Secrets management
for AI agents.
An agent needs credentials to do anything useful, and anything an agent can read it can be talked into leaking. seekrit stores secrets end-to-end encrypted and delivers them as spawn-time env vars, MCP-managed names, or placeholders swapped at the network boundary — so the model never holds the raw key.
Values are decrypted on your machine at launch and exist only in the child process's environment — no .env file for the agent to read, commit, or quote back. Replacing .env →
How do I give an AI agent an API key securely?
Three levels, ordered by how little the agent gets to see. Most teams start at the first and move down as the agent gets less trusted or longer-lived — the full walkthrough has the decision table.
Spawn-time injection
seekrit run -- <your agent>The agent sees: Env vars in the child process, nothing on disk
Good for: Local dev, CI, MCP servers, anything reading process.env
Guide →MCP + agent plugin
npx plugins add seekritdev/agent-pluginThe agent sees: Secret names and metadata; skills for using a value without printing it
Good for: Coding agents — Claude Code, Cursor, anything speaking MCP
Guide →Placeholder brokering
seekrit proxy run --preset openaiThe agent sees: {{seekrit:NAME}} — a string that names a key without being one
Good for: Untrusted, long-running, or sandboxed agents
Guide →How do I keep API keys out of MCP config files?
.mcp.json and claude_desktop_config.json both take an env block, and nearly every MCP server README says to paste your key into it — which makes the config a plaintext secrets file that gets committed, synced, and read by every agent with workspace access. Wrap the server instead: seekrit run -- npx <mcp-server> keeps the config value-free and injects the credential only into the server's own process at launch.
Which agent stacks does it plug into?
Coding agents
One plugin install gives Claude Code, Cursor, and anything else speaking MCP both servers plus the skills for using a secret without printing it.
Agent guides →Frameworks
LangGraph, CrewAI, Mastra, Pydantic AI, the AI SDK, LlamaIndex, and both Agents SDKs — three commands to get keys in, then tighten from there.
Framework guides →Sandboxes
E2B, Modal, Daytona, Vercel Sandbox, Cloudflare Sandbox and Computer — pass secrets in at spawn, or keep the credential outside the sandbox entirely.
Sandbox guides →Why does zero-knowledge matter for agent secrets?
Keeping a credential away from your agent is worth little if the service storing it can read it anyway. seekrit is zero-knowledge: secrets are encrypted in your browser or CLI before they leave the machine, the server holds ciphertext it cannot decrypt, and the egress policy that decides where a credential may go is signed on your side — bytes the service can store but not rewrite.
Start with one command
Free for one developer, no card. Put a key in, run your agent through seekrit run, and delete the .env it replaced.