# What is OAuth Token Exchange?

> Token exchange is an OAuth standard, RFC 8693, that lets a service trade a token it holds for a new token meant for the next service, either posing as the user or openly acting on the user's behalf.

Explainer by Boston Identity, 2026-10-15. 7 min read. https://insights.bostonidentity.com/glossary/token-exchange

- What it is: protocol
- Defined by: IETF, RFC 8693 (Standards Track, 2020)
- Who it's for: Teams wiring services and agents to APIs

## What is OAuth Token Exchange?

A front-end service gets a call carrying a user's access token and must call a back-end service for that same user. Token exchange is the standard way to get a credential for that second call: the service sends the token it holds to an OAuth authorization server and gets back a new one fit for the back end [S1 §1]. OAuth is the web's common framework for letting one application act at another without sharing passwords. RFC 8693, published by the Internet Engineering Task Force (IETF) in January 2020, lets its authorization server act as a security token service (STS), which checks the tokens it is given and issues new ones [S1 §1]. The new token might be scoped more narrowly for the downstream service, or be a different kind of token altogether [S1 §1].

The RFC separates two ways one party can act for another [S1 §1.1]. Under impersonation, service A gets a token that still names user B, and anyone receiving it is in effect dealing with B. Under delegation, A keeps its own identity and the token says A acts for B; in the RFC's words, "In a sense, A is an agent for B" [S1 §1.1]. Which of the two a deployment issues is left to policy at the authorization server [S1 §1.1].

Think of it like this: Think of a hotel front desk: you show a booking in your name and get a key card for one room, and if a colleague collects it, a delegation-style card would name you both. The comparison stops at narrowing, because a hotel always cuts a key for one room, while the RFC leaves how much the new token can do to the server's policy.

Not the same as: Not the same as forwarding the user's own token to the next service. In an exchange the authorization server checks the token it is handed and issues a new one under the policy for the service the client names [S1 §2.3].

## How does OAuth Token Exchange work?

The RFC's worked example has a front service holding a user's access token, an authorization server acting as the STS, and a back-end service the front service needs to call [S1 §2.3]. Whoever asks for the exchange is called the client, even when it is itself a server [S1 §1].

[Figure: The front service hands the user's token to the authorization server as subject_token, adds an actor_token only when it wants delegation, and calls the back end with the token it gets back.]

1. **A call reaches the front service with a token.** A request arrives at the front service with the user's OAuth access token in its Authorization header [S1 §2.3]. To reach the back end, the front service takes the client role and trades that token for one suited to the backend call [S1 §1].
2. **The client asks the authorization server for a new token.** The client posts to the authorization server's token endpoint with grant_type set to the token-exchange value [S1 §2.1]. It must send subject_token, a token for the party the request is made for, and subject_token_type, which says what kind of token that is [S1 §2.1]. It may name the target as a resource address or as an audience, a logical name for the service the token is for, and may ask for a scope and a token type (requested_token_type) [S1 §2.1]. If the server requires it, the client authenticates the normal OAuth way [S1 §2.1].
3. **The client adds an actor token to ask for delegation.** For delegation, the client also sends actor_token, a token for the party that will act, with actor_token_type [S1 §2.1]. The actor token is optional, but without it delegation is impossible [S1 App. A.1.1]. A may_act claim inside the subject token can name who may act for the subject, and the server can use it to decide whether to allow the request [S1 §4.4].
4. **The authorization server checks the tokens and applies policy.** The server must validate the subject token, and the actor token if there is one, by the rules for each token type [S1 §2.1]. In the RFC's example it reads the resource parameter to pick the policy for the target [S1 §2.3]. An invalid or unacceptable token must get the invalid_request error; a target the server will not issue for should get invalid_target [S1 §2.2.2].
5. **The authorization server returns the new token.** A successful reply carries the new token in access_token, with issued_token_type and token_type, and should give its lifetime in expires_in [S1 §2.2.1]. The name access_token is historical: the issued token need not be an OAuth access token [S1 §2.2.1]. If the scope granted differs from the scope requested, the reply must include it [S1 §2.2.1]. In a delegation issued as a JSON Web Token, sub names the user and an act claim names the acting party, nesting to show earlier actors [S1 §4.1].
6. **The front service calls the back end.** The front service presents the new token to the back-end service, and in the RFC's example that token is valid for one minute [S1 §2.3]. Unless the token type has its own rules, such as one-time use, the exchange does not change whether the original tokens are valid. Renewing the input token is not expected to extend the new one, and passing on a revocation is left to each implementation [S1 §2.1].

## Why does OAuth Token Exchange matter now?

Most of the agent delegation designs we have read come down to this one grant type [BI-FIND-2020-0004](https://insights.bostonidentity.com/ref/BI-FIND-2020-0004). The Model Context Protocol (MCP) now has an enterprise login extension in which an employee's AI client uses token exchange to trade the single sign-on token from the identity provider for an assertion the MCP server's own authorization server accepts [BI-FIND-2026-0678](https://insights.bostonidentity.com/ref/BI-FIND-2026-0678). The OAuth identity-chaining draft crosses into another company's domain by running an RFC 8693 exchange first and presenting the result there as a signed grant [BI-FIND-2026-0185](https://insights.bostonidentity.com/ref/BI-FIND-2026-0185).

Uber gives every agent hop a fresh token for one audience that lives five minutes, across thousands of internal agents, and reports that 99 percent of exchanges finish in under 40 milliseconds [BI-FIND-2026-0081](https://insights.bostonidentity.com/ref/BI-FIND-2026-0081). Auth0's Agent Gateway, in beta, uses token exchange to reach the software company's own APIs and MCP tools [BI-FIND-2026-0265](https://insights.bostonidentity.com/ref/BI-FIND-2026-0265). In June, Okta moved its agents' requests to Salesforce and MCP servers onto OAuth token exchange [BI-FIND-2026-1188](https://insights.bostonidentity.com/ref/BI-FIND-2026-1188).

At the Identiverse conference in June, practitioners settled on delegate, never impersonate: the agent gets its own identity, and the person's authority passes to it by token exchange at each hop [BI-FIND-2026-0469](https://insights.bostonidentity.com/ref/BI-FIND-2026-0469). MCP's core authorization specification, as of its 28 July 2026 revision, still has no token exchange and no actor claim [BI-FIND-2026-0189](https://insights.bostonidentity.com/ref/BI-FIND-2026-0189). One forecast we track says it gains a delegated-identity mechanism, such as a token-exchange profile, by September 2027 [BI-CALL-2026-0036](https://insights.bostonidentity.com/ref/BI-CALL-2026-0036).

## Where does it fit?

- Compared with Security Token Service: A security token service is the role: a service that checks tokens and issues new ones. Token exchange is the OAuth protocol for asking one; web services used the older WS-Trust protocol for the same job [S1 §1].
- Compared with Transaction Tokens: A transaction token is minted once, when a request comes in at the external-facing endpoint, and carried unchanged down the internal chain with no actor chain. Uber's token exchange issues a fresh token at every hop instead [BI-FIND-2026-0171](https://insights.bostonidentity.com/ref/BI-FIND-2026-0171).
- Compared with OAuth Identity and Authorization Chaining Across Domains: Identity chaining is built on token exchange: the client exchanges its token for a grant addressed to one authorization server in another domain, then presents it there for a local access token [BI-FIND-2026-0185](https://insights.bostonidentity.com/ref/BI-FIND-2026-0185).
- Compared with OAuth: Ordinary OAuth trades a person's approval for an access token. Token exchange adds a grant type that takes a token in and gives a different token back [S1 §1].

## What goes wrong?

Narrowing is not automatic. Nothing in the RFC requires the new token to carry less authority than the old one, and its security section only suggests the scope claim and short lifetimes to limit abuse [S1 §5] [BI-FIND-2020-0002](https://insights.bostonidentity.com/ref/BI-FIND-2020-0002). Karl McGuinness says that in production deployments today, child tokens routinely inherit the parent's full scope unless someone narrows them on purpose [BI-FIND-2026-0228](https://insights.bostonidentity.com/ref/BI-FIND-2026-0228). The RFC also leaves the tokens' format, their security properties and the trust model to profiles and to each deployment, including whether tokens are signed, encrypted or bound to a key [S1 §1]. Client authentication is a deployment choice too, and the RFC warns that leaving it out lets anyone holding a compromised token turn it into other tokens [S1 §2.1].

Impersonation hides the agent: whoever receives the token is dealing with the user [S1 §1.1]. Delegation records the agent, but for access decisions a consumer must look only at the top-level claims and the current actor, and earlier actors in nested act claims are informational only [S1 §4.1]. Ashay Raut of Amazon says an implementer may keep the actor chain in the token or drop it [BI-FIND-2026-0239](https://insights.bostonidentity.com/ref/BI-FIND-2026-0239), and George Fletcher says implementers already put the user and the agent the wrong way round between sub and act [BI-FIND-2026-0512](https://insights.bostonidentity.com/ref/BI-FIND-2026-0512).

Raut also argues that an exchange for every sub-agent would slow the whole job [BI-FIND-2026-0242](https://insights.bostonidentity.com/ref/BI-FIND-2026-0242), and no shipping standard carries the original person and the limits added along the way through a chain of agents [BI-FIND-2026-0070](https://insights.bostonidentity.com/ref/BI-FIND-2026-0070).

## What should a security leader ask?

1. When our services or agents exchange a token, does the authorization server narrow the scope and audience, or can the new token do everything the old one could?
2. Which of our agents act as the user rather than as themselves on the user's behalf, and can our logs tell the two apart?
3. Does our token endpoint require client authentication for token exchange, and which clients may exchange which users' tokens?
4. Are the tokens our exchange issues signed, encrypted or bound to the holder's key, or are they plain bearer tokens?
5. For a vendor: does your exchanged token name the user in sub and the agent in act, and do you keep or drop the earlier actors?

## Sources

- [S1] [OAuth 2.0 Token Exchange (RFC 8693)](https://www.rfc-editor.org/rfc/rfc8693) (IETF, Standards Track RFC, January 2020)

## Evidence

- [BI-FIND-2020-0004](https://insights.bostonidentity.com/ref/BI-FIND-2020-0004): Every delegation design in this corpus reduces to one grant type from a January 2020 RFC written for services, by engineers at Microsoft, Ping Identity, Yubico and Visa, that never mentions an agent. The standard defines an HTTP and JSON security token service: a client posts a subject token and, optionally, an actor token to the token endpoint and receives a new token for a named audience, with typed token identifiers so a SAML assertion can be traded for a JWT. The transaction-tokens draft mints its token through this exchange, the identity-chaining draft runs it before a JWT bearer grant to cross domains, Uber's exchange is conceptually based on it and its builder says the newer standards are not ready to implement as written, and the working group's agent draft composes it rather than replacing it. The agent-specific work is all profile; the base has not changed in six years. (from "The token exchange RFC that agent designs cite says its record of earlier actors is not for access decisions")
- [BI-FIND-2026-0678](https://insights.bostonidentity.com/ref/BI-FIND-2026-0678): MCP now has a stable way for a company's identity provider to stand between an employee's AI client and every MCP server. The client trades the employee's single sign-on ID token at the IdP for an ID-JAG, using OAuth token exchange, and hands that to the MCP server's own authorization server as a JWT assertion; the IdP checks group, role and conditional-access policy before issuing, and the employee sees one corporate login instead of a consent screen per server. The ID-JAG must be addressed to that authorization server, and when the client names a resource the access token must be restricted to it. The extension is opt-in: a server signals support in its authorization-server metadata. (from "MCP's enterprise login extension says IdP revocation is immediate. Its own spec does not back that up")
- [BI-FIND-2026-0185](https://insights.bostonidentity.com/ref/BI-FIND-2026-0185): Chaining defines no new token. A client in domain A takes the token it already holds to its own authorization server and, through RFC 8693 token exchange, receives a signed JWT grant whose audience is one authorization server in domain B; it presents that grant there under RFC 7523's JWT-bearer profile and gets a domain-B access token. The precondition is arranged in advance: B trusts A's signing keys, typically through A's published key set. The audience 'is restricted to a single authorization server in trust domain B to prevent an authorization server from presenting the client's authorization grant to an authorization server in a different trust domain', so a grant cannot be relayed on to a domain C. Uber's production exchange is 'conceptually based on' the first of those RFCs; ZeroID's delegation engine is built on it; the draft shows what the second RFC adds when the far end is someone else's server. (from "The cross-domain OAuth draft is one step from an RFC, and it never mentions agents")
- [BI-FIND-2026-0081](https://insights.bostonidentity.com/ref/BI-FIND-2026-0081): A large enterprise has per-hop token exchange with a full actor chain running in production for AI agents, and has published the numbers and the tokens. At Uber every agent hop takes a fresh audience-scoped token that lives five minutes (the two payloads in the post expire 300 seconds after issue) and carries the attested chain from the engineer down to the current sub-agent, each entry with its own issue time. Thousands of internal agents use it, and Uber reports a P99 of under 40 milliseconds on the exchange API even when one task makes dozens of tool calls; its own chart runs mostly under 10 ms, with one spike touching 40. Multi-hop delegation is no longer only a whiteboard problem. (from "Uber built the delegation chain the rest of the field calls unsolved")
- [BI-FIND-2026-0265](https://insights.bostonidentity.com/ref/BI-FIND-2026-0265): In Auth0's Agent Gateway design the agent never holds a credential that works anywhere but the gateway. It receives one short-lived token the gateway alone accepts; the gateway then checks policy before each tool call and obtains the real credential itself, from Token Vault for a customer's outside accounts such as Slack or Teams, and by token exchange for the software company's own APIs and MCP tools. Each call is tied to an agent, a customer organization and a user, so an agent working for one customer cannot use another's connection. A kill switch disables an agent or cuts its downstream access. The product is in beta by application, with no price or date. (from "Auth0's agent gateway keeps every credential away from the agent; the post shows no step where a person approves")
- [BI-FIND-2026-1188](https://insights.bostonidentity.com/ref/BI-FIND-2026-1188): Cloudflare and Okta both shipped ways in June for an agent or bot to authenticate with no person in the loop. Cloudflare's temporary accounts (June 19) let an agent with no Cloudflare credentials deploy Workers to a temporary preview account, with no sign-up, browser login or API token. A week later its MCP server portals accepted Access service tokens, so autonomous agents reach upstream MCP servers without a browser OAuth flow; the setup turns per-user authorization off, and the portal then uses the admin credential instead of a per-user grant. Okta's agent-to-agent connections, in early access in preview from June 17, let service apps call AI agents without user context, and its Auth0 unit made machine-to-machine access for third-party apps generally available on June 2 for callers that run without a user. In the same month Cloudflare shipped OAuth clients that act for a user after consent, and Okta moved its agents' requests to Salesforce and MCP servers onto OAuth token exchange, so the no-user routes sit beside delegated ones. (from "Google's agent registry went GA in June while Okta, SailPoint and Ping pulled agents into their own lists")
- [BI-FIND-2026-0469](https://insights.bostonidentity.com/ref/BI-FIND-2026-0469): Across its eight panels the summit settles one rule and names one unsolved problem. The rule is to delegate, never impersonate: an agent gets its own identity, the person's authority passes to it by token exchange (an OAuth step that swaps one token for another) at each hop from one party to the next, and the agent may do only what every identity in the chain is allowed to do. The unsolved problem is the second hop. Once an agent starts other agents, no shipping standard carries the person who started the job and the limits added along the way. An Amazon engineer works on the transaction-token draft, a draft standard for short-lived tokens that carry a request's context between services. He says the record of the chain is optional and survives only if the HTTP header has room. His own group is drafting a way for a parent agent to issue its children's identities, the opposite of the rule every other panel states. Uber built its own record of the chain and published it; nobody on the stage had built one. (from "Eight panels agree an agent acts for a person, never as that person. When agents hand work to other agents, nobody had an answer.")
- [BI-FIND-2026-0189](https://insights.bostonidentity.com/ref/BI-FIND-2026-0189): As of the 28 July 2026 revision, an MCP client is an OAuth 2.1 client and an MCP server is an OAuth 2.1 resource server, and that is the whole identity model. Authorization is optional; HTTP transports SHOULD use it, and 'implementations using an STDIO transport SHOULD NOT follow this specification, and instead retrieve credentials from the environment'. The authorization server is out of scope. There is no actor claim, no on-behalf-of flow, no token exchange and no transaction-token profile: nothing in the token says whom the agent is acting for, which is what the Identiverse summit panel meant by 'not a policy system' and what WorkOS meant by the gateway being undefined. Anything of that kind is routed to the new extension model, optional and additive, in a separate repository. This page is the baseline against which the summit's forecast, that the specification gains a delegated-identity mechanism within a year, will be graded. (from "MCP's authorization spec makes the agent a plain OAuth client and leaves delegation to an extension")
- [BI-CALL-2026-0036](https://insights.bostonidentity.com/ref/BI-CALL-2026-0036): Within 12 months the Model Context Protocol's authorization specification gains an explicit delegated-identity mechanism: a token-exchange profile, an actor or on-behalf-of claim, or a transaction-token profile, so that a downstream MCP server can see both the calling agent and the user it acts for. It settles wrong if, by September 2027, the specification still models the client as a plain OAuth client with no delegation element. (from "The rules cover one agent acting for one person; they break when that agent hands work to other agents")
- [BI-FIND-2026-0171](https://insights.bostonidentity.com/ref/BI-FIND-2026-0171): A transaction token is minted once, at the edge, and carried unchanged down the chain. The external-facing endpoint trades the inbound access token at the transaction token service for a short-lived JWT whose subject is the user or workload that started the transaction, whose req_wl claim names the workload that asked for it, and whose txn claim is the transaction's own identifier; every later workload forwards the same token in a dedicated Txn-Token header and verifies the signature itself. There is no actor chain. Uber's production design re-exchanges at every hop and grows an act_chain of user, on-call agent and investigation agent; ZeroID intersects scope at each delegation. The OAuth group chose the other design: one token, one transaction, and a replacement path it explicitly warns 'potentially negates the primary purpose'. For agents that spawn agents, this draft attributes the transaction to its initiator and says nothing about who acted in the middle. (from "OAuth's transaction token goes unchanged down a service chain and names who started it, not who acted in between")
- [BI-FIND-2020-0002](https://insights.bostonidentity.com/ref/BI-FIND-2020-0002): The two controls the agent designs lean on hardest, narrowing authority at each hop and controlling who may perform an exchange, are both left to the authorization server's discretion by the RFC. Nothing in it requires the issued token's scope to be a subset of the subject token's: scope in the request is optional and service-defined, the response rule only says when scope must be echoed back, and the security section suggests scope as an abuse mitigation without a normative keyword. Client authentication at the token endpoint is likewise a deployment decision, with the RFC noting only that omitting it lets anyone holding a stolen token launder it into more tokens. McGuinness's reading that token exchange supplies the plumbing and not the constraint is the text; Highflame's attenuation by intersection and Uber's rule that only its security token service may issue an agent token are policies those deployments added on top. (from "The token exchange RFC that agent designs cite says its record of earlier actors is not for access decisions")
- [BI-FIND-2026-0228](https://insights.bostonidentity.com/ref/BI-FIND-2026-0228): A delegated mission's child authority must be a subset of the parent's, and nothing in the token plumbing makes it so. McGuinness's reading is that OAuth token exchange, RFC 8693, carries a delegated token from one party to the next but leaves the narrowing to the implementation, and that in production exchange deployments today child tokens routinely inherit the parent's full scope unless someone explicitly attenuates them. He adds two things the corpus had not said: delegation degrades governability, because a child runs on a different runtime with weaker telemetry, so each hop needs its own trust-quality rating and a tighter budget regardless of its nominal bounds; and a sub-agent spawned at runtime has no clear issuer of child authority at all. Uber's production chain carries who asked through every hop and does not describe narrowing; Highflame's open-source layer makes attenuation by intersection an explicit feature, which is what a default would not need. (from "Containing an agent is not a backup to defining its task; where the agent cannot see everything, it comes first")
- [BI-FIND-2026-0239](https://insights.bostonidentity.com/ref/BI-FIND-2026-0239): The record of who delegated to whom is optional in practice, and what decides whether it survives is room in the HTTP header. Amazon principal engineer Ashay Raut, who works on the OAuth transaction-token draft, says an implementer doing token exchange may keep the actor chain in the token or drop it, and that a bank or a hospital that needs it should keep it somewhere else and decide for itself where. That matches the text: RFC 8693 treats earlier actors as informational only. Uber carries the full chain in every token; nothing obliges the next implementer to. (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-2026-0242](https://insights.bostonidentity.com/ref/BI-FIND-2026-0242): Standards work is under way to let an agent hand identities to the sub-agents it spawns without going back to a central issuer each time. Raut describes it as work his group is doing, with agents treated as workloads, and gives the reason plainly: a round trip to an authority service and a token exchange for every sub-agent would slow the whole job. That runs against the rule the rest of this corpus treats as settled. Uber lets only its token service mint an agent token, and an earlier panel at the same summit said an agent should never mint its own credentials. (from "Keeping an agent's delegation chain is optional, and room in the HTTP header decides whether it survives")
- [BI-FIND-2026-0070](https://insights.bostonidentity.com/ref/BI-FIND-2026-0070): Agents that spawn agents are the unsolved problem. Once an orchestrator hands work to sub-agents, the identity of the person who started the job is lost, separation-of-duties rules are broken by accident, and no shipping standard carries the original actor and the accumulated constraints through the chain. Every technical source in this corpus that models multi-agent systems says the same. (from "The rules cover one agent acting for one person; they break when that agent hands work to other agents")

---

© 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/token-exchange. 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
