Letting an outside AI SRE read your Bucker data
An AI SRE, an incident tool's agent or a coding agent that you already run elsewhere can read your
cross-pillar incident context from Bucker under an observer credential. It is read-only, it has
a spend limit, it ends on a date you choose, and every grant that makes it up is in your audit log
with your name on it. Bucker is the context source; the outside agent is yours (or your
processor's), not Bucker's.
Status. The profile is built and enforced in the domain layer; the dashboard screen and the API route that apply it are not shipped yet. Until they are, an operator applies it from a maintenance script running as an organization owner's own session.
What an observer can and cannot do
| It can | It cannot |
|---|---|
List the read tools of the core, issues, monitors and apm toolsets (their write tools stay hidden and refused, because it holds no write scope) |
Call any tool that writes, or hold any write scope, granted or still pending |
Fetch an evidence bundle (evidence_bundle_get): one bounded object for an alert, change, check run or issue, with priced handles to the rest of the incident |
Read person-level analytics rows (pii:read is human-only and cannot be granted) |
| Read through the MCP surface within the scan and token budgets you set | Exceed those budgets (it is refused, and paused until the period turns over), or read over the REST API, where no budget meters a read |
| Read the projects you pinned it to | Read any other project, or any other organization |
Everything served to it passes the agent-tier scrub, and every field that came from outside Bucker
(telemetry, customer-chosen keys, free text) is labelled untrusted in the response envelope's
trust.untrustedFields, because the receiving model belongs to someone else. Treat what comes back
as data, not instructions.
Set up
- Create an agent identity for the outside agent (Settings, Agents). Choose the capability tier
telemetry:read(orcontext:hydrate) and an org role ofMEMBERorVIEWER. A tier that carries a write scope is refused: the profile will not apply to it, nor to an agent holding a write grant, including one still inside its 24-hour activation delay. Do not issue it a credential yet. - Pin it, if you want to. Link the agent only to the teams of the projects it may read. If you pin projects when applying the profile, the apply is refused while the agent can still reach a project outside the pin, and tells you which.
- Apply the observer profile as yourself. Applying it is the same human-only act as any other
widening of an agent (
agent_grant.widen, you needmanage_agents, from your own session): it cannot be done by an agent, a delegated token or an OAuth client. You give the receiving vendor's name (it appears in each grant's reason and on the consent summary), a reason, and optionally a lifetime (90 days by default, 365 at most), a rows-scanned budget (QUERY_SCAN, default 200,000, at most 2,000,000) and a token budget (TOKENS, default 500,000, at most 5,000,000). The budgets are written first, then the four toolset grants and theevidence_bundle_getgrant. Applying it again is safe: grants you already hold live are skipped, and a budget already set is kept unless you name a new limit, so a re-apply never quietly raises a cap someone lowered. - Issue the agent a credential for the MCP audience, with only the read scopes it needs
(
issue:read,event:read,project:read) and the audiencebucker-mcp. The budgets are charged on the MCP surface only, and abucker-mcpcredential cannot be used against the REST API, so it cannot read around them. The apply is refused while the agent holds any live credential that can reach the REST API; revoke it and mint the MCP one. Hand the credential to the outside agent over a channel you trust. Bucker shows it once. - Point the outside agent at Bucker's MCP server. It exchanges the
bpat_credential for a short-lived token atPOST /agents/token(see MCP setup) and connects with that. Nothing was added to the MCP connection's scope ceiling for this: the three read scopes above were already inside it, and the ceiling stays frozen pending its security review.
The agent should begin at evidence_bundle_get with the id of the alert, change or issue it is
investigating. The bundle lists the next reads that reach the rest of the incident, with the exact
arguments and the price of each, so the agent can choose the cheapest set.
What it sees, for your consent screen
Issues and their evidence; traces, spans and release health; alert rules, SLOs and alert events; change events and incident links. It never sees person-level rows, billing or money fields, credentials, or another organization.
Ending it
Revoke the grants (narrowing is immediate, and the agent can do it to itself), delete the credential, or suspend the agent. Expiry and a grantor who has left or lost the scope end it on the agent's next request: nothing waits for a sweep. Lowering a budget is also a narrowing; raising one more than twofold alerts the organization's owners.
Checking it
The posture read (observerPosture) reports, live, whether the agent's whole ceiling is
read-only (pending grants included), any credential it holds that reaches the REST API (an
observer should hold none: one minted after the apply shows up here and marks the profile
incomplete), which toolsets and tools it holds, and how much of each budget is spent. Every MCP
call it makes is charged to its budgets and recorded as a spend row, never as the data.