Keycard logo

Keycard

Category:

Runtime Authorization for AI Agents and MCP Servers - Keycard

What is Keycard?

Keycard is runtime authorization infrastructure that gives AI agents their own identity and issues short-lived, task-scoped credentials on the fly. Instead of handing agents risky, permanent API keys, it checks permissions in real time, helping teams safely ship autonomous agents into production with clear audit trails.

Features

Overview

Keycard is runtime authorization infrastructure for AI agents, MCP servers, APIs, and internal developer tools. Built by Keycard Labs, Inc., it gives each agent its own application identity and checks every credential request against policy.

The platform targets privilege accumulation and unclear attribution in autonomous systems. Instead of inherited human credentials or static grants, Keycard issues short-lived tokens tied to the agent, its runtime, the task, and the human user.

Teams can enforce policy in three places: SDKs inside application code, an optional Agent Gateway for MCP servers, or an infrastructure sidecar proxy. Organizations federate their existing identity providers and write access rules in Cedar or Terraform.

Keycard acts before a credential exists, not after traffic is inspected. Agents talk directly to target resources unless the optional Gateway is used.

Pricing

Keycard bills by transactions, counted whenever it issues, validates, or exchanges a credential. The Starter plan costs $0 per month with a hard cap of 5,000 transactions, unlimited users, agents, and apps, and 7-day telemetry retention, but no SSO. The Team plan costs $500 per month for 100,000 transactions, plus $1 per 1,000 extra, and adds SSO, Zone Policy, 90-day retention, and audit log export. Enterprise pricing is custom with an annual commitment and adds SCIM, BYOC or on-prem deployment, customer-managed KMS, and a 99.95% uptime SLA.

* Disclaimer: Please note that pricing information may not be up to date. For the most accurate and current pricing details, refer to the official website.

Key Features

  • ✓

    Evaluates agent, runtime, task, and delegated user together

  • ✓

    Issues short-lived tokens only after policy approval

  • ✓

    Policy as code with Cedar, RBAC, ABAC, and ReBAC

  • ✓

    Agent Gateway consolidates MCP servers with tool-level allow or deny

  • ✓

    OCSF audit export to Amazon S3, Splunk, and Panther

Use Cases

01

User-Level Limits on Shared Agents

One internal assistant serves several departments. Keycard checks the user’s directory group on each request, so the agent reaches PII stores only for authorized users.

02

Governing Coding Agent Tool Access

Security teams connect Claude Code or Cursor to repositories, trackers, and deployment tools. A Gateway or the Keycard CLI limits which MCP servers and tools each session can call.

03

Secretless Production Agent Deployments

Platform engineers run agents on Render, Fly.io, AWS, or Google Cloud without embedded API keys. Keycard exchanges the host’s OIDC workload token for short-lived resource credentials.

04

Centralizing Auth Across MCP Servers

Organizations running GitHub, Linear, Sentry, and PagerDuty MCP servers route calls through one Gateway. Developers avoid rewriting auth middleware inside each server.

05

Investigating and Halting Agent Activity

Security operations teams track which agent accessed which resource for whom in their SIEM. After odd behavior or a role change, one revocation stops further token issuance.

Strengths & Weaknesses

Strengths

+

It removes long-lived static credentials by minting task-scoped tokens only after policy checks pass.

+

It federates Okta, Microsoft Entra ID, Google, Auth0, and Ping Identity instead of requiring an identity migration.

+

SDK, Gateway, and sidecar options let teams avoid routing all runtime traffic through Keycard.

+

Cedar and Terraform let teams version, test, and roll back access policies.

+

The company holds SOC 2 Type II certification and offers BYOC, on-premises, and customer-managed KMS options.

Weaknesses

–

The free plan has a hard cap of 5,000 transactions, which may limit testing of high-frequency agent loops.

–

No self-serve tier sits between the free plan and the $500 per month Team plan.

–

Access requires a work-email invite request or a booked demo, not open self-registration.

–

Official SDKs cover only Python, TypeScript, Go, and Ruby, so other languages need the Gateway or a sidecar.

Who Is This For?

●

AI and agent developers: Engineers building assistants or MCP servers who want to reach third-party APIs without writing custom OAuth token plumbing.

●

Platform and infrastructure engineers: Teams deploying agent workloads who want secretless workload authentication and Terraform-managed policies.

●

Enterprise security and IAM architects: Security leads enforcing least privilege and keeping audit trails across human and non-human identities.

●

AI operations and engineering managers: Leaders scaling Cursor or Claude Code across large organizations who need central visibility and instant revocation.

Frequently Asked Questions

How does Keycard differ from an API gateway or proxy?

Most tools inspect traffic after a credential exists. Keycard evaluates policy first, so a denied request produces no token at all.

Does Keycard replace our existing identity provider?

No. It federates your current provider and adds runtime agent policy and dynamic token issuance on top.

What happens when a user lacks access to a requested resource?

Cedar policies such as agents-inherit-user-scope evaluate agent and user together. Keycard denies the request, issues no token, and logs the denial.

Can it govern local coding agents like Claude Code or Cursor?

Yes. Connect them through an Agent Gateway or launch them with the Keycard CLI to apply tool-level access policies.

How do deployed agents authenticate without static secrets?

Keycard verifies native OIDC tokens from AWS, Google Cloud, Vercel, GitHub Actions, Render, and Fly.io. It then exchanges them for short-lived credentials.

Can agents keep working when the user is offline?

Yes. Agents can continue tasks for an absent user under previously authorized grants. Revoking a grant blocks further credential issuance.

Which authorization models does Keycard support?

It supports role-based, attribute-based, and relationship-based access control. Each model is available on every plan, including the free tier.

Which protocols and standards does it implement?

Keycard supports OAuth 2.1, Token Exchange, SPIFFE, WIMSE, MCP, and A2A delegation. It also works with standard APIs and CLIs.

What are the steps to get started?

Request access, create an organization and Zone, and federate your identity provider. Then register agents as Applications, write Cedar policies, and add an enforcement point.

Keycard integrates with Okta, Microsoft Entra ID, Google, Auth0, Ping Identity, OIDC, Active Directory, LDAP, and SCIM. Its catalog covers GitHub, Slack, Linear, Notion, Gmail, PagerDuty, Sentry, Neon, Google Calendar, and Snowflake. It supports Claude Code, Cursor, OpenAI Agents, and the Claude Agent SDK. SDKs exist for Python, TypeScript, Go, and Ruby, plus a CLI and a Terraform provider. Audit events export to Amazon S3, Splunk, and Panther in OCSF format.

Integrations