Skip to content

Let your agents use secrets without ever seeing them.

ASM is a self-hosted secrets manager for Claude Code and other MCP clients. An agent asks for a secret with a project and a reason, you approve it, and the value goes straight into the process the agent starts. It never shows up in the conversation.

Runs in one Docker container on your own server. Open source under the Apache License 2.0.

Claude Code asks ASM for NPM_TOKEN with the project my-app and the reason 'Publish v1.4.0 to npm'. The request is approved on a phone with Face ID, then asm run injects the token into npm publish. The value never appears in the conversation.
Demo: an agent requests a secret, you approve it on your phone, and the CLI hands it to the process.

How it works

Three steps between an agent needing a token and a command that has it. You decide at step two.

  1. 1

    The agent asks

    Claude Code finds the secret through MCP and requests it with the project it is working on and a concrete reason.

    ● asm · request_secret (MCP)
      secret  NPM_TOKEN
      project my-app
      reason  "Publish v1.4.0 to npm"
      Waiting for approval…
  2. 2

    You approve

    The request shows up live in the browser. Press A or D, or answer the push notification on your phone.

    Accessreceived 11:00 AM

    claude@macbook wants NPM_TOKEN for my-app.

    Reason: Publish v1.4.0 to npm

  3. 3

    The CLI hands it over

    The agent only gets a one-time grant. asm run redeems it and sets the value as an environment variable for that one process.

    $ asm run --grant grt_7Kq2X -- npm publish
    npm notice Publishing to registry.npmjs.org
    + my-app@1.4.0
    ✓ NPM_TOKEN was set for npm only

Built around the approval

A web UI for requests, secrets and the audit log, an MCP endpoint for agents, and a CLI that runs your commands.

Every request waits for you.

Requests arrive live with the device, the project and the reason. Changes an agent proposes to a secret wait the same way. Anything you leave unanswered expires after five minutes.

Working through a long task? Auto mode approves for a limited time and only for the vaults you pick. A standing approval trusts a single device for days. Every automatic approval lands in the audit log.

The ASM requests page with two open requests: claude@build-01 wants to change CLOUDFLARE_API_TOKEN, claude@macbook wants STRIPE_SECRET_KEY for shop-api. Each has Deny and Approve buttons and an expiry timer.

Vaults for you and your team.

Everyone has a personal vault, and you can share vaults with other people on the same server. Every change creates a new version with who made it, for which project and why; older values stay viewable and can be restored.

Agents can create secrets and archive them with a reason. They can't delete anything, and changing a value needs your approval.

The secrets list of the shared vault Shop with DATABASE_URL, STRIPE_SECRET_KEY, REDIS_URL and others; the details panel of DATABASE_URL is open with the masked value.

An audit log you can check.

Every request, decision and change is recorded with device, project and reason. Each entry contains the hash of the previous one, so edited or deleted entries show up as a broken chain.

Filter by device, secret, project or event, and export the log as CSV.

The ASM audit log showing 'Chain intact' and an expanded request.approved entry with project, reason, device, entry hash and previous hash.
  • One-time grants

    An approval gives the agent a grant ID, not the value. Grants are valid for 15 minutes and, by default, for a single use.

  • Auto mode

    Approves automatically for a limited time (30 minutes by default) and only for the vaults you select.

  • Standing approval

    Trust one device for days, for all secrets, selected vaults or a list. When it expires, ASM asks again, or asks once and extends it. An agent can propose one too; you narrow it and confirm it yourself.

  • Approve from your phone

    A push notification opens a small approval page; approve with Face ID. Only that page needs to be reachable from outside.

  • Passkeys

    Sign in with a passkey or a password. Revealing a value or pairing a device needs a sign-in from the last 10 minutes.

  • Devices you control

    Every Claude Code or CLI pairing is a device you approve in the browser and can revoke at any time. Link a CLI to its Claude Code device, and only that CLI can redeem the grants Claude receives.

  • Several users

    Admins add accounts and can deactivate them, which signs the person out and revokes all of their devices.

  • Self-hosted with Docker

    One container, one SQLite database, your master key. No external service involved.

  • English and German

    The web UI follows your browser language. Switch it any time under Settings.

The value stays out of the conversation

With the default CLI delivery, the model never sees the secret. That keeps it out of the transcript, out of logs and away from the model provider.

What the agent receives
status     approved
grant_id   grt_01M7MFA73EMVYEDYZ7XRJ2RN0G
run        asm run --grant <id> -- <command>
What the process receives
NPM_TOKEN=npm_••••••••••••••••••••••••
# set for this one process only,
# never written to the transcript
Plaintext is opt-in, per secret
An agent can only get a value in plain text if you allow it on that secret. It is off by default, and agents can never change the setting.
Every request needs a decision
Reads and changes wait for you and expire after five minutes. Auto mode and standing approvals are the only exceptions, both limited in time and in scope.
Archive instead of delete
Agents can archive a secret with a reason, never delete or restore it. Archived secrets can't be read or changed until a person restores them.
Encrypted per secret
AES-256-GCM with a separate key for each secret, encrypted with your master key. Keep the master key apart from the data.
Revocable devices
Refresh tokens rotate, and reusing an old one locks the device.
Tamper-evident audit log
A hash chain detects modified or deleted entries. Export it regularly to also catch entries cut off at the end.

The details are in the security model in the README.

Quick start

You need a server with Docker and, for passkeys and push, an HTTPS domain behind a reverse proxy.

  1. 1Build the image and create the master key

    Back the key up separately from the data. Without it, all secrets are lost.

    git clone https://github.com/eltomato89/asm.git && cd asm
    docker build -t asm .
    
    mkdir -p secrets && chmod 700 secrets
    docker run --rm asm node apps/server/dist/scripts/generate-key.js \
      > secrets/asm_master_key
    chmod 600 secrets/asm_master_key
    sudo chown 1000:1000 secrets/asm_master_key
  2. 2Start it and create the admin account

    Set ASM_PUBLIC_URL in docker-compose.example.yml to the address you'll reach ASM under, then start it.

    Open the address, enter the setup token from the log and create the first account. Add a passkey and more users under Settings.

    docker compose -f docker-compose.example.yml up -d
    docker compose -f docker-compose.example.yml logs asm \
      | grep "ASM setup token"
  3. 3Connect Claude Code

    On first use, Claude Code opens the browser: sign in, check the pairing and approve it. Claude can then search, request, create, change and archive secrets.

    claude mcp add --transport http asm https://asm.example.com/mcp
  4. 4Install the CLI

    Straight from your own server, hash-verified, no sudo. Needs Node.js 22.12 or later.

    curl -fsSL https://asm.example.com/cli/install.sh | sh
    asm login https://asm.example.com

Before you rely on it

ASM v1 runs as a single instance, project names are shared by all users of a server, and the audit hash chain can't tell when the newest entries were cut off. Read the limitations, and see the README for reverse proxy setup, push approval and backups.