How Asgard AI works

Asgard AI sits between your AI agents and your real systems. The agent does the work through scoped, audited, revocable access — while your API keys, tokens and database credentials never leave the vault.

The core idea: capabilities, not credentials

Most "AI connector" setups hand the agent a secret — an API key, an OAuth token, a connection string. Once the agent holds it, you've lost control: the secret can leak in a prompt, a log or a model provider, and the agent can do anything that credential allows.

Asgard AI inverts this. The agent never receives a secret. It asks Asgard AI to perform an operation; Asgard AI performs it on the agent's behalf using a credential that stays encrypted in the vault, and returns only the result.

A request, step by step

1. The agent calls Asgard AI. Your agent connects over MCP or REST using a scoped Passport (an Asgard AI-issued key bound to one identity, not an upstream secret).

2. Policy is checked. Asgard AI resolves what that agent is allowed to do — which connection, which operations, which parameters — and applies the active guardrails.

3. The real call is made from the vault. Asgard AI retrieves the upstream credential, makes the actual API or database call, and the agent never sees the secret.

4. Content is filtered. Request and response are scanned for secrets and sensitive data; parameters can be pinned or stripped so the agent can't reach beyond its grant.

5. Everything is logged. Who, what, when, which resource, and the outcome — written to an immutable audit trail you can review and export.

The building blocks

Vault — credentials are encrypted at rest and never returned to the agent.
Envoys & Passports — each agent gets an identity and a scoped key, so access is least-privilege by default and attributable per agent.
Grants — you decide exactly which operations an Envoy may call on each connection, down to parameter constraints.
Guardrails — nine built-in controls including DLP/redaction, prompt-injection scanning on tool results, human-in-the-loop approvals and action quotas.
Audit & kill-switch — a complete action log, plus one-click revocation of any agent or the whole tenant, with automatic quarantine on a denial burst.

Works with the AI clients you already use

Any MCP- or REST-capable client connects: Claude, ChatGPT, Cursor, GitHub Copilot, Microsoft Copilot, Windsurf, Cline, Gemini and more. You choose which clients may connect, and which Envoy each one acts as.

Keep reading: Use cases → ISO 42001 → Asgard AI vs native connectors → Book a demo →