F5 / Shape API Reference

Solves F5/Shape challenges. The endpoint is mounted under /f5 (http://host:3000/f5/...).

Like DataDome and unlike Akamai's sensor channel, F5 solving is single-shot HTTP — one request, one answer, no WebSocket session.

All providers are described in one machine-readable spec: download openapi.yml · browse interactively.

POST /f5/solve takes the two script bodies the protected page served you and a description of the request you want decorated, and returns the header bag the VM set on that request. You then send the request yourself, with those headers on it.

Authentication

None. The container is licence-gated at startup, not per-request — there's no API key on this endpoint. The boundary is the network you run it on. See Authentication.

The solver makes no outbound request

There is no proxy field on this endpoint, and that is not an omission. A solve fetches nothing: the launcher body, the VM body and the token body all have to arrive in the request body, and the result is a record of headers rather than a sent request.

So the proxy that matters is the one you use. Fetch the page, fetch the VM script, and send the decorated request through one pinned session.

POST /f5/solve

Request body

Field names are snake_case.

Field	Type	Required
page_url	absolute http(s) URL	yes
launcher_body	base64 — the launcher served inline from page_url	yes
vm_url	absolute http(s) URL the VM body was served from	yes
vm_body	base64 — the VM script body	no — omit it when one body is both launcher and VM, see below
request_url	absolute http(s) URL — the request to decorate	yes
request_method	HTTP method token, e.g. POST	yes
request_body	string, verbatim — not base64	no — absent means ""
profile_id	string	no
js_profile	JsProfilePayload	no
timeout	integer ms, 1–120000	no — default 20000
snapshot_points_ms	array of integer ms offsets, max 8	no — see Series
token_url	absolute http(s) URL	yes — see Supplying a token
token_body	base64 of the response body for token_url, non-empty	yes

A minimal request:

{
  "page_url": "https://www.example.com/checkout",
  "launcher_body": "<base64 of the inline launcher>",
  "vm_url": "https://www.example.com/assets/vm.js",
  "vm_body": "<base64 of the VM script>",
  "request_url": "https://www.example.com/api/order",
  "request_method": "POST",
  "request_body": "{\"sku\":\"A1\"}",
  "token_url": "https://www.example.com/api/session?id=…",
  "token_body": "<base64 of the response body>",
  "profile_id": "chrome-151-macos"
}

What you have to capture yourself

The solver has no HTTP client, so everything it reads comes from you:

1. Load the protected page and keep the inline launcher it serves → launcher_body.

2. Follow the script the launcher pulls and keep its body → vm_body, and the URL it came from → vm_url.

3. Decide which request you want decorated — the one the site's own script would have put the headers on → request_url, request_method, request_body.

4. Make the request the VM makes for its session token, and keep the response body → token_body, and the URL → token_url. See Supplying a token.

If the deployment serves one body that acts as both launcher and VM, send it once, as launcher_body, and leave vm_body out. Absent means fused. (Sending the same base64 in both fields is still understood.)

Two encoding rules that bite

launcher_body, vm_body and token_body are canonical base64 and are checked by round-trip, not by a character class. base64url, unpadded input, or anything with stray characters is refused by name with a 400 rather than silently decoded to a truncated prefix — a short VM body would otherwise compile and fail as an ordinary SyntaxError naming nothing.

request_body is not base64 — it reaches the VM exactly as you send it. Leave it out, or send "", for a request with no body; request_method already says whether one is expected.

Response

{
  "correlationId": "a1b2c3d4",
  "headers": {
    "x-abcd-a": "...",
    "x-abcd-b": "...",
    "x-abcd-c": "...",
    "x-abcd-d": "...",
    "x-abcd-f": "...",
    "x-abcd-z": "..."
  },
  "letters": ["a", "b", "c", "d", "f", "z"],
  "prefix": "x-abcd",
  "scope": "session",
  "tokenSupply": "delivered"
}

Field	Meaning
headers	Header name → value, spelled exactly as the VM spelled the name. Copy them onto your request verbatim.
letters	Which of the six attested letters (a,b,c,d,f,z) this solve produced. A short list is a real answer, not an error.
prefix	The family prefix as this deployment spells it, with no assumed x-. null when the VM set no attested header at all.
scope	session or target — see Retrying.
tokenSupply	delivered or unmatched — see Supplying a token.
correlationId	Quote this in a support ticket; it ties your response to the server-side log.
payloads	Present only when you asked for a series.
series_checks	Present only when you asked for a series.
series_observer_cost	Present only when you asked for a series.

The prefix and header names above are illustrative — read them off prefix and the keys of headers rather than hard-coding them, because they vary by deployment.

A partial answer is still an answer

The contract is that every letter is either produced with a value the VM set, or refused with a named reason — never invented. Six headers of which two were composed by us would be worse than four honest ones, so a solve that returns four returns four.

Supplying a token

F5 deployments have the VM fetch a server-issued value mid-collection and echo it into its payload. The solver can't make that request, so you make it yourself and hand over the response. Both fields are required:

{
  "token_url": "https://www.example.com/api/session?id=…",
  "token_body": "<base64 of the response body>"
}

A solve without a token would put a placeholder where a real session carries the value, so it is refused with 400 rather than run. token_body must be non-empty, decode to UTF-8 and be at most 64 KiB.

Only the origin and pathname of token_url are used

The query string carries your session's identity. It's discarded at the API boundary and reaches neither the solve nor the log. The URL is parsed for a string comparison and is never dereferenced — nothing fetches it.

That's also why matching is on origin + pathname: a whole-URL match would find nothing, since our URL and your browser's differ in the query by construction.

Read tokenSupply

The response tells you whether your body was used:

• delivered — the VM requested that origin + pathname and was given your body.
• unmatched — the VM never requested it. Your token_url is wrong, or this deployment doesn't make that request; the solve ran without your body.

Treat unmatched as a failed integration, not a result. Without this field a wrong token_url produces a completely ordinary-looking solve and you'd have no way to learn the parameter did nothing.

Tokens are per-session

A token body issued to one session is not a neutral input to another. Fetch it in the same session as the page, the launcher and the VM, and send the request you decorate through that session too.

Asking for a series

snapshot_points_ms fires the decorated request once at each wall-ms offset, inside one run, and returns one payload per point in payloads:

{ "snapshot_points_ms": [0, 150, 400] }

Rules the validator enforces:

• 1–8 points.
• Non-negative integers, strictly increasing. Two equal points name one place in the run; that's refused here rather than driven and refused after.
• The last point must be less than timeout, since the run waits it out in wall time. Otherwise you'd get a timeout, which reads as a slow subject rather than an unsatisfiable request.

Each entry of payloads carries the headers for that emission plus where in the run it came from:

Field	Meaning
index	Fire order, 0-based. Not an ordinal the VM knows about.
headers, letters, prefix	As the top level, but for this emission.
requested_at_ms	The offset you asked for.
observed_at_ms	Wall-ms at which it actually fired.
subject_handler_calls	Order and compare on this.
observed_handler_calls	The raw progress reading at this emission.
handler_calls_after_emit	The reading immediately after it returned.
displaced_handler_calls	How much of this point's reading was spent by the series' own earlier emissions. 0 on payloads[0].

Asking for more points moves the later ones

Every emission is synchronous and costs the run real progress, so points are not free samples — they slow the subject down, measurably. A dense series reaches a given wall-ms offset with the subject further behind than a sparse one does, and the payload at that offset is correspondingly less complete.

displaced_handler_calls per point and series_observer_cost for the whole series exist so you can see it happening. Ask for the fewest points that answer your question — three or four is a normal ask.

This is also why subject_handler_calls, not observed_handler_calls, is the field to order on: the raw reading counts our own emissions as well as the subject's work, so across differently shaped series it can invert. The two agree only on payloads[0].

There is no "give me payload #2"

The VM carries no emission ordinal — a payload exists only because a request was decorated — so the only thing that can be asked for is "fire again at T", and the only honest answer is "here's what came out, and here's where it landed".

And all the points come from one run, which is why they arrive together. One index per call would mean a fresh run per call, and pasting three such answers together would look like a session and would not be one.

When a series was requested, the top-level headers and letters arepayloads[0]. They are not a separate solve, and they are not what a no-series request would have returned. Send payloads in order and ignore the top-level bag.

series_checks lists which cross-snapshot checks ran, so "checked and held" is never confused with "never checked". A 200 with the field present means every check listed passed; a series that failed one is a failed solve and answers with an error. The verdicts themselves aren't on the wire.

Errors

Every failure carries a stable code you can branch on, plus correlationId.

{
  "code": "f5.solve.failed",
  "error": "F5 solve error",
  "correlationId": "a1b2c3d4",
  "scope": "session"
}

Code	Status	Meaning
f5.request.invalid	400	Your request didn't satisfy the parameter contract. error names the field.
f5.solve.timeout	500	The solve didn't finish inside timeout. Retryable.
f5.solve.failed	500	No usable headers. Check scope.

f5.request.invalid is the only code whose error text is specific — it describes your request, which is the only thing you can act on. Every other code carries static text; the cause is in the server log under correlationId. Don't parse the prose or branch on it, and don't expect the solver's internal reasoning on the wire — ask an operator to pull the log line.

Retrying

scope appears on f5.solve.failed and answers one question: can a later attempt change this?

• session — no. Stop; don't loop. This is also the default when no scope was measured, because failing closed costs one attempt rather than an infinite retry.
• target — the refusal is about the request you asked to decorate. A different request may succeed.

Limits

Request body	16 MB (app-wide)
timeout	1–120000 ms, default 20000
snapshot_points_ms	at most 8 points
token_body	at most 64 KiB decoded

The default timeout is a safety net, not a pacing mechanism — a solve runs to quiescence on its own and normally finishes well inside it. A ~317 KB VM body is ~423 KB once base64'd, comfortably inside the body limit.

See also

• How to integrate · Via the API
• Akamai · DataDome