MCP server

Drive AI-search visibility from the assistant you already work in

The Attensira MCP server exposes 18 tools at https://mcp.attensira.com/mcp over Streamable HTTP. Your assistant reads mention rate, citation rate, sources, competitors and the inbox directly, and can add prompts, create automations and ask the agent — in the same conversation, without a workspace id anywhere in the request.

Ten tools are read-only and free to call. Three spend credits. Two delete data. That split is published, not implied — see the tool reference for every parameter, or the REST API if your own code is doing the work rather than an assistant.

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

What the MCP server does

It puts your visibility data inside the assistant you already work in. Instead of switching to a dashboard, reading a number and carrying it back by hand, the assistant queries the workspace directly and then changes it — adding prompts, creating an automation, asking the agent — without leaving the work in front of you.

Concretely: while editing a pricing page you can ask which prompts mention a competitor and not you, then add three prompts covering the gap, in one conversation. While reviewing a release you can pull mention rate for the last 30 days, compare it against the previous window, and check whether the delta clears a 95% two-proportion z-test before anyone writes it into a report. The tools return {value, n} for a rate and {value, real} for a delta precisely so that check is possible rather than assumed.

The 18 tools, by group

Six groups. Ten of the eighteen are read-only, which is the set an assistant spends most of its time in.

Attensira MCP tools grouped by what they do
Group	Tools	Notes
Find and measure	`search, get_analytics, get_prompt, get_page`	All read-only. get_analytics takes an area — visibility, prompt, page, source, competitor, topic, platform or bot.
Track	`add_prompts, remove_prompt, add_competitor`	add_prompts spends credits; remove_prompt is destructive; add_competitor changes no readings and costs nothing.
Inbox	`list_inbox, resolve_inbox_task`	list_inbox is read-only. Rows are needs_you or in_flight — worth calling before concluding the agent is idle.
Automations	`list_automations, get_automation, create_automation, run_automation, delete_automation`	Creating is free. run_automation spends credits and takes an idempotency_key. delete_automation is destructive.
Agent	`ask, get_session`	ask spends credits and may return status: running — poll get_session rather than treating that as an error.
Account	`get_account, get_ai_traffic_install`	Both read-only. get_account returns the plan, the credit balance, the tracked platforms and the scopes in force.

Full parameters, defaults and return shapes are in the tool reference.

Three tools cost credits. The ten read-only tools are free

Credits meter agent work, not reading. Pulling numbers Attensira has already collected costs nothing, no matter how often you pull them. That is the whole billing model in one sentence: you pay for work done, not for seats or dashboards.

Which MCP tools spend credits or destroy data
Tool	Effect	Why
`add_prompts`	Spends credits	Each new prompt is read across your tracked models from the next run onward.
`run_automation`	Spends credits	Triggers an automation immediately instead of waiting for its schedule.
`ask`	Spends credits	Runs an agent chat on an agent model.
`remove_prompt`	Destructive	Stops tracking a prompt; its history is no longer collected.
`delete_automation`	Destructive	Removes the automation, its schedule and its run history.

Pass an idempotency_key on every run_automation call. It maps to the Idempotency-Key header upstream, so a retried or duplicated call does not produce a second billed run.

How to connect a client

1
Add the server URLPoint your client at https://mcp.attensira.com/mcp over Streamable HTTP. In Claude Code that is: claude mcp add --transport http attensira https://mcp.attensira.com/mcp
2
AuthenticateRun /mcp in Claude Code, or click through the authentication prompt in Cursor or VS Code. A browser opens, you sign in and pick the workspace this connection should reach. For a headless caller, send Authorization: Bearer atn_live_<your key> instead.
3
Choose read, or read and writeThe consent screen names the application, the workspace and the access it wants. Read covers metrics, prompts, pages, sessions, automations and account details. Read and write adds everything that changes the workspace or spends credits.
4
Confirm with get_accountAsk the assistant to call get_account. It returns the workspace, the plan, the credit balance and the scopes in force without you supplying a workspace id — which confirms the credential is bound correctly before you rely on it.

Claude Code, in one command

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

Then run /mcp, select attensira, and authenticate. Add --scope user to make the server available in every project rather than only the current one. Cursor and VS Code take the same URL in .cursor/mcp.json or .vscode/mcp.json; with OAuth those files carry no secret, so the in-repo path is safe.

Claude Desktop, Claude.ai, ChatGPT, Windsurf, Cline, Codex CLI and n8n connect directly. Zed and other stdio-only clients need the mcp-remote bridge, which runs locally and forwards to the HTTPS endpoint. Copy-pasteable config for each is in Connect a client.

OAuth or a static key — both, permanently

OAuth compared with a static API key
	OAuth	Static API key
Best for	Interactive clients with a human and a browser — Claude, Claude Code, Cursor, VS Code, ChatGPT	Headless callers — n8n, CI jobs, scripts, servers
Secret on disk	None. The client stores a token it can refresh	Yes. A long-lived `atn_live_` key in a config file or secret store
Expiry	Access tokens last 60 minutes and refresh silently; you re-approve every couple of months	Never, until you revoke it
Where you revoke	Settings → Connections	Settings → Developer

You can run both at once. Claude Code on a laptop over OAuth and an n8n workflow on a static key are two independent credentials against the same server, and revoking either leaves the other alone. A key is displayed exactly once because only a hash is stored — a lost key is replaced, not recovered.

What keeps an assistant inside the lines

The honest objection to handing an assistant write access to your workspace is: what stops it doing something stupid? Three answers, all structural rather than settings.

No tool takes a workspace id. The credential names the workspace, so a request cannot be retargeted. A URL that cannot express a workspace cannot be pointed at someone else's.
Read and write are separate scopes. A write tool called without attensira:write returns 403 with insufficient_scope — the credential is valid, it simply may not do this. Scope cannot be widened after the fact; you revoke and redo it.
Anything irreversible stops for a person. Publishing, sending mail or spending with a third party raises a row in the inbox rather than proceeding, and an approval left unanswered for seven days closes the run having changed nothing.

Anyone can register an MCP client and pick its display name. The consent screen marks an application Attensira does not recognise as unverified. If you did not just start a connection yourself, close the tab.

What it does not do

Stating the limits is the point, not a disclaimer. These are the ones that change how you should read a result.

It does not stream whole answers through search. Every string on a hit is capped at 400 characters, so a complete model answer can never come back that way. Fetch the prompt, page or session for the full record.
Null is not zero. A null rate means Attensira did not measure that thing in that window. A per-platform row carrying tracked: false was never queried — rendering it as 0% invents a failure that did not happen.
It does not change what gets measured. Sampling stays at most once per prompt per model per country per workspace-local calendar day, and adding a prompt produces its first data on the next such day. There is no backfill.
There is no blended visibility score. Mention rate and citation rate are reported separately and deliberately never collapsed. There is no sentiment measurement and no brand rank, so competitor rates share your denominator and do not sum to 100%.
It speaks Streamable HTTP only. No stdio and no plain SSE transport. A stdio-only client needs the mcp-remote bridge.

Which plans include it

Every plan. The billing ladder gates two capabilities — Knowledge & Skills and Crawler Logs — plus one model surface, Gemini, and the MCP server is none of them. The plan changes what there is to read rather than whether you can read it: prompt slots, sampling depth and how long history is kept. Three tools meter credits on every plan, and the monthly credit allowance rises with the tier.

The exact slots, sampling depth, retention and credit allowance per tier are on the pricing page, which is generated from the same ladder the billing system reads.

Frequently asked questions

How many tools does the Attensira MCP server expose?

Eighteen, across six groups: find and measure, track, inbox, automations, agent, and account. Ten are read-only. Two delete data permanently — remove_prompt and delete_automation. Three spend credits — add_prompts, run_automation and ask. Every tool has an exact twin in the REST API, so the answer never depends on which door you came through.

Which MCP tools cost credits?

Exactly three: add_prompts, run_automation and ask. The other fifteen do not, and the ten read-only tools are free to call as often as you like. Credits meter agent work, not reading — pulling numbers you have already collected costs nothing. Creating an automation is also free; each run it performs is what spends.

Do I need an API key, or can I sign in with OAuth?

Both work on the same endpoint, permanently. OAuth is the better path wherever a browser can open — Claude, Claude Code, Cursor, VS Code, ChatGPT — because there is no key to paste into a config file, access tokens last 60 minutes and refresh silently, and revoking one client leaves the others working. Use a static atn_live_ key for headless callers such as n8n, CI jobs and scheduled scripts.

Can an assistant read another workspace's data by mistake?

No tool takes a workspace id. Your credential names the workspace, so there is no path segment, query parameter or argument anywhere in the tool set that could be pointed at someone else's data. To work across two workspaces you connect twice — one OAuth connection or one key per workspace, as two server entries.

What stops an assistant deleting my prompts?

Scope. Read tools need attensira:read, which every credential has; the eight write tools additionally need attensira:write, which you either approve at the OAuth consent screen or choose when minting a key. A write tool called without it returns 403 with error insufficient_scope. Scope is decided when you connect and cannot be widened afterwards — you revoke and redo it.

Why does my client show no Attensira tools?

The server speaks Streamable HTTP only — no stdio and no plain SSE transport. A client that launches local processes needs the mcp-remote bridge. Some clients also only load servers added before the conversation started, and some cache the tool list until a restart. A healthy connection lists 18 tools; https://mcp.attensira.com/healthz answers without a key if you need to confirm reachability first.

Which plans include the MCP server?

All of them. MCP is not one of the capabilities the billing ladder gates — those are Knowledge & Skills and Crawler Logs, both Growth and above — and it is not affected by the one gated model surface, Gemini, which also unlocks on Growth. What varies by plan is what there is to read: prompt slots, sampling depth and retention.

Where to go next

The REST APIThe same 18 capabilities as plain HTTP, for a cron job, a CI step or a dashboard.IntegrationsEvery way Attensira connects to the rest of your stack, and what each one is allowed to do.AI crawler logsThe bot data behind get_analytics area=bot, and why it has to be measured at the edge.PricingPrompt slots, sampling depth, retention and credits per plan, generated from the billing ladder.