Issue scoped, short-lived credentials to your AI agents. Test their access, verify their behavior, and audit every call — from one control plane.
Fullmakt is a credential broker for AI agents. It issues scoped, short-lived credentials to your agents on demand, verifies their access on every request, and records a tamper-proof audit of every call — so your secrets never reach the model and every action stays attributable.
If you are shopping by category: this is AI agent access control — the layer people also look for as secrets management for AI agents, non-human (machine) identity, or MCP server security. Fullmakt is not a replacement for your identity provider or your secret store. It sits between your agents and your APIs, decides what each agent may do on every single call, and proves what it did.
If any of these is the thing keeping your agents out of production, this is the layer that fixes it.
API key management for LLM agents, without handing the key over: the agent gets a scoped token that expires in an hour, and your real key stays in the vault and is injected server-side. How scoped credentials work →
An audit log of every AI agent API call — allowed and denied, with the agent, the policy decision and the endpoint on each line — so a compliance or incident question has an answer. What gets logged →
MCP server security that does not stop at the connection handshake: policy is evaluated on every tool call, so a rewritten tool description cannot quietly widen what the agent may do. The rug-pull problem →
Every agent is a distinct non-human identity with its own credentials, so revoking one takes a click — no rotating a shared secret and redeploying everything that used it. Agent offboarding →
Secrets management for AI agents assumes the secret never reaches the model in the first place: responses are sanitized before the agent sees them, and sensitive fields are masked in the audit trail. The observability trap →
A valid token is not a wise call. Per-request policy, SSRF guards and hold-for-human-approval rules bound the blast radius when an agent is confused or prompt-injected. The confused-deputy problem →
Your agent works with a short-lived scoped token. Your real keys stay in the vault and are injected server-side, only after the policy allows the call — and everything is audited.
Want the wire-level details — Agent Card, token exchange, MCP endpoints? Read the agent docs.
Broker credentials, prove they work, and prove what they did — without writing your own auth layer.
Mint short-lived, scoped credentials for each agent. Rotate, revoke, and expire on demand — secrets never touch the model.
Simulate agent calls against your APIs before they ship. Confirm scopes, expirations, and downstream behavior in a sandbox.
Continuously verify agent identity, scope, and policy compliance on every request. Block anything that drifts off-policy.
Every issuance, call, and revocation is logged with cryptographic chaining. Answer "which agent did what, when, on whose behalf?" instantly.
Define what each agent can touch — by endpoint, method, rate, or data class. Least-privilege by default, configurable per workflow.
Drop-in support for MCP servers, OAuth 2.1, and OIDC providers. Bring your existing IdP — agents inherit the same trust model.
Fullmakt speaks the protocols your agent already does. Point it at one URL and it can discover the rest on its own.
An Agent2Agent server with a well-known Agent Card. Your agent builds out collections, policies, and identities — while vault credentials stay human-only.
Every collection doubles as a remote MCP server: its requests become tools, brokered through your policies with short-lived credentials.
Client-credentials tokens scoped to a workspace or a single collection, expiring in an hour. Standard well-known metadata for spec-following clients.
curl https://fullmakt.ai/.well-known/agent-card.jsonFull walkthrough in the agent docs — or create an account and open Governance → Connect.
No plans, no minimums, no feature gates. Usage is metered and billed at the end of each month — our infrastructure cost plus a small margin. Try everything first in the no-card sandbox.
Every feature included. You're charged per operation:
Add a card when you're ready. Remove it or leave anytime — any remaining usage is settled on a final invoice.
Try the sandbox — no cardJoin teams brokering credentials for AI agents with Fullmakt — least privilege, fully audited, by default.
Create Free AccountA credential broker issues scoped, short-lived credentials to AI agents on demand, so the agent (and the model behind it) never holds your long-lived API keys or secrets. Fullmakt mints, rotates, and revokes these credentials, enforces what each agent may do, and records every call in a tamper-proof audit log.
Identity providers authenticate users and services. Fullmakt sits one layer out: it brokers per-agent access to your APIs — scoped by endpoint, method, rate, and data class — and proves what each agent did. You can keep your existing IdP; agents inherit the same trust model through OAuth 2.1, OIDC, and MCP.
No. The agent receives a short-lived, scoped credential rather than your real API keys. Secrets stay on the control plane, are never exposed to the model, and can be rotated or revoked at any time.
Through two standard protocol surfaces: an Agent2Agent (A2A) management server, discovered via the Agent Card, and per-collection MCP servers that turn your requests into tools. Both authenticate with OAuth 2.0 client credentials that a human provisions in the app — the full walkthrough is in the agent docs.
Fullmakt is pay-as-you-go — there's no flat plan and no minimum. You're billed at the end of each month only for what you use: per request sent through the broker, per AI-generated collection, and per endpoint generated, at our infrastructure cost plus a small margin. Every feature is included. Try it all first in the no-card sandbox; when you're ready, add a card. Remove the card or leave anytime and any remaining usage is settled on a final invoice.
Stop giving the agent the key. Store the real credential in Fullmakt and give the agent an OAuth client instead: it receives a one-hour scoped token, and Fullmakt injects the real credential server-side when it forwards the call. The key never appears in the agent's environment, its prompt, or its logs.
Every call an agent makes through Fullmakt is recorded — which agent, which endpoint, what the policy decided, and whether a human approved it — including the calls that were denied. The log is append-only and sensitive fields are masked, so it can be handed to an auditor as is.
Neither on its own — it is the access-control layer between them. A secrets manager stores credentials; a non-human identity tool says which machine is which. Fullmakt uses both to decide, per call, whether this agent may make this request right now — and it can read credentials straight out of the vault you already run.
Yes. Each collection can be exposed as an MCP server, so any MCP client — Claude Desktop, claude.ai connectors, or your own — sees your saved requests as tools and connects over OAuth. Agents that speak A2A instead use the Agent Card and the management surface.
Yes — the workbench is a full API client for HTTP, GraphQL, WebSocket, gRPC and MCP, and it imports Postman collections. The broker is what you turn on when an agent, rather than you, starts making the calls.
Have a question or feedback? Send us a message and we'll get back to you.