Files
Jakob Wennberg f929b4b1d2 feat(mcp): lazy authentication so a client can connect before an account exists (#1814 PR 2) (#1892)
* feat(mcp): lazy authentication so a client can connect before an account exists

Second PR of agent-first onboarding (#1814). A client with no token may
now initialize, list the default catalog and call the three documentation
tools (search_tools, list_skills, load_skill). Every other request keeps
the transport-level 401 + WWW-Authenticate, which is what Claude, Claude
Code and Codex turn into their Connect prompt; with #1855 the account is
created inside that prompt, so the first protected tool call is the whole
signup trigger.

- The JSON-RPC body is parsed before auth so the method and tool name can
  decide whether a token is required. A tokenless unparseable body keeps
  the old 401 answer.
- Anonymous callers get an 'anonymous' actor, an empty scope set, a
  not-connected variant of the initialize instructions, and the full
  default catalog from tools/list (the agent has to be able to name a
  protected tool to trigger the challenge).
- Anonymous traffic is rate-limited per truncated IP via checkRateLimit;
  truncateIp moves to lib/api/ip.ts so the MCP server can use it without
  importing the v1 wrapper (which pulls lib/init and would cycle).
- gnubok_list_skills is now company-independent and skips its two context
  lookups when there is no company (anonymous or not yet onboarded).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018wCdzRTatKiDByKB8hCNT6

* fix(mcp): gnubok_list_skills keeps its company_id argument as an optional-company tool

Making list_skills company-independent (so anonymous callers can run it)
silently dropped its company_id argument: a multi-company user asking
for another company's skill list got the key default instead. Optional-
company tools now advertise company_id and resolve (membership-checked)
it when an authenticated caller names one; anonymous callers cannot.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018wCdzRTatKiDByKB8hCNT6

---------

Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-25 11:38:29 +02:00

40 lines
1.7 KiB
TypeScript

/**
* Truncate a client IP to a privacy-preserving prefix for rate limiting and
* forensic logging: IPv4 to its /24, IPv6 to its /48. Enough for abuse
* correlation, not enough to identify a point of presence.
*
* Honors `x-forwarded-for` when set (Vercel / proxies); behind Vercel the
* leftmost value is rewritten by the edge so we accept it as authoritative.
*
* Lives in its own module (no route or init imports) so both the v1 REST
* wrapper and the MCP server can use it without an import cycle.
*/
export function truncateIp(ip: string | undefined): string | undefined {
if (!ip) return undefined
// IPv4: validate octets are 0-255, then drop last octet → "203.0.113.0/24".
// Out-of-range octets indicate a spoofed or malformed header; refuse to
// log a pseudo-IP that would pollute abuse-pattern analysis.
const v4 = /^(\d{1,3})\.(\d{1,3})\.(\d{1,3})\.(\d{1,3})$/.exec(ip)
if (v4) {
const octets = [v4[1], v4[2], v4[3], v4[4]].map((s) => Number.parseInt(s, 10))
if (octets.every((o) => o >= 0 && o <= 255)) {
return `${octets[0]}.${octets[1]}.${octets[2]}.0/24`
}
return undefined
}
// IPv6: keep first 3 hextets → "2001:db8:abc::/48"
const v6 = /^([0-9a-f]{1,4}:[0-9a-f]{1,4}:[0-9a-f]{1,4}):/i.exec(ip)
if (v6) return `${v6[1]}::/48`
return undefined
}
/**
* The client IP as the request presents it: leftmost `x-forwarded-for`
* entry, else `x-real-ip`, else undefined.
*/
export function requestClientIp(request: Request): string | undefined {
const forwarded = request.headers.get('x-forwarded-for')
if (forwarded) return forwarded.split(',')[0]?.trim() || undefined
return request.headers.get('x-real-ip') ?? undefined
}