01
AI agent
Receives status, key names, and the next action — not plaintext or child-process output.
Built for AI agents
Secrets your AI can use. Values it can't see.
Ask once in chat. manage-env generates what it can, collects provider keys through your local terminal, syncs approved CI targets, and runs the app — returning only status and key names.
no values Open source
AI chat · local MCP
You
Set up this repo's env. Generate what you can, ask me only for provider keys, then run the app.
AI agent
I'll handle it with local manage-env tools. Values stay out of chat and tool results.
setup_project
4 generated locally · 2 config defaults ready
request_value_from_user
Local prompt · STRIPE_SECRET_KEY, DATABASE_URL
sync_to_target
GitHub Actions synced · key names only
run_with_secrets
App exited 0 · 1.8s · no process output returned
The agent sees statuses, key names, and next steps — never a secret value.
What goes where
Work moves forward. Values stay on trusted paths.
02
Your machine
Decrypts locally and injects only when a process or an approved sync target needs the value.
03
manage-env server
Stores ciphertext, wrapped keys, and value-free metadata that it cannot turn into plaintext.
Why it's safe to hand to an agent
MCP tools inject secrets into a child process and return only exit codes and key names. There is no get-secret tool — by design, a value can't travel back through a tool result.
no get_secret tool
2
Plaintext stays on trusted client paths
Values are decrypted client-side and sent only to the local process or provider you explicitly approved. Our server stores encrypted data it cannot decrypt.
3
The server only sees encrypted data
Diff, history, and sync status run on keyed fingerprints, not plaintext. We compare versions without ever seeing a value.
Don't take our word for it. See encryption run in your own browser.
XChaCha20-Poly1305 · Argon2id · end-to-end encrypted
One-time setup
Connect once. Ask your agent from then on.
Use the terminal for the first trust setup and local prompts. After that, the main interface is the AI request in your editor.
01
Agent-led repo setup
Your AI can discover required keys, initialize the project, and set up what is safe to create automatically.
02
No-reveal MCP execution
run_with_secrets injects values into the process and returns only status, duration, and key names.
03
Trusted CI sync
Sync to GitHub Actions, Vercel, or Cloudflare only after committed target declarations and human trust approval.
Claude CodeCursorGitHub ActionsVercelCloudflare
Install from GitHub$menv signup
$menv connect
$menv init me/my-app
Tell your AI
Set up this repo with manage-env. Generate safe random/config values, ask me locally for provider-issued secrets, sync trusted CI targets, then run the app without returning secret values.
Let AI handle env work without seeing env values.
Create an account, connect your first repo, and give your agent one clear instruction.