Files
yaak-mountain-loop/crates-server/yaak-send-proxy
Gregory Schier 253b939b34 Add the browser send proxy and web sender
crates-server/yaak-send-proxy: a stateless executor over yaak-http's
HttpTransaction. It takes a rendered request, streams timeline events,
the response head, body chunks and the resulting cookies back as NDJSON,
and keeps nothing. Private/loopback/link-local/metadata ranges are refused
after DNS on every hop (an AddressFilter on the resolver plus a per-hop URL
check), with size caps, a timeout ceiling, a rate limit, host allow/deny
lists and an optional token.

The web host now sends through it: the wasm worker resolves and renders
the request (render_http_request moved into yaak-models so it builds for
wasm; re-exported from its old paths), the tab posts it, and stores what
comes back where the desktop stores it. Requests needing auth plugins or
template functions are refused with the reason until plugins run in the
browser.
2026-08-17 06:51:48 -07:00
..

yaak-send-proxy

The network half of Yaak in a browser.

A tab can't see an HTTP response the way a desktop app can: CORS hides most headers (2 of 8 in a typical response), redirects are followed silently, and there is no timeline. So the tab renders the request and posts it here, and this process puts it on the network with the desktop's own engine (yaak-http) and streams back everything that happened — every header, every redirect hop, DNS timing, the body — for the tab to store.

It is a stateless executor. It keeps nothing: no database, no files, no sessions, no cookies between calls. Every byte it sees comes from the tab in the request, and every byte it returns is stored by the tab. Restart it any time.

Running it

cargo run -p yaak-send-proxy

Listens on 127.0.0.1:9227. Then run the web build against it:

YAAK_TARGET=web npm run dev --workspace @yaakapp/yaak-client

The tab looks for the proxy at http://127.0.0.1:9227 unless VITE_YAAK_SEND_PROXY_URL says otherwise at build time.

Every flag has a YAAK_PROXY_* environment variable, so a container needs no arguments; --help lists them all.

Flag Default What
--bind 127.0.0.1:9227 Listen address. 0.0.0.0:9227 inside a container.
--allowed-origins * CORS origins, comma-separated. A hosted instance should name its web origin.
--token unset Require Authorization: Bearer <token>. Unset means anonymous, which is what the hosted funnel wants alongside the rate limit.
--allow-private-networks off Let sends reach private, loopback and link-local addresses. Off by default; see below.
--allow-hosts empty Only these hosts (api.example.com, *.example.com). Empty means any host not denied.
--deny-hosts empty Never these hosts. Checked before the allow list.
--max-request-bytes 16 MiB Largest rendered request accepted from the tab.
--max-response-bytes 64 MiB Largest upstream body relayed before the send is cut off.
--max-timeout-secs 60 Ceiling on a send's timeout; a request asking for more (or none) gets this.
--rate-limit-per-minute 120 Sends per client IP per minute; 0 disables.
--max-concurrent 256 Sends in flight at once.
--trust-forwarded-for off Take the client IP from X-Forwarded-For. Only behind a load balancer that sets it.

What it refuses, and why

A hosted proxy is, by construction, a machine that makes HTTP requests on behalf of strangers. Left alone that is an open relay into whatever network it sits on. So by default it refuses to connect to:

  • loopback (127/8, ::1), private (10/8, 172.16/12, 192.168/16, fc00::/7), link-local (169.254/16 — where cloud metadata lives — and fe80::/10), carrier-grade NAT, multicast, reserved and unspecified ranges, and IPv4 addresses tunnelled inside IPv6 forms (::ffff:a.b.c.d, NAT64);
  • anything not http:// or https://.

The check runs on the resolved addresses, after DNS, for every hop of a redirect chain, so a public hostname that points at an internal address is caught, and so is a Location: header that points at one. It also refuses body types that would read files on the proxy's disk (binary, multipart file fields), since no browser tab could legitimately mean those.

Refusals are logged with the reason. A self-hosted instance on a private network that legitimately needs to reach the services next to it turns the range check off with --allow-private-networks, and can narrow that with --allow-hosts.

Self-hosting

One binary, no dependencies. Build it and run it wherever you like:

cargo build --release -p yaak-send-proxy
YAAK_PROXY_BIND=0.0.0.0:9227 \
YAAK_PROXY_ALLOWED_ORIGINS=https://yaak.example.com \
YAAK_PROXY_TOKEN=change-me \
./target/release/yaak-send-proxy

Put TLS in front of it (a reverse proxy) — the token travels as a header. If the reverse proxy buffers responses, tell it not to: the reply is a stream and the X-Accel-Buffering: no header it sets is honoured by nginx-shaped ones.

The wire

POST /v1/http/send with a JSON body:

{
  "request":  { "url": "https://…", "method": "GET", "headers": [], "body": {}, "bodyType": null, "urlParameters": [] },
  "settings": { "validateCertificates": true, "followRedirects": true, "timeoutMs": 0, "sendCookies": true, "storeCookies": true },
  "cookies":  [  ] 
}

request is a Yaak HttpRequest in the desktop's own model shape with every template already rendered by the tab; the proxy builds the URL, headers and body from it exactly the way the desktop does after rendering. cookies is the jar's contents (or null for no jar).

The reply is application/x-ndjson, one JSON frame per line, in the order things happened:

type When Carries
event as the engine produces them one timeline event, in the desktop's http_response_event.event shape
response once, when the final hop's headers arrive status, all headers, request headers as sent, remote address, HTTP version, timing
body as the body is read a decompressed chunk, base64
done last, on success elapsed, byte counts, and the cookie jar as the send left it
error last, on failure the reason, and any cookies collected before the failure

Refusals that happen before anything is sent (a blocked destination, a bad body, rate limit, missing token) are plain HTTP errors (403, 400, 429, 401) with {"error": "…"}, not streams.

Why a streamed HTTP response and not a WebSocket: one POST is stateless by construction, cancellable by closing the connection, readable with curl, and needs no upgrade handling on either side. A WebSocket only earns its keep when traffic is bidirectional, which a single send is not.

GET /v1/health reports the version and the effective limits.

What comes later

Not built, by design, but the router is shaped for it: a WebSocket relay (/v1/ws/relay) and a gRPC relay (/v1/grpc/relay) would be long-lived, bidirectional endpoints on the same binary, behind the same destination policy, limits and token. They differ from this endpoint in holding per-connection in-memory state while a connection is open (never persisted), which brings connection limits and a larger abuse surface — the reason they are separate work.