immediately. Existing pull requests stay exactly as they are; what stops is everything future, and tracking on existing inbox rows freezes where it was.
OAuth connections. Revoke at Settings → Connections. Refreshes stop immediately, but an already-issued access token stays valid for up to 60 minutes. A connection idle 60 days, or 180 days old, forces re-consent.
API keys. Revoke at Settings → Developer. Keys are never cached in the API layer, so revocation takes effect on the next request. One key reaches both MCP and REST, and one key belongs to one workspace, chosen at mint time and never movable.
Plugins. Revoke the grant and the agent stops seeing that plugin's tools at all.

Frequently asked questions

How does Attensira connect to GitHub?

Through a first-party GitHub App you install into your own organisation or account. It is not an OAuth app and it is not brokered through a third-party connector: the App is ours, the installation is yours, and you can see and revoke it from GitHub's own settings at any time. You grant it only selected repositories, and GitHub — not Attensira — holds that list.

Can the agent merge a pull request?

No. Four rules bound the whole of its write access: it never merges a pull request, never pushes to your default branch, never pushes twice to the same branch, and never touches CI configuration or workflow files. One finding, one branch, one push. A red build or a merge conflict is surfaced in the inbox and left for a person — there is no follow-up commit coming.

What permissions does the GitHub App ask for?

Exactly three: Contents read and write, Pull requests read and write, and Metadata read. There is no Actions permission, no Administration permission, and no access to secrets. If GitHub's consent screen shows you something not on that list, you are not installing the Attensira App.

Can I connect my own internal tools to the agent?

Yes. Register your own MCP server with a name and a URL, choose auth mode none or apiKey, and Attensira calls tools/list and registers what it finds. Discovered tools are namespaced mcp__<slug>__<tool>, so two servers can both expose a search tool and Attensira's built-ins can never be shadowed. OAuth is refused as an auth mode — there is no token store, so registering an OAuth server returns an explicit error rather than failing later with an undiagnosable 401.

What stops the agent using a tool I did not give it?

A granted plugin is the only way a tool enters the agent's toolset. Until you grant one, the agent cannot see its tools at all. Grants are per workspace and revocable, and a granted plugin is available to the whole workspace rather than to one person or one chat.

Does Attensira integrate with my CMS?

Publishing goes through GitHub pull requests. There is no documented CMS publishing connector today, and we would rather say so than list one that does not exist. If your site lives in a repository, the GitHub App covers it; if it does not, the REST API and MCP server let your own code or assistant do the writing on your side.

How do I get AI crawler data into Attensira?

With edge middleware, on Growth and Business. AI crawlers do not execute JavaScript, so a client-side beacon has never seen one and never will — a crawler fetch only exists to be counted on the server, before any page renders. Attensira generates install instructions carrying your workspace's own values in Settings; a snippet copied from elsewhere sends data nowhere.

Where to go next

The MCP server18 tools, the credit split, and how to connect Claude Code, Cursor or n8n.The REST APIThe same capabilities as plain HTTP, for a cron job or a dashboard.AI crawler logsWhat the edge middleware collects once it is deployed, and how to read it.PricingWhich capabilities the billing ladder gates, and what each tier includes.