Atlas

The operational findings Atlas raises on the operations view, as the key's member sees them in the app: admins see the workspace; managers see findings on the campaigns, creators and tasks they can see. Financial findings (domain financials, or financial sensitivity) are returned only to an admin key that also holds finance:read; otherwise they are left out (a page can then be shorter than limit). Filter by status, severity or domain; page with the returned next_cursor (limit ≤ 50, default 25). Read-only: acknowledging or resolving a finding happens in the app. Admin and manager members only. Requires the Atlas plan feature (403 feature_not_available otherwise).

MCP tool

list_atlas_findings

Scopes

atlas:read

Members

admin, manager

Idempotency-Key

Ignored (read)

Rate weight

1

Untrusted fields

data[].title data[].summary

The workspace's PENDING approvals as the key's member sees them in the app's approvals list (/agents?approvals=1), newest first: what is waiting, who or which key asked, when it lapses, and whether this member could decide it in the app (can_decide). An approval past its expires_at is left out (it can no longer be decided). Without the Atlas plan feature only the approvals Public API credentials queued are listed, as in the app. Payloads are not returned. Read-only: there is no approve or reject over the API — a person decides in TLNT; to follow something THIS key queued, poll get_action / GET /actions/{action_id}. Page with next_cursor (limit ≤ 50, default 25; a page can be shorter than limit when some approvals are not visible to the member).

MCP tool

list_atlas_approvals

Scopes

atlas:read

Members

admin, manager, scout, viewer

Idempotency-Key

Ignored (read)

Rate weight

1

Untrusted fields

data[].requested_via.name

Runs Atlas once, as the key's member, with the same budget, throttle and policy as asking Atlas in the app: it uses one Atlas run credit from the workspace's monthly Atlas allowance (the same allowance the app draws on; a failed run is billed like a failed run in the app), counts against the per-workspace and per-member Atlas run limits (429 atlas_run_throttled with Retry-After when busy), and is refused before any model call when the allowance is used up (402 usage_limit_exceeded). Atlas only uses tools — and only sees workspace context — in areas this key's scopes cover (no inbox tools without inbox:read, no proposed sends without email:send; money, finance tools and the agency's saved memory notes only on an admin member's key with finance:read). What Atlas does follows this credential's approval settings: an email reply, outreach email or inbox thread change runs directly when the credential runs that class directly (sends within its daily recipient cap; past it, or for a member who cannot run it, it waits for approval in TLNT); paid Discovery searches always wait for a person; everything else follows the workspace's own Atlas rules, as in the app. Every action is listed in `actions` with its action_id and status — get_action / list_actions report its outcome. `answer` is model-generated text (wrapped as untrusted content): verify specifics with the read tools before acting on them. Each call is a fresh, single-turn run (no conversation history). 403 atlas_disabled when Atlas is switched off for the workspace, 403 feature_not_available without the Atlas plan feature. Requires an Idempotency-Key (400 idempotency_key_required without one): a retry with the same key returns the first answer instead of running (and billing) again, and while the first request is still running — or never answered — the key answers 409 idempotency_in_progress (Retry-After: 10) for up to 15 minutes. The run gets at most 240 seconds from the moment the request arrives, including any wait for a free run slot. Admin and manager members only. Rate weight 5.

MCP tool

ask_atlas

Scopes

atlas:run

Members

admin, manager

Idempotency-Key

required

Rate weight

5

Untrusted fields

data.answer data.actions[].title data.actions[].detail