Files
Maven/internal/webfetch/webfetch.go
kami 95ae900a58 Talk MCP: a client for external tool servers (#251)
docs/plans/06-mcp-support.md asks for the host direction — Maven connects OUT
to MCP servers and consumes what they offer. This is the client half: the
protocol, the transports, the connection manager, the config block. Nothing is
wired into a turn yet, and nothing here exposes Maven's own capabilities to an
outside caller.

internal/mcp:
  - hand-rolled JSON-RPC 2.0 (the wire format is four fields, and the repo
    vendors its deps, so a library would cost more than it saves);
  - two transports: a stdio subprocess on this box, and streamable HTTP, which
    accepts a plain JSON reply or an SSE frame because servers disagree about
    which they send;
  - Client: initialize handshake, tools/list, tools/call, resources/list,
    resources/read. Text content only — everything downstream is a sentence;
  - Manager: lazy dial, per-server failure that never blocks boot or the other
    servers, backoff reconnect, Status for a web surface, graceful Close;
  - the allowlist encoding: a discovered tool becomes the store row
    "vikunja_list_tasks" with cmd ["mcp","vikunja","list_tasks"], scope
    "mcp:vikunja". No new column, no migration, and ProposeTool, EnableTool,
    the act matcher and the confirm turn all keep working untouched.

Constraints held, in code rather than in prose:
  - OFF unless configured, and a server is dark until "enabled": true.
  - A url server goes through internal/webfetch, so the SSRF guard, the size
    cap, the redirect cap and the per-host rate limit apply. Reaching loopback
    needs allow_private on THAT server, and each server gets its own fetcher so
    one loopback exemption cannot become a hole for a public endpoint.
  - readOnlyHint decides destructive: no hint means "assume it mutates", which
    will route the call through the existing confirm turn. Guessing wrong in
    that direction only costs a question.
  - The catalogue stays small on purpose — allow_tools, and max_tools=12 per
    server. The resident model is a 1.7B with a 4096-token context; a tool name
    it half-remembers is a wrong act.
  - Only the tool name and the router's arguments are sent. There is no API
    here through which a note, a fact or the persona block could travel.

webfetch grows Post (JSON-RPC cannot be a GET) and surfaces response headers
for Mcp-Session-Id. It shares Get's guards exactly: a body buys a caller
nothing, a POST to the LAN is refused for the same reason a GET is.

Verified against the real Vikunja MCP server on homesrv
(http://localhost:9100/mcp): handshake, three discovered tools with update_task
correctly NOT read-only, a live list_projects call, a tool excluded by
allow_tools refused, and the same server refused outright once allow_private
was dropped. Tests cover both transports (the stdio one against a real
subprocess), SSE and JSON framing, session echo, reconnect, and the config
validation.
2026-08-01 04:22:52 +04:00

349 lines
12 KiB
Go

// Package webfetch is the one door Maven uses to read something off the
// network, and it is a narrow one.
//
// "Never phones home" stopped being a hard constraint on 2026-07-31, but what
// replaced it is not "she may fetch anything": local sources come first (Kiwix
// on the box), external fetching is off unless configured, and only the
// utterance ever leaves — never his notes, facts or history. That policy is
// enforced by the callers. What THIS package enforces is the part that must be
// code rather than a paragraph in a plan, because it protects the homelab from
// its own assistant:
//
// - http/https only — no file://, no ftp://, no gopher;
// - no private address, ever: loopback, RFC1918 (which is what makes the
// 10.42.0.0/24 wireguard tunnel and the 192.168.1.0/24 LAN unreachable),
// link-local incl. the 169.254.169.254 cloud metadata address, CGNAT,
// unique-local v6. Checked in the dialer's Control hook, so it holds for
// every address the resolver returns AND for every hop of a redirect
// chain — a DNS name that resolves to 127.0.0.1 is refused at connect
// time, which a pre-flight lookup could not promise (rebinding);
// - an allowlist, when one is configured, and a denylist that always wins;
// - a response size cap, a total timeout, a redirect cap;
// - one request per host per interval, so a poll loop with a bug is slow
// rather than an outbound flood.
//
// Everything above is on by default with sane numbers: a zero Config is a
// usable, conservative fetcher. There is no cache and no retry — a feed poll
// or a page read that fails is simply not answered this round.
package webfetch
import (
"bytes"
"context"
"errors"
"fmt"
"io"
"net"
"net/http"
"net/url"
"strings"
"sync"
"syscall"
"time"
)
// Defaults. Small on purpose: this reads feeds and article pages, not ISOs.
const (
DefaultTimeout = 20 * time.Second
DefaultMaxBytes = 2 << 20 // 2 MiB
DefaultMaxRedirects = 3
DefaultHostInterval = time.Second
DefaultUserAgent = "Maven/1.0 (self-hosted personal assistant)"
)
// Errors callers distinguish. Everything else is wrapped transport error.
var (
ErrScheme = errors.New("webfetch: only http and https are allowed")
ErrBlocked = errors.New("webfetch: host is not allowed")
ErrPrivate = errors.New("webfetch: refusing to connect to a private address")
ErrTooLarge = errors.New("webfetch: response exceeds the size cap")
ErrRedirects = errors.New("webfetch: too many redirects")
ErrStatus = errors.New("webfetch: non-2xx status")
)
// Config are the limits. Every zero value means "the default above", so
// Config{} is safe; the only field that changes behaviour by being empty is
// AllowHosts (empty ⇒ any public host that is not denied).
type Config struct {
// AllowHosts — when non-empty, the ONLY hosts that may be fetched. An
// entry matches the host itself and its subdomains ("example.com" allows
// "news.example.com"). This is the knob to reach for when a capability
// should read two feeds and nothing else.
AllowHosts []string
// DenyHosts — same matching, checked first and always winning.
DenyHosts []string
Timeout time.Duration // whole request, including redirects and body read
MaxBytes int64 // response body cap
MaxRedirects int // 0 ⇒ default; negative ⇒ no redirects followed
HostInterval time.Duration // minimum spacing between requests to one host
UserAgent string
// AllowPrivate disables the private-address guard. It exists for tests
// (httptest listens on 127.0.0.1) and for an explicitly configured
// on-box mirror. Nothing in deploy/mavend.json sets it, and it should
// stay that way: with it on, any URL Maven is handed becomes an SSRF
// probe of the LAN and the wireguard range.
AllowPrivate bool
}
// Response is a fetched body, already bounded by MaxBytes.
type Response struct {
URL string // final URL after redirects
Status int
ContentType string
Body []byte
// Header — the response headers, one value each (the first). Populated for
// every request; MCP needs Mcp-Session-Id, nothing else reads it.
Header map[string]string
}
// Fetcher performs guarded GETs. Safe for concurrent use; the per-host rate
// limiter is shared, which is the point of sharing one Fetcher.
type Fetcher struct {
cfg Config
http *http.Client
mu sync.Mutex
last map[string]time.Time // host → when we last dialed it
}
// New builds a fetcher from cfg, filling in defaults.
func New(cfg Config) *Fetcher {
if cfg.Timeout <= 0 {
cfg.Timeout = DefaultTimeout
}
if cfg.MaxBytes <= 0 {
cfg.MaxBytes = DefaultMaxBytes
}
if cfg.MaxRedirects == 0 {
cfg.MaxRedirects = DefaultMaxRedirects
}
if cfg.HostInterval <= 0 {
cfg.HostInterval = DefaultHostInterval
}
if cfg.UserAgent == "" {
cfg.UserAgent = DefaultUserAgent
}
f := &Fetcher{cfg: cfg, last: map[string]time.Time{}}
dialer := &net.Dialer{Timeout: 10 * time.Second}
if !cfg.AllowPrivate {
// The guard lives here rather than in a pre-flight net.LookupHost so
// that it sees the address actually being connected to: every A/AAAA
// the resolver handed back, on every redirect hop, with no window in
// which the name could be re-pointed at the LAN.
dialer.Control = func(_, address string, _ syscall.RawConn) error {
host, _, err := net.SplitHostPort(address)
if err != nil {
return err
}
ip := net.ParseIP(host)
if ip == nil || IsPrivateIP(ip) {
return fmt.Errorf("%w: %s", ErrPrivate, host)
}
return nil
}
}
f.http = &http.Client{
Timeout: cfg.Timeout,
Transport: &http.Transport{DialContext: dialer.DialContext},
CheckRedirect: func(req *http.Request, via []*http.Request) error {
if len(via) > f.cfg.MaxRedirects {
return ErrRedirects
}
// A redirect is a fresh URL and gets the full check: an allowed
// host must not be able to bounce us onto a denied one.
return f.checkURL(req.URL)
},
}
return f
}
// Get fetches rawURL. The body is capped: a larger response is an error, not a
// truncation, because half an XML document is worse than none.
func (f *Fetcher) Get(ctx context.Context, rawURL string) (*Response, error) {
return f.do(ctx, http.MethodGet, rawURL, nil, nil)
}
// Post sends body to rawURL and returns the reply, under exactly the same
// guards as Get: scheme rule, host lists, the dialer's private-address check on
// every hop, the size cap and the per-host rate limit.
//
// It exists for JSON-RPC over HTTP (internal/mcp), which cannot be expressed as
// a GET. That an outbound request now carries a body does not widen the
// address policy one bit — a POST to the LAN is refused for the same reason a
// GET is, unless AllowPrivate was set for that specific fetcher.
//
// hdr is merged over the defaults; a caller may not override User-Agent or
// Accept-Encoding, because identity encoding and an honest UA are part of the
// contract with whatever is on the other end.
func (f *Fetcher) Post(ctx context.Context, rawURL, contentType string, body []byte, hdr map[string]string) (*Response, error) {
if contentType == "" {
contentType = "application/json"
}
if hdr == nil {
hdr = map[string]string{}
}
merged := make(map[string]string, len(hdr)+1)
for k, v := range hdr {
merged[k] = v
}
merged["Content-Type"] = contentType
return f.do(ctx, http.MethodPost, rawURL, body, merged)
}
func (f *Fetcher) do(ctx context.Context, method, rawURL string, body []byte, hdr map[string]string) (*Response, error) {
u, err := url.Parse(strings.TrimSpace(rawURL))
if err != nil {
return nil, fmt.Errorf("webfetch: bad url %q: %w", rawURL, err)
}
if err := f.checkURL(u); err != nil {
return nil, err
}
if err := f.waitTurn(ctx, u.Hostname()); err != nil {
return nil, err
}
var rdr io.Reader
if body != nil {
rdr = bytes.NewReader(body)
}
req, err := http.NewRequestWithContext(ctx, method, u.String(), rdr)
if err != nil {
return nil, err
}
for k, v := range hdr {
req.Header.Set(k, v)
}
req.Header.Set("User-Agent", f.cfg.UserAgent)
req.Header.Set("Accept-Encoding", "identity")
resp, err := f.http.Do(req)
if err != nil {
// http.Client wraps our sentinels in *url.Error; unwrap so callers can
// still tell "blocked" from "the network is down".
for _, sentinel := range []error{ErrPrivate, ErrBlocked, ErrRedirects, ErrScheme} {
if errors.Is(err, sentinel) {
return nil, err
}
}
return nil, err
}
defer resp.Body.Close()
respBody, err := io.ReadAll(io.LimitReader(resp.Body, f.cfg.MaxBytes+1))
if err != nil {
return nil, err
}
if int64(len(respBody)) > f.cfg.MaxBytes {
return nil, fmt.Errorf("%w (%d bytes)", ErrTooLarge, f.cfg.MaxBytes)
}
if resp.StatusCode < 200 || resp.StatusCode > 299 {
return nil, fmt.Errorf("%w: %d", ErrStatus, resp.StatusCode)
}
out := &Response{
URL: resp.Request.URL.String(),
Status: resp.StatusCode,
ContentType: resp.Header.Get("Content-Type"),
Body: respBody,
Header: map[string]string{},
}
for k := range resp.Header {
out.Header[k] = resp.Header.Get(k)
}
return out, nil
}
// checkURL applies the scheme rule and the host lists. The address rule is the
// dialer's job (see New).
func (f *Fetcher) checkURL(u *url.URL) error {
switch u.Scheme {
case "http", "https":
default:
return fmt.Errorf("%w: %q", ErrScheme, u.Scheme)
}
host := strings.ToLower(u.Hostname())
if host == "" {
return fmt.Errorf("%w: no host", ErrBlocked)
}
if HostMatches(host, f.cfg.DenyHosts) {
return fmt.Errorf("%w: %s is denied", ErrBlocked, host)
}
if len(f.cfg.AllowHosts) > 0 && !HostMatches(host, f.cfg.AllowHosts) {
return fmt.Errorf("%w: %s is not on the allowlist", ErrBlocked, host)
}
// A literal private address is refused here as well as in the dialer, so
// the error is the specific one even when no connection is attempted.
if !f.cfg.AllowPrivate {
if ip := net.ParseIP(host); ip != nil && IsPrivateIP(ip) {
return fmt.Errorf("%w: %s", ErrPrivate, host)
}
}
return nil
}
// waitTurn blocks until this host's rate-limit interval has elapsed. It holds
// no lock while sleeping, so two hosts never wait on each other.
func (f *Fetcher) waitTurn(ctx context.Context, host string) error {
for {
f.mu.Lock()
now := time.Now()
earliest := f.last[host].Add(f.cfg.HostInterval)
if !now.Before(earliest) {
f.last[host] = now
f.mu.Unlock()
return nil
}
f.mu.Unlock()
wait := time.NewTimer(earliest.Sub(now))
select {
case <-ctx.Done():
wait.Stop()
return ctx.Err()
case <-wait.C:
}
}
}
// HostMatches reports whether host equals one of pats or is a subdomain of one.
// Exported because the crawler applies the same rule to links it decides not to
// follow, before it ever builds a request.
func HostMatches(host string, pats []string) bool {
host = strings.ToLower(strings.TrimSuffix(host, "."))
for _, p := range pats {
p = strings.ToLower(strings.TrimSpace(strings.TrimPrefix(p, "*.")))
if p == "" {
continue
}
if host == p || strings.HasSuffix(host, "."+p) {
return true
}
}
return false
}
// cgnat is 100.64.0.0/10 — carrier NAT, not covered by net.IP's helpers and not
// somewhere a personal assistant has business connecting.
var cgnat = &net.IPNet{IP: net.IPv4(100, 64, 0, 0).To4(), Mask: net.CIDRMask(10, 32)}
// IsPrivateIP reports whether ip is somewhere Maven must never reach out to:
// the box itself, the LAN, the wireguard range (10.42.0.0/24 ⊂ 10/8), the cloud
// metadata address (169.254.169.254 ⊂ link-local), or anything unroutable.
func IsPrivateIP(ip net.IP) bool {
if ip.IsLoopback() || ip.IsPrivate() || ip.IsUnspecified() ||
ip.IsLinkLocalUnicast() || ip.IsLinkLocalMulticast() ||
ip.IsInterfaceLocalMulticast() || ip.IsMulticast() {
return true
}
if v4 := ip.To4(); v4 != nil && cgnat.Contains(v4) {
return true
}
// IPv4-mapped/compatible forms of the above are handled by To4() inside the
// stdlib helpers; what is left is v6 unique-local (fc00::/7).
if len(ip) == net.IPv6len && ip.To4() == nil && ip[0]&0xfe == 0xfc {
return true
}
return false
}