# What is Transaction Tokens?

> A short-lived, signed token that says who started a request and what it is for. It is issued once, when the request arrives, and each internal service that handles it gets the same token and can check the original intent.

Explainer by Boston Identity, 2026-10-16, updated 2026-10-17. 7 min read. https://insights.bostonidentity.com/glossary/transaction-tokens

- What it is: protocol
- Defined by: IETF OAuth WG draft (-11), not yet a standard
- Who it's for: Teams running service chains or agents behind an API gateway

## What is Transaction Tokens?

Transaction Tokens, Txn-Tokens for short, come from a draft by the OAuth working group of the Internet Engineering Task Force (IETF). OAuth is the open standard most APIs use for delegated access. The document is a draft, not yet a standard, and its revision 11 is dated 30 July 2026 [S1]. A Txn-Token is a JSON Web Token (JWT), a signed set of claims. It names the user or workload behind a transaction and carries an authorization context that should stay the same for the whole transaction [S1 §4]. A workload is any running piece of software, such as a microservice or a managed database [S1 §3].

The draft is about what happens once a request is inside. One outside request often sets off a chain of calls between internal services, and a compromised network lets an attacker call services with no real transaction behind them, pose as a user or a service, change parameters, or steal and reuse access tokens [S1 §1]. With a Txn-Token, each service checks for itself who started the call and what it was for, and the signature shows whether anyone changed that context on the way [S1 §1]. The token is valid only inside one trust domain, a group of systems that share one set of security controls [S1 §3, §9.2].

Think of it like this: Think of a work order clipped to a job as it moves through a factory: every station reads the same order, sees who placed it and what was asked for, and no station may rewrite it. The comparison breaks on theft, because anyone holding a copy of a Txn-Token can present it, so a stolen one is limited mainly by its short life and narrow scope.

Not the same as: Not an OAuth access token. An access token grants authority to a client, while a Txn-Token carries the context of one transaction, and the draft says a workload must not use one as an access token or to prove its own identity [S1 §13.12, §13.13].

## How does Transaction Tokens work?

Three actors take part: the external endpoint where the request enters (often an API gateway), the Transaction Token Service (TTS) that issues the tokens, and the internal workloads that do the work. Each trust domain must have exactly one logical TTS [S1 §3, §10].

[Figure: A Txn-Token is minted once at the external endpoint and carried unchanged through the internal workloads.]

1. **A caller reaches the external endpoint.** An outside client calls the external endpoint with an ordinary credential, such as an OAuth access token [S1 §8.1]. A transaction can also start inside the domain. A scheduled job, for example, may present a self-signed JWT in place of an inbound token [S1 §5.1, §8.2].
2. **The endpoint asks the TTS for a token.** The endpoint sends the TTS a token exchange request, the OAuth method in RFC 8693 for trading one token for another [S1 §11.1]. The request names the trust domain as the audience and states the transaction's purpose as the scope. It passes the inbound token as the subject token, with the request's context and details alongside. The endpoint and the TTS must authenticate each other [S1 §11.5].
3. **The TTS checks the request and mints the token.** The TTS must verify the requesting workload and the subject token, and decide whether that workload may have the token it asked for [S1 §11.3]. The new token's scope must be equal to or less than the subject token's, and when the TTS can't tell that scope it must refuse [S1 §13.6, §13.14]. It must set a unique transaction id (txn), the subject (sub), the trust domain as audience and an expiry, and the token names the requesting workload (req_wl) [S1 §9.2, §11.3]. The TTS should also evaluate the request's details, and it may build the fixed transaction details (tctx) from them. The response must not include a refresh token [S1 §11.4].
4. **The endpoint passes the token inward.** The endpoint calls the first internal workload and sends the Txn-Token in an HTTP header of its own, named Txn-Token [S1 §8.1, §12.1]. The standard Authorization header must not carry it, because workloads may need that header for service-to-service authorization [S1 §12, §16.4].
5. **Each workload checks the token and passes it on.** Every workload that receives the token must check its signature, confirm that its audience is the workload's own trust domain, and confirm that it has not expired [S1 §12.2]. The workload then decides whether the token covers the requested action, and the draft leaves that decision to the deployment. Any call it makes further down must carry the same token, unmodified [S1 §12.2]. Responses travel back up the chain to the outside caller [S1 §8.1].

## Why does Transaction Tokens matter now?

The draft is close to done. Revision 11 carries the status 'WG Consensus: Waiting for Write-Up', and the working group plans to send it to the IETF's steering group in December 2026 [BI-FIND-2026-0169](https://insights.bostonidentity.com/ref/BI-FIND-2026-0169). George Fletcher, one of its authors, says it is still some months from becoming an RFC [BI-FIND-2026-0514](https://insights.bostonidentity.com/ref/BI-FIND-2026-0514). We forecast that the group submits it by 31 March 2027 [BI-CALL-2026-0052](https://insights.bostonidentity.com/ref/BI-CALL-2026-0052).

Shipping products have not caught up. Fletcher points to the open-source Keycloak server and the open-source Tokenetes token service as places to start, and he does not yet know of vendors that build transaction tokens into their stacks [BI-FIND-2026-0514](https://insights.bostonidentity.com/ref/BI-FIND-2026-0514). Buyers ask anyway: two buyers on one conference panel said a vendor of agent or non-human identity tools is expected to speak to transaction tokens [BI-FIND-2026-0468](https://insights.bostonidentity.com/ref/BI-FIND-2026-0468).

The debate about AI agent identity keeps reaching for this draft, though the draft never mentions an AI agent [BI-FIND-2026-0170](https://insights.bostonidentity.com/ref/BI-FIND-2026-0170). Ashay Raut argues that most estates authenticate at the edge and let internal services trust each other, so one compromised service or agent reaches everything behind it [BI-FIND-2026-0240](https://insights.bostonidentity.com/ref/BI-FIND-2026-0240). Where the agent itself goes in the token is still open. Fletcher says no profile yet names the claim for it [BI-FIND-2026-0512](https://insights.bostonidentity.com/ref/BI-FIND-2026-0512), and Raut's own draft adds a claim that records the originator, the agent acting now and a hop count [BI-FIND-2026-0530](https://insights.bostonidentity.com/ref/BI-FIND-2026-0530).

## Where does it fit?

- Compared with OAuth Token Exchange: Token exchange (RFC 8693) is OAuth's general method for trading one token for another. Txn-Tokens are a profile of it: a workload that asks a remote TTS over HTTP must use token exchange, with a token type the draft defines [S1 §11, §11.1].
- Compared with OAuth Identity and Authorization Chaining Across Domains: Identity chaining carries a user's identity and authorization from one trust domain to another. A Txn-Token stops at its own domain's edge, and the draft says a workload that needs the context elsewhere should use identity chaining [S1 §14].
- Compared with WIMSE Workload Proof Token: A Workload Proof Token proves that a calling workload holds a key, while a Txn-Token must never be used to authenticate a workload [S1 §13.12]. The two fit together: the draft lists the proof token as one way a workload can authenticate to the TTS [S1 §11.5], and the proof token can carry a hash of the transaction token in the request [BI-FIND-2026-0378](https://insights.bostonidentity.com/ref/BI-FIND-2026-0378).

## What goes wrong?

- A stolen Txn-Token works for whoever holds it. The draft says the token is not resistant to replay, so its lifetime must be kept short [S1 §13.1]. Fletcher's guide is the shortest lifetime that covers 99.99% of transactions [BI-FIND-2026-0513](https://insights.bostonidentity.com/ref/BI-FIND-2026-0513).
- A workload can refuse a txn value it has already seen, but that may not be practical when copies of one service share no state [S1 §13.2].
- A Txn-Token must not contain the access token presented at the external endpoint, because an attacker could extract it and replay it to any service that accepts it, even after the Txn-Token expires [S1 §13.4].
- A fake TTS would collect the access tokens that workloads send it, so workloads must authenticate the TTS and must get its address from a trusted source [S1 §13.7, §13.9].
- A TTS may accept an old Txn-Token as the subject of a new request. It must not widen the permitted actions or change txn, sub or aud, and it should limit how often a token is replaced with a longer life [S1 §13.15].
- One implementer found that revision -08 said nothing about a subject token with no scope claim, and his code let any scope through; revision -10 now says the TTS must reject that request [BI-FIND-2026-0893](https://insights.bostonidentity.com/ref/BI-FIND-2026-0893).
- How a workload turns the token into an allow or a deny is outside the draft, so the token alone enforces nothing [S1 §12.2].
- Complete tokens must not be logged verbatim, and claims such as the caller's IP address may count as personal information [S1 §15.1, §15.2].

## What should a security leader ask?

1. Which internal services accept a request today only because it came from inside our network, with no check of who started it?
2. If we run a Transaction Token Service, which workloads may ask it for tokens, and who approves that list?
3. How long would our Txn-Tokens live, and does any service refuse a transaction id it has already seen?
4. When an AI agent acts in a chain of calls between our services, which claim in the token names it, and does that match what our vendors chose?
5. Does any product we are evaluating issue transaction tokens as the IETF draft specifies, and against which revision?

## Sources

- [S1] [Transaction Tokens (draft-ietf-oauth-transaction-tokens-11)](https://datatracker.ietf.org/doc/html/draft-ietf-oauth-transaction-tokens) (IETF OAuth Working Group, Internet-Draft, revision -11 (30 July 2026); a draft, not yet a standard)

## Evidence

- [BI-FIND-2026-0169](https://insights.bostonidentity.com/ref/BI-FIND-2026-0169): Transaction Tokens have cleared the OAuth working group. Revision 11, dated 30 July 2026, carries the status 'WG Consensus: Waiting for Write-Up', and the group's milestone is to submit the draft to the Internet Engineering Steering Group in December 2026. The authors are Atul Tulshibagwale of CrowdStrike, George Fletcher of Practical Identity and Pieter Kasselman of Defakto, with Ping Identity's Brian Campbell as a named contributor, and the document history from revision 04 to 11 reads like a spec settling: scope-widening prohibited, cross-domain guidance added, terminology aligned with the WIMSE workload-identity work. This is the draft the Identiverse summit's practitioner told the room to read first, and the one Uber says its production token exchange is 'conceptually based on' via RFC 8693. (from "OAuth's transaction token goes unchanged down a service chain and names who started it, not who acted in between")
- [BI-FIND-2026-0514](https://insights.bostonidentity.com/ref/BI-FIND-2026-0514): Fletcher puts the base specification past working-group last call and in shepherd write-up, with the normative text settled and "a number of months" before it is an RFC. Products lag it: he cites the Keycloak open-source server's implementation and the open-source Tokenetes token service as places to start, and says he does not yet know of vendors shipping transaction tokens built into their stacks. Crossing to a partner such as Salesforce is the job of his separate chaining profile, which trades the token back home for something the partner's server accepts. (from "Transaction tokens are nearly a standard, and there is still no agreed place in them for the agent")
- [BI-CALL-2026-0052](https://insights.bostonidentity.com/ref/BI-CALL-2026-0052): By 31 March 2027 the OAuth working group submits draft-ietf-oauth-transaction-tokens to the IESG, visible on the datatracker as 'Publication Requested' or later. The group has consensus, the shepherd write-up is the only step left, and the milestone says December 2026; three months of slack covers a normal slip. Wrong if the datatracker still shows the draft at 'Waiting for Write-Up' or earlier on that date. (from "OAuth's transaction token goes unchanged down a service chain and names who started it, not who acted in between")
- [BI-FIND-2026-0468](https://insights.bostonidentity.com/ref/BI-FIND-2026-0468): What changes when the purchase is an agentic or NHI tool rather than classic IAM is the vetting list, on the account of one panel's two buyers. The incumbent platforms get checked first. A vendor is expected to speak to transaction tokens, CAEP and Shared Signals, identity fabric and intent, and to answer what has to be in place before the tool works, because the buyers' own experience is that they buy the Cadillac and implement the Chevy. The standards that two days earlier the draft authors pitched from the same stage are now a buyer's checklist. (from "Vendors all claim to cover AI agents and non-human identities, so buyers test them in their own environment")
- [BI-FIND-2026-0170](https://insights.bostonidentity.com/ref/BI-FIND-2026-0170): The draft never mentions an AI agent. Its threat model is a compromised internal network, an insider or a supply-chain defect invoking workloads outside any real transaction, impersonating a user or a service, tampering with parameters, or replaying a stolen access token further down the chain, and its actors are workloads and the people whose requests start a call chain. Yet it is the standard the agent-identity conversation keeps reaching for: the summit panel's practitioner named it alongside the agent-auth draft, Uber tracks it, and the Ping event's doctrine that authority must pass through every hop without being lost or inflated is, in the draft's language, 'context preserved throughout the call chain'. What the OAuth group has standardised is the container. Whether an agent's task fits in it, and who fills the scope claim for an agent that chooses its own next call, is the question the agent proposals still own. (from "OAuth's transaction token goes unchanged down a service chain and names who started it, not who acted in between")
- [BI-FIND-2026-0240](https://insights.bostonidentity.com/ref/BI-FIND-2026-0240): The people writing the transaction-token draft now present it as the way to tell that an AI agent is in a call chain. Raut's argument is that most estates authenticate at the edge and let internal workloads trust each other, so one compromised service or agent reaches everything behind it; a transaction token carries the call's context from the edge to the last workload, and policy can then act on it. He does not say the token marks an agent, and the draft we read never mentions one: it is minted once at the edge and has no actor chain. Detecting the agent is left to whatever context the edge chose to put in. (from "Keeping an agent's delegation chain is optional, and room in the HTTP header decides whether it survives")
- [BI-FIND-2026-0512](https://insights.bostonidentity.com/ref/BI-FIND-2026-0512): Fletcher's rule for agents is not to hand them the user's credentials or access token, and to keep the agent's authentication apart from the authorization it carries. Where the agent then appears in a transaction token is open: the token can carry it, but "there is no profile that says we're going to call that claim X", and when a gateway rather than the agent requests the token the original caller has no standard place. He points to Ashay Raut's agents draft, which adds an actor chain and agent context, says Karl McGuinness's actor proposal probably needs merging with it, and notes that implementers already put the user and the agent the wrong way round between sub and act. (from "Transaction tokens are nearly a standard, and there is still no agreed place in them for the agent")
- [BI-FIND-2026-0530](https://insights.bostonidentity.com/ref/BI-FIND-2026-0530): Raut's draft answers where an AI agent goes in a transaction token, and the answer is three fields, not a chain. The act claim, copied once from the access token, names the agent the user delegated to and never changes; in an autonomous flow it is absent and the agent is the subject. A new agentic_ctx claim then carries the originator, the agent acting now and a hop count that rises by one at each agent hand-off. The agents in between are not recorded anywhere in agentic_ctx, and "a path of agent identifiers" is explicitly left undefined. Only the comma-separated list of requesting workloads from the base specification keeps a trail. (from "Raut's draft records the first and current agents in the transaction token, and only counts the ones between")
- [BI-FIND-2026-0378](https://insights.bostonidentity.com/ref/BI-FIND-2026-0378): A Workload Proof Token proves that the caller holds the key in its credential, and that is the limit of what it proves. The signed JWT carries the target URL as audience (query and fragment stripped), a hash of the workload's identity token, and hashes of any transaction token or other token in the request; it does not cover the HTTP method, the path beyond the audience, headers or the body. Section 3.3 states the consequence without softening it: the proof 'does not protect major portions of the request and response and therefore does not provide protection from an actively malicious middle box'. A workload that chooses this proof over the HTTP message-signature profile is trusting every proxy between it and the recipient with the content of the call. (from "The IETF's simplest workload proof proves the key and leaves the request unprotected, and the draft says so itself")
- [BI-FIND-2026-0513](https://insights.bostonidentity.com/ref/BI-FIND-2026-0513): The design trades proof for deployability. Transaction tokens are bearer tokens "for deployment and performance reasons", so a stolen one is limited by its narrow scope and a lifetime the spec deliberately leaves open; Fletcher's guide is the shortest time that covers 99.99% of transactions, and a token that ever shows up at the front door should be dropped and alarmed on. The token carries data and decides nothing: each service is its own decision point, and fine-grained policy is "not a requirement to start". Revocation and narrowing across hops never come up. (from "Transaction tokens are nearly a standard, and there is still no agreed place in them for the agent")
- [BI-FIND-2026-0893](https://insights.bostonidentity.com/ref/BI-FIND-2026-0893): Transaction tokens -08 required the token service to keep a new token's scope "equal or less than" the scope in the incoming subject token, but said nothing about a subject token with no scope claim, which Agnihotri notes is normal for an OpenID Connect ID token. Read literally, his implementation skipped the check and let any requested scope through, which he calls "arguably the unintentional-scope-increase case" the rule exists to prevent; he had no test for it. He raised it on the OAuth list on 24 June and in a GitHub issue. The issue closed on 7 July, and revision -10 (29 July 2026) added a section saying the service MUST reject the request when the subject token's scope cannot be determined and MUST NOT treat it as unconstrained. (from "One implementer found a transaction-token rule that let any scope through; the OAuth group fixed it in weeks")

---

© 2026 Boston Identity LLC. All rights reserved. Published by Boston Identity LLC.
Quote up to 100 words with the credit "Boston Identity, IAM Insights" and a link to https://insights.bostonidentity.com/glossary/transaction-tokens. Claim numbers may be cited freely. Training, fine-tuning and text-and-data-mining use of this page is permitted.
Terms: https://insights.bostonidentity.com/terms · How to cite: https://insights.bostonidentity.com/cite
