1Password MCP Server

1Password MCP Server

Give your AI assistant a locked door to 1Password — not a pile of passwords in the chat.

@takescake/1password-mcp is a community Model Context Protocol (MCP) server. It lets Claude, Cursor, VS Code Copilot, OpenAI Codex, Gemini, and other MCP clients manage vaults and credentials through a 1Password Service Account.

Built on the MCP TypeScript SDK v2 with protocol negotiation for 2026-07-28 (and legacy clients). Secrets stay in 1Password; agents prefer op:// references and op_run so plaintext never has to land in the model transcript.

Not an official 1Password product. Community-built, Apache 2.0 licensed.

---

Who this is for

---

Why teams pick this server

• Security-first defaults — password_read returns metadata, and item_get hides secret-bearing fields (passwords, SSH keys, OTP seeds, card numbers), unless you opt in with reveal: true.
• op_run (the MCP equivalent of op run) — inject op://vault/item/field into a local command’s environment; resolved secrets are redacted from the returned stdout/stderr on a best-effort basis (including common encodings).
• Full vault toolkit — list, search, get, edit, create logins & notes, rotate passwords, archive, or delete.
• Guided prompts — password generation, credential rotation, vault audit, and secret-reference helpers.
• Browsable resources — vault and item catalogs over onepassword://… URIs (no secrets in resource payloads).
• Modern MCP — stdio transport, Zod 4 schemas, MCP 2026-07-28 negotiation with legacy client compatibility.

---

What you get

Tools (15)

Grouped the way agents and humans actually use them.

Discover

Read (safe by default)

Create & update

Use secrets without revealing them

Soft-delete & destroy

Prompts (4)

Resources (3)

Upgrading from 4.x: resource URIs used to start with 1password://, which MCP clients could never read (a URI scheme can't start with a digit). Replace any hard-coded 1password:// URIs with onepassword://.

---

Before you start

You need two things:

1. Node.js 20 or newer
2. A 1Password Service Account with access to the vault(s) you want the AI to use

Create a service account (plain English)

1. Sign in to your 1Password account on the web.
2. Open Developer → Service Accounts (or follow 1Password’s guide).
3. Create a service account and grant it only the vaults you want automation to touch (for example an Automation or CI vault — not your personal banking vault).
4. Copy the token once. Treat it like a master key.

---

Quick start

Claude Desktop / Cursor / VS Code / most IDEs

Add this to your MCP config (exact file depends on the app):

{
  "mcpServers": {
    "1password": {
      "command": "npx",
      "args": ["-y", "@takescake/1password-mcp"],
      "env": {
        "OP_SERVICE_ACCOUNT_TOKEN": "YOUR_SERVICE_ACCOUNT_TOKEN"
      }
    }
  }
}

Restart the app, then try: “List my 1Password vaults.”

macOS Keychain (no token in the config file)

Store the token in Keychain, then point the server at it:

{
  "mcpServers": {
    "1password": {
      "command": "npx",
      "args": ["-y", "@takescake/1password-mcp"],
      "env": {
        "OP_KEYCHAIN_SERVICE": "op-service-account-claude-automation",
        "OP_KEYCHAIN_ACCOUNT": "your-macos-username"
      }
    }
  }
}

Token resolution order: CLI (--service-account-token / --token) → OP_SERVICE_ACCOUNT_TOKEN → macOS Keychain. OP_KEYCHAIN_ACCOUNT is optional when the service name alone is unique. Avoid the CLI flags: command-line arguments are visible to other local processes, and the server logs a warning at startup if you use them.

OpenAI Codex (TOML)

Option A — token in config:

[mcp_servers."1password"]
command = "npx"
args = ["-y", "@takescake/1password-mcp"]

[mcp_servers."1password".env]
OP_SERVICE_ACCOUNT_TOKEN = "YOUR_SERVICE_ACCOUNT_TOKEN"

Option B (recommended) — config only names the env var:

[mcp_servers."1password"]
command = "npx"
args = ["-y", "@takescake/1password-mcp"]
env_vars = ["OP_SERVICE_ACCOUNT_TOKEN"]

Set OP_SERVICE_ACCOUNT_TOKEN in your shell or CI. Note: codex mcp add ... --env OP_SERVICE_ACCOUNT_TOKEN=... writes the secret into Codex config; prefer env_vars when you can.

On macOS you can omit the token env and use OP_KEYCHAIN_SERVICE (+ optional OP_KEYCHAIN_ACCOUNT) instead.

Optional: restrict the server to certain vaults

By default the server can use any vault the service account can see. To allow-list vaults:

{
  "env": {
    "OP_SERVICE_ACCOUNT_TOKEN": "YOUR_SERVICE_ACCOUNT_TOKEN",
    "OP_MCP_ALLOWED_VAULTS": "Automation, CI"
  }
}

Names or IDs work (case-insensitive). Same setting via --allowed-vaults. The allow-list applies server-wide, to every tool and resource that touches a vault, not only op_run / op_check_ref:

• vault_list and onepassword://vaults show only allowed vaults.
• Tools that take a vaultId, and the onepassword://vaults/{vaultId}/items resource, must be given the vault’s ID (not its name); vaults outside the list are rejected.
• op:// references are checked both as written and by the vault they actually resolve to.
• When an allow-list is set, each guarded call makes one extra vaults.list read (mind service-account rate limits), and the server fails closed if vaults can’t be listed.

Upgrading from 4.0.3 or earlier? If you set this expecting it to affect only op_run / op_check_ref, it now restricts everything.

This is defense in depth: scope the service account’s own vault access in 1Password first.

---

For agents: how to handle secrets

Follow this order every time:

1. Discover with vault_list → item_lookup / item_list (metadata only).
2. Confirm a reference with op_check_ref — never reveal just to see if a path exists.
3. Use a secret in a command or API call with op_run and op://vault/item/field in env.
4. Reveal with password_read / item_get + reveal: true only when the human explicitly needs the value in chat.
5. Rotate with password_generate → password_update (keep returnSecret: false unless asked).
6. Prefer item_archive over item_delete unless permanent removal is required.

op_run sketch

{
  "argv": ["curl", "-sS", "https://api.example.com/health"],
  "env": {
    "API_TOKEN": "op://Automation/Example API/credential"
  },
  "timeout_ms": 60000
}

Prefer argv over a shell command string when you can — fewer quoting surprises.

---

Configuration reference

Environment variables

CLI flags

--service-account-token <token>   1Password service account token (avoid: visible to other local processes)
--token <token>                   Alias for --service-account-token
--log-level <level>               error | warn | info | debug (default: info)
--integration-name <name>         Custom integration name for the 1Password SDK
--integration-version <version>   Custom integration version
--allowed-vaults <list>           Comma-separated vault allow-list (names or IDs), applied server-wide

---

Security & privacy

Read this before pointing the server at a vault you care about.

• LLM privacy — Anything revealed to the model may be sent to your AI provider and retained under their policies.
• MCP is not end-to-end encrypted for secrets in flight — Values are plaintext inside the MCP workflow and toward the model. They are encrypted at rest in 1Password once stored.
• Best fit — Automation credentials: CI tokens, bot accounts, disposable env secrets.
• Avoid — Banking, primary personal logins, recovery codes, or anything you cannot afford to expose to a model provider.
• Token = master key — Scope the service account tightly; rotate immediately if leaked; never commit tokens or MCP configs with secrets.
• Prefer the env var or Keychain for the token — --token / --service-account-token puts it in the process arguments, which other local processes can read (the server warns at startup). A same-user process can generally read the server’s environment too (for example /proc/<pid>/environ on Linux), so on macOS Keychain is the strongest option: it keeps the token out of config files and the process environment.
• Prefer references — op://… + op_run beat pasting passwords into prompts or files.
• op_run runs arbitrary commands as your user — Keep your MCP client’s approval prompts on for it; don’t auto-approve it.
• Redaction is best effort — op_run masks resolved secrets in returned output, including common encodings (base64, JSON- and URL-encoded forms, multi-line values). That protects against accidental disclosure. It is not a sandbox: a command can deliberately transform or transmit a secret it was given.
• Notes are not concealed — item_get returns notes as-is (like op item get). Don’t keep secrets in the notes of vaults an agent can read.
• Least privilege — Dedicated automation vaults beat sharing your whole account. OP_MCP_ALLOWED_VAULTS is a second fence, not a substitute.
• Reporting vulnerabilities — Open a public issue; see SECURITY.md.

---

Protocol & compatibility

---

Development

git clone https://github.com/CakeRepository/1Password-MCP.git
cd 1Password-MCP
npm ci
npm run build
npm test
npm run lint

Watch mode: npm run dev.

Project layout

src/
  index.ts                 # Entrypoint — MCP stdio + protocol negotiation
  server.ts                # buildServer() — registers tools, prompts, resources
  config.ts                # CLI / env / Keychain / allow-list
  client.ts                # 1Password SDK client
  logger.ts                # Structured logs on stderr (stdout is protocol)
  secret-ref.ts            # op:// parsing & reference checks
  vault-access.ts          # Server-wide vault allow-list enforcement
  redaction.ts             # op_run output redaction
  utils.ts                 # Result helpers, password generation
  tools/                   # All 15 MCP tools
  prompts/                 # Interactive workflow prompts
  resources/               # onepassword:// resources
tests/

See CONTRIBUTING.md. Maintainers / agents: agents.md.

---

Changelog

See CHANGELOG.md for version history, including the 5.0.0 release (resource URIs moved from 1password:// to onepassword://, plus security hardening; read its Changed notes before upgrading), the 4.0.0 MCP v2 / 2026-07-28 migration, and the 3.0.0 op_run / reveal-opt-in security changes.

---

License

Apache License 2.0