AI agent secret management, explained plainly.
Practical guides for scoping API keys, provisioning coding agents, replacing risky env-file habits, and keeping credential resolution auditable.
Fix an incident
Pasted an API key into chatIf you pasted an API key into ChatGPT, Claude, Cursor, Codex, or another AI chat, treat it as exposed, revoke or rotate it, and move future agent access to scoped runtime resolution.Pasted API key into ChatGPTIf you accidentally pasted an API key into ChatGPT, treat it as exposed, revoke or rotate it, check usage, and avoid pasting the replacement key back into chat.Shared API key with ClaudeIf you shared an API key with Claude or Claude Code, rotate the exposed key, check where it was used, and move future Claude workflows to scoped runtime access.Claude Code read my .envIf Claude Code read your .env file, treat every key inside it as exposed: rotate them at each provider, check usage, then move real keys out of readable files so the next session cannot repeat it.Surprise bill from a leaked keyA surprise bill usually means a leaked API key is being used right now. Kill the key first, then cap spend, contact the provider about the charges, find the leak path, and rebuild so keys stop living where they leak.API key exposure checklistA practical checklist for what to do after pasting an API key into ChatGPT, Claude, Cursor, Codex, or another AI tool.Rotate an OpenAI API keyIf an OpenAI API key was pasted into chat, shared with an AI agent, or exposed in local files, rotate it, update legitimate runtimes, and prevent repeat exposure.
Set up safe agent access
Give an agent an API key safelyGive an AI agent API-key access safely by creating a named agent identity, granting the credential it needs, resolving it at runtime, and auditing each use.Give Claude Code an API key safelyGive Claude Code API-key access safely with a named agent, one project secret, runtime resolution, and raw keys kept out of chat.Give Cursor an API key safelyHow to give Cursor an API key safely: handle credentials for a Cursor agent with a project-scoped identity, a direct secret grant, runtime resolution, and audit history.Give Codex an API key safelyGive Codex API-key access safely with a scoped agent identity, the needed project secret, runtime resolution, and raw provider keys kept out of chat and local files.Share API keys safelyThe safer way to share API-key access with AI tools is scoped runtime resolution: store the key once, grant the named agent or workflow the access it needs, and audit each use.Are .env files safe for agents?.env files are useful for local configuration, but AI agents need tighter handling for real API keys because agents can inspect files, run commands, and repeat workflows. The safer alternative is scoped runtime resolution from a central credential layer.
Tool guides
Claude Code secrets managementSecrets management for Claude Code keeps raw API keys out of chat by using scoped agent credentials, project grants, runtime resolution, and audit history.Credentials for Cursor agentCredentials for a Cursor agent should be scoped and resolved at runtime, not pasted into prompts or broad .env files. How to give Cursor an API key safely, per project.Codex agent secretsCodex agents can use sensitive APIs more safely with scoped secret resolution instead of raw keys pasted into chat, shell history, or local files.MCP credential managementManage MCP credentials and secrets without copying API keys into config files, prompts, or broad local environments. ScopeHold gives MCP tools scoped runtime access with audit trails and revocation.Best MCP credential managementCompare raw .env files, MCP config files, password-manager CLIs, traditional secret managers, custom MCP proxies, and ScopeHold for MCP credential management.
Concepts and comparisons
Best AI agent credential managementCompare pasted keys, .env files, password-manager CLIs, traditional secret managers, short-lived tokens, and scoped runtime resolution for AI agent credential management.API key management for AI agentsAPI key management for AI agents means storing keys centrally, granting them narrowly, resolving them at runtime, and auditing each use by agent identity.Agentic AI secret managementAgentic AI secret management gives autonomous or semi-autonomous agents scoped, auditable access to credentials without turning every task into a manual key-sharing workflow.Non-human identityNon-human identity for AI agents means giving each runtime a distinct identity, access scope, credential path, and audit trail instead of borrowing a human user's secrets.Scoped credentialsScoped credentials limit what an AI agent can resolve by workspace, project, provider, secret, and direct grant, reducing blast radius without slowing every task.How ScopeHold protects agent secretsScopeHold keeps AI agent secrets out of chat, logs, and local files by combining scoped agent identities, one-time provisioning, encrypted secret custody, MFA before human reveal, audit logs, rate limits, and plan-limit enforcement.
For products that hold customer credentials
MCP server credential custodyThe MCP specification defines client-to-server authorisation, not downstream credential custody. Here is where MCP server credentials actually live, and how to manage them.How AI products collect credentialsA neutral survey of how 17 AI products take and hold customer credentials. Six patterns from public documentation, and the finding that stands out: of 17 products, one documents custody after intake.Answering the credential questionnaireThe six questions a security reviewer asks when your product holds customer credentials: where stored, what can read it, scoping, rotation, offboarding, evidence. Honest answer patterns at three maturity levels.