signalfort · from outside
Triage your vulnerabilities with an AI assistant over MCP
Connect Claude, Cursor, or VS Code to signalfort. The assistant sees your findings, helps decide what to fix, and queues a recheck after the fix — through a key you scope yourself.
- ✓Any MCP client: Claude, Cursor, VS Code
- ✓Key with scopes and a domain allowlist
- ✓No access to your server
- ✓Every call is logged
signalfort in numbers
Nothing to install. Enter a domain — and the check starts.
Giving AI access to security feels risky
Usually “connect an assistant” means handing over a broad token or letting it onto the server. For your own security data, that is too much.
- —A key with full access to the whole org
- —Server access, SSH, or an agent on the host
- —No clarity on what the assistant read and did
- —Secrets and data leaking into the model chat
What the assistant does
It works with signalfort data — your sites, findings, and CVEs — not your server.
Sees your sites
Hosts, subdomains, stack, SSL and domain expiry — the same as in the cabinet.
Reads the queue
Open findings and CVE/BDU matches, sorted by KEV, EPSS, and risk.
Helps decide
Affects us / not, accept the risk, mark fixed — with a reason in the note.
Queues a recheck
After a fix: retest a finding or rescan a domain. FSTEC/CRA report in one call.
Connect in three steps
The key lives in the cabinet. You set scopes and the domain list; read-only by default.
- 01
Create a key
Settings → Integrations. Pick scopes (reader = read-only) and, if needed, a domain list.
- 02
Add it to the client
The server is mcp.signalfort.ru/mcp. Add it to Claude, Cursor, or VS Code as a remote MCP server with an Authorization: Bearer <key> header. A ready-made config is in the cabinet.
- 03
Ask about your sites
“Which findings on acme.com are known-exploited?” — the assistant reads live data and answers.
Revoke the key at any time — one button
Direct access vs MCP
Both let an assistant work with your security. The difference is what it can reach.
- Broad rights across the whole org
- SSH or an agent on the host
- Secrets in the model chat
- No record of what the assistant did
- 1.Scope per taskread-only by default
- 2.Domain allowlistthe key sees only those
- 3.No server accesssignalfort data only
- 4.Every call loggedwho, what, and when
A decision from chat — in the cabinet
What the assistant decides over MCP shows up in the cabinet and in the time-to-decision metric.
A BDU match landed in the findings queue
The assistant checked the stack and marked it “not affected”
- ·status: not affected
- ·reason: product is nginx
- ·in the time-to-decision metric
Who it suits
People who already live in an editor or assistant chat and do not want another tab.
Security / DevSecOps
Triage the findings queue right in your assistant, without clicking through the cabinet.
IT and development
Ask “what to fix on this host” and queue a recheck after the fix — from your editor.
Leadership
A short “what here actually affects us” in the assistant’s own words.
MCP is on Business
Read API, webhooks, and MCP for AI assistants are included in Contour Business.
Radar
₽0
Checks from the outside and monitoring. No API or MCP.
Contour
₽1,490/mo
Active checks, confirmation, SLAs. No MCP.
Business
₽4,490/mo
5 domains, read API, webhooks, MCP for AI assistants, FSTEC/CRA reports.
What MCP is and why it is here
MCP (Model Context Protocol) is an open standard by which an AI assistant gets tools. signalfort runs an MCP server, and any compatible client — Claude, Cursor, VS Code — calls its tools on behalf of your key.
The point is simple: your findings show up where you already work. No switching to the cabinet just to ask “what’s exposed”.
What the assistant can and cannot do
Can: read your sites, findings, and matches, decide on them, queue a recheck, and export a report — the same actions as in the cabinet.
Cannot: reach your server, read the contents of your data, or get test accounts for deep checks. Secrets are never returned to the assistant.
Why it is safe
A key carries scopes and a domain list. Read-only by default; writing (decisions, scans, reports) turns on through separate scopes you grant deliberately.
Different customers’ data is separated: a key sees only your org and only the allowed domains. Every tool call is logged — which key, what, and when.
A finding carries a minimal fact (port, version, response type), not the contents of your forms or a data dump. More on confirmation: does a CVE affect you.
Where to start
MCP is part of Contour Business. You create the key in the cabinet: Settings → Integrations.
Connect your site first — a free Radar check — then turn on MCP on Business.
FAQ
Does the assistant get access to our server?
No. MCP serves signalfort data — your sites, findings, and matches. The assistant does not connect to your server, SSH, or files. Scanning still happens only from the outside and only on verified domains.
What if the key leaks?
A key is limited by scopes and a domain list, and read-only by default. You can revoke it in the cabinet with one button; every call is in the log.
Can the AI break or delete something?
The action set is fixed: decisions on findings, recheck, scan, report. There is no tool to delete domains or data. Writing works only if you granted the key those scopes.
Does our security data go into the model?
What passes through the assistant is what it asked for: findings and matches as facts (port, version, status), without the contents of your data. Test accounts and secrets are never returned.
Which clients are supported?
Any MCP client: Claude, Cursor, VS Code, and others. The tools are self-describing — the client sees which arguments to pass.
Is it on the free plan?
MCP, read API, and webhooks are part of Contour Business. Radar and Contour do not include them.