01
Create a read-only credential
Create the credential in OpenAI with the minimum read scope, paste it into devguard, and the connection is live. The integration page in the product spells out the exact permissions.
devguard reads your OpenAI organization through the Admin API: who is a member, which roles they hold, and whether admin keys are current, access governance for the AI tooling your team already uses.
4 read-only checks, run on a schedule. Each failure becomes a finding linked to a control — every check is listed, none are “coming soon”.
Connectivity check: confirms the Admin API key authenticates and reads the organization users endpoint (HTTP 200). A non-200 (e.g. 401) indicates an invalid key or a non-admin key.
admin-connectivity
Access-review check: fails if any organization member has a role outside {owner, reader}, or if the number of owners exceeds the approved maximum. Enumerates members with cursor pagination. Also fails if no members are returned, to avoid a vacuous pass.
member-access-review
Fails if the number of organization Admin API keys exceeds the approved maximum, bounding admin-key sprawl. Enumerates admin keys with cursor pagination. Also fails if none are returned, to avoid a vacuous pass.
admin-key-inventory
Fails if any organization project has a status outside {active, archived}, supporting a data-segregation inventory review. Enumerates projects with cursor pagination. Also fails if none are returned, to avoid a vacuous pass.
project-inventory
Bring your current platform's test list — we'll map it against these checks, gap by gap
Setup is minutes, not a migration project — and every step stays on your side
Create the credential in OpenAI with the minimum read scope, paste it into devguard, and the connection is live. The integration page in the product spells out the exact permissions.
The first run starts right after connecting, then repeats on a schedule. Results are timestamped, so your evidence has a history, not a snapshot.
Passing results attach as evidence to the controls they support; failures become findings you triage. When the audit comes, the proof is already connected.
Open any page for its full check register.
You create and scope every credential yourself, and you can revoke it at any time. Checks only read; results and evidence are stored in Switzerland and stay exportable.
Yes. The checks authenticate with a credential you create and scope yourself, and only ever read configuration. devguard holds no write access to OpenAI and remediates nothing silently.
A failed check becomes a finding linked to the control it supports, with the evidence attached. You decide the treatment — nothing pages your engineers and nothing is auto-remediated.
Checks run from Swiss infrastructure, and results and evidence are stored in Switzerland. Everything is exportable — your evidence stays yours.
20 minutes on how you run certification today. Bring your current test or control list and we'll map it against these checks — honestly, gap by gap.