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.
`--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 it refuses, always, 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, IPv4 addresses carried inside IPv6 forms (::ffff:a.b.c.d, the well-known NAT64 prefix, 6to4), and the whole NAT64 local-use range;
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. There is no switch to turn this off: the proxy's private network is the cloud's, not the user's, so a localhost or LAN API can never be reached through it — that is what the desktop app is for.

Deploying

One binary, no dependencies:

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

There is no authentication: an instance is anonymous and protected by the per-client rate limit and the destination policy, which is what the hosted funnel wants. Anything more (a shared token, per-user quotas) is a later slice and would sit in front of send_http in main.rs. Put TLS in front of it (a reverse proxy). 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, capacity) are plain HTTP errors (403, 400, 429, 503) 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.

The TypeScript side of this contract is generated from src/wire.rs by ts-rs into bindings/ (run cargo test -p yaak-send-proxy after changing a frame) and published to the tab as @yaakapp-internal/send-proxy, so a change to the wire on one side is a type error on the other.

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 and limits. 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.