Team-scoped gateway keys tag org, team and user identity for AI spend dashboard

A call under this design leaves the laptop with a team-scoped key aimed at `https://gateway.reclaimllm.com/v1/{org_slug}` ^[8]. The gateway resolves the org, substitutes a master provider credential held in its vault ^[1], checks the requested model against policy ^[6], stamps the record with org, team and user identity server-side ^[3], and forwards the request to Anthropic, OpenAI or Google ^[7]. Attribution is therefore a property of the credential, not of the person typing. The cost report has exactly the resolution of the key issuance policy. One key per team gets team lines where there were none ^[11]. One key for all of engineering moves the same blindness one hop closer to the code.

Two of the post's statements need a bridge it does not supply. Tagging is described as durable and includes user identity ^[3], while the keys themselves are scoped to teams such as `frontend-team` and `data-platform` ^[2]. Either keys are minted per developer in practice, or the client passes a user field with the request. If it is the second, the best property of the design goes soft: a server-side tag cannot be forged by a developer's machine, a client-supplied one can. That distinction is the reason to collect the record at all for a SOC 2 reviewer ^[13].

The post's own title promises real-time auditing. The number underneath it is a set of PostgreSQL materialized views refreshed every 15 minutes ^[4], which works out to 96 refreshes a day ^[10]. For monthly chargeback that is generous. For an agent stuck retrying a frontier model, it is up to a quarter hour of billable surprise before any aggregate tab shows it. The Sessions tab is described as a filterable raw session list for drilling behind a metric spike ^[5]; whether it reads base tables or the same 15-minute views is not stated, and that is the detail that decides if it is usable mid-incident.

Adoption cost sits in the clients, not the gateway. Each surface from Part 1 (Gemini CLI, Antigravity, Claude Code, Cursor, Codex, API proxies) has to accept a base URL you control and a key you mint ^[14]. That is six separate configuration questions, and the post enumerates the tools without enumerating their base-URL support.

Credit where it is due: running DLP before execution rather than after ^[14], and tagging server-side rather than trusting the caller ^[3], are both the harder and the correct side of the choice. The vault is an honest trade rather than a free win. Many laptops each holding a master key ^[12] becomes one store holding master credentials for OpenAI, Anthropic, Gemini and Azure at once ^[1]. Smaller attack surface, larger blast radius, and worth it provided the vault has a better on-call rotation than the laptops do. In my context I would still route everything through the gateway, and I would mint keys at the granularity I intend to bill at, because that granularity is the only thing the dashboard can ever report.