A2A tokens in the backend
The module AI > A2A tokens issues, shows and revokes the bearer
tokens of the A2A server role
(ADR-294). Administrators use it for every client record.
Editors use it, when an administrator allows it, for tokens that act as
their own account. The command nrllm: stays for
installations without a browser.
For administrators
The module lists every client record: title, kind (technical user or an editor's own account), owner, creation, expiry, status and the time of the last task. Reaching it takes the module ticked in the group's module list; an administrator needs no more.
- Generate token, Rotate token
- Per row of a client record with a technical backend user. The token is shown once in a panel above the list, with a Copy button. It is not stored, not logged and not shown again. Rotating ends the previous token at once; the client keeps its tasks, which belong to the record. The route that does this checks the administrator flag on the server and carries the core's request token, so another site cannot trigger it.
- Revoke
- Per row of any record. The record is disabled and its token hash is deleted: the token stops working at once and cannot come back. Use the Enabled field of the record to switch a client off temporarily instead.
An administrator cannot generate or rotate the token of an editor's own
record, and nrllm: refuses it as well. Such a token acts
as the editor, and only the editor issues it. An editor holds at most ten
active tokens (enabled, with a token, not expired); a revoked or expired token
does not count.
Every issuing, rotation and revocation is logged at level info by the
channel of A2a with the actor, the record and the time. TYPO3
drops that level by default, so the extension writes the channel to its own
file, var/; an installation changes
that through $GLOBALS like any other channel. In
switch-user mode the administrator who switched is named as well. A token
and its hash are never part of a log line.
Letting editors issue their own tokens
An editor may create A2A tokens that act as their own backend account, see them, and revoke them. Nobody else's. This is off until an administrator acts, and it is limited by default.
- On the editor's backend group, tick the permission
Issue A2A tokens for my own account
(
a2a_) under AI (nr_llm) permissions, and the module A2A tokens in the group's module list. The module switch alone grants nothing.token_ self_ service - On the same group, open the tab A2A tokens and set the limits below.
Without step 2 the permission allows a token that runs in a draft workspace, expires within 90 days, and calls no profile. Such a token can do nothing, so the editor sees that an administrator has to allow a profile first.
The limits
All four are fields of the backend group. A backend group is an administrator-only record, so an editor cannot relax their own limits. Every field starts at its strict value.
| Field | Default | Meaning |
|---|---|---|
| Allow tokens in the live workspace | Off | Off: the token must be pinned to a draft workspace the editor can work in, so what an agent changes is a draft the editor reviews. On: a token may be pinned to live. A token in live cannot call a writing profile. |
| Maximum validity of a token (days) | 0 (means 90) | How far ahead the expiry may lie. At most 730, whatever the database holds. |
| Allow tokens that never expire | Off | Off: every token needs an expiry within the maximum. On is an explicit choice: such a token is a permanent credential without a login. |
| A2A profiles tokens may call | Empty | The publication profiles a token may be allowed. Empty: none. |
The draft workspace needs the system extension workspaces. Without it
there is only live, and tokens exist only where a group allows live.
Why these are fields on the group, and not user TSconfig, is recorded in ADR-294: the form validates them, they sit next to the permission, and a misspelled TSconfig key would read as "not set".
When a user has several groups
The limits are merged over the groups of the user that hold the permission; TYPO3 resolves them, subgroups included. A group that sets a relaxation but does not hold the permission contributes nothing.
- Live is allowed when any of these groups allows it.
- The maximum validity is the largest value any of them sets. A group that sets none does not widen what another sets; if none sets one, it is 90.
- No expiry is allowed when any of them allows it.
- The profiles are the union.
When a limit is checked
Twice, on the server, and never by the form alone.
- When the token is created
- A request beyond a limit is refused and says which. The editor cannot pick another owner: the owner is the signed-in user. The workspace must be one the editor can work in.
- On every request
- The limits are read from the owner's groups as they are now. A token whose owner lost the permission, became an administrator, cannot work in the pinned workspace any more, or whose expiry lies beyond the maximum, is refused like an unknown token. The same check runs when a worker starts a queued run of the token and on every resume of a suspended run. A run that is already executing is not interrupted; it stops at its next resume. A profile the groups no longer allow is taken out of the token, and the remaining profiles stay callable. When the groups allow it again, the token works again.
Lowering the maximum validity affects tokens that already exist: a token whose expiry now lies further ahead than the maximum is refused until the maximum is raised again. The editor creates a new token with a shorter validity.
The module shows such a token as Withdrawn, with the reason.
Existing clients with a technical backend user are not affected by any of this.
What a token means for security
Warning
A bearer token bypasses the interactive login of the backend, including multi-factor authentication and passkeys. Whoever holds the token acts as its owner, with everything the owner's permissions allow in the pinned workspace and the allowed profiles, until it expires or is revoked.
That is why the defaults are what they are:
- Expiry is mandatory. A token that is forgotten in a script or a chat stops working on its own, within 90 days unless an administrator raised the maximum.
- A draft workspace is the default. An agent that acts with the editor's rights writes drafts that the editor publishes, not live content.
- No profile by default. A token can call only what an administrator published for that group.
- Revocation is immediate, by the editor or by an administrator, and the token itself is never stored: the database holds its prefix and its SHA-256.
Treat a token like a password: paste it only where the agent reads it.
Not implemented: OAuth 2.0 Authorization Code
The better long-term answer is the OAuth 2.0 Authorization Code flow with PKCE. The editor would then authorise the agent interactively; the login, multi-factor authentication included, happens at that step, and the access the agent gets can be cancelled by the group. It needs an authorisation endpoint that nr_llm does not have, and it is recorded in ADR-294 as a possible later step.
For editors
Open AI > A2A tokens. You see your own tokens and nobody else's.
- Give the token a title, for example the agent that will use it.
- Choose the workspace, the validity in days and the profiles. Only what your groups allow is offered, and the limits are checked again when you create the token.
- Choose Create token. The token appears once. Copy it with the button or by hand, then choose I have stored the token. If you lose it, revoke it and create a new one.
The column Last task shows when the newest task of the token was created. A token that only reads existing tasks does not show up there.
Choose Revoke to end a token at once.