AI agents and the MCP server Beta

Without an account

These tools need no sign-in and read no account:

• check_service_status: Check whether a service is down
• check_service_status_url: Check whether a website is down
• recommend_monitoring: Recommend what to monitor for a website
• explain_features: Explain what FlareWarden does
• search_help: Search FlareWarden's help docs

check_service_status reports what FlareWarden's public status page for a service shows: the worse of the vendor's own status page and FlareWarden's checks. check_service_status_url checks a URL from several regions; each host is checked at most once every 5 minutes, and a repeat returns the earlier result with its age.

recommend_monitoring previews a site's monitoring: the services it relies on and their status now, when its SSL certificate expires, a summary of its DNS, whether it is up, and what FlareWarden would set up, in monitors. It stores nothing about who asked.

An agent asks first whether you already have an account. Someone with no account who wants a site monitored gets the link to sign up at https://app.flarewarden.com/auth/register?utm_source=ai_assistant&utm_medium=mcp straight away, and the preview only if they ask for it. The link's one tag is the same for everyone and tells us the sign-up came through an AI assistant. Signing up carries nothing from the conversation: no site, address or redirect. Back in the conversation, the agent's next account call starts sign-in, then it plans the site and sets it up once they confirm.

With your account

Connected to your account, an agent works in one FlareWarden account as you. It can report on your sites, incidents, status pages and usage; add, change and pause monitoring; run status pages; post incidents and schedule maintenance; and manage the team, alert emails and webhooks, as far as your role and the permissions you grant allow. There are two ways to connect:

• Sign in (Claude, Claude Code, ChatGPT, Cursor, Codex and other apps that support it). You add the server to the app, then sign in on FlareWarden and approve the connection: you choose the account, the permissions and a daily monitor budget.
• A personal access token (scripts, or any app you'd rather give a token). You create it on your account's AI Agents & API page and send it in the Authorization header.

Any member of an account may connect an agent, whatever their role. To let an agent read and never change anything, connect it to https://mcp.flarewarden.com/mcp/readonly instead.

Permissions

Each connection has permissions, chosen when you approve an app or create a token. A tool needs the permission it belongs to, and the agent sees only the tools its permissions and your role allow. When you approve an app, the permissions it asked for that your role can use are preselected; a token, or an app that asks for none, starts from the column below. You can untick any of them but Read.

The rules every agent works under

Never more than you can do on the website

Every tool, reading or writing, goes through the same permission checks as the page or button you would use: your role, the sites and status pages you have been given, and whether your access to the account is suspended. A change the website would refuse is refused, in the website's words. FlareWarden reads your membership and role again on every call, so a role change applies to the next call, and an agent of someone removed from the account stops working.

New monitors are planned, then created when you confirm

The plan tools (plan_site, plan_dependency, plan_content_check, plan_cron_monitor, plan_dns_watch and plan_status_page) change nothing. They return the choices to put to you, with FlareWarden's recommendation and what each costs in monitors. The agent is told to apply a plan only once you confirm, by passing its id to create_site or the matching add_ tool with the plan's confirm sentence: FlareWarden's own words for exactly what applying it does, which your client's approval prompt shows. A sentence that is not the plan's as it stands is refused before anything is created, so a plan changed after you saw it is shown to you again. Applying the same plan again returns the first result rather than adding the monitors twice, and a plan can be applied for 24 hours. A status page is public as soon as it exists, so create_status_page refuses a plan in which you have not confirmed publishing.

Incident posts and maintenance are drafted, then published

draft_incident_post and draft_maintenance post nothing. They return the exact text, the status pages it will appear on, and who would be notified, counted by the same code that sends the notices. The agent is told to call publish_incident_post or schedule_maintenance with the draft's id and its confirm sentence, which quotes the full text and who is notified, only after you say yes. Where your client can show a form, publishing an incident post also asks you there.

Secrets never pass through the chat

• An API token a dependency check needs is entered on a secure FlareWarden page the agent links you to, never typed into the chat.
• A webhook an agent adds shows its signing secret once, on a FlareWarden page only you can open.
• Slack, Discord, Google Chat and Microsoft Teams alerts are connected on FlareWarden's own page. An agent that passes one of their webhook addresses is refused and given that page's link.

Deletes and billing stay in the dashboard

No tool deletes a site, a check, a status page or a team member, and none changes your plan or billing. Do those in the dashboard.

A daily monitor budget per connection

Each connection may add at most 25 monitors in any 24 hours unless you change it, counted as your plan counts them. You can set it from 1 to 1000 on the AI Agents & API page, and your plan's limit still applies. Posting an incident or scheduling maintenance adds no monitors, so it is never held to the budget. Separately, agents can create at most 50 sites in an account in an hour.

The owner can turn AI agents off

The account owner's Allow AI agents switch, on the AI Agents & API page, applies to every member. Turning it off disconnects every agent and revokes every personal token at once, and while it is off nobody can connect an agent or create a token.

Every change is in the audit log

A change an agent makes is recorded in the account's audit log as yours, marked with the name of the connection it came through.

Rate limits

Calls are limited per connection, and the tools that need no sign-in have limits of their own. A connection past its limit is answered HTTP 429 with a Retry-After header.

Text from other sites

Some results carry text a third party wrote, such as a vendor's incident titles or a website's page title. It is only ever in fields named site_text, as quoted data, never in a result's summary or its suggested next steps.