mcp-1password

Status: Public Beta

This package is under active development (v0.x). The API and CLI flags may change between minor versions. The underlying @1password/sdk dependency is also a beta release. Pin an exact version (mcp-1password@x.y.z) in production-like environments.

A Model Context Protocol server that exposes 1Password to AI agents with opaque-by-default secret handling. Secrets are never revealed unless you explicitly opt in.

Features

• Read and search vaults, items, and environments with secrets redacted by default.

• Create tracked item requests, then update, archive, and delete items and vaults when write/destructive capabilities are enabled.

• Manage group permissions on vaults when permission mutation is enabled.

• Reveal plaintext secrets only on explicit request with a per-call acknowledgement.

• Generate plaintext passwords only with a reason and explicit acknowledgement.

• Run pre-approved scripts with injected 1Password-backed environment values.

• Optionally run unrestricted local shell commands under explicitly approved workspace roots, after local browser confirmation.

• Use stdio by default, or a local/single-user HTTP transport protected by a bearer token.

• Write a JSONL audit log for sensitive actions at ~/.onepassword-mcp/audit.jsonl.

Related MCP server: op-mcp

Requirements

• Node.js >= 20.10

• 1Password desktop app for --auth-mode=desktop; this requires the 1Password beta channel with SDK integration enabled.

• 1Password Connect for --auth-mode=connect; the POC only accepts a localhost Connect host.

• 1Password CLI (op) only when scripts need CLI authentication or you configure script runner CLI auth; Connect-backed envSecretRefs do not require op.

Enable Desktop Integration

Desktop auth requires the 1Password beta channel and SDK integration:

1. In 1Password, switch to the beta channel: Settings -> Updates -> Beta channel.

2. Enable SDK integration: Settings -> Developer -> Connect with 1Password SDKs.

Installation

# Public beta install
npm install -g mcp-1password@beta

# Connect-only binary after global install
OP_CONNECT_TOKEN="<connect-token>" \
mcp-1password-connect --connect-host=http://127.0.0.1:8090

# Run on demand without a global install
npx -y mcp-1password@beta --auth-mode=desktop --account="My Account"

During beta, prefer mcp-1password@beta or an exact version instead of relying on the default npm tag.

Quick Start

Connect-Only Binary

Use mcp-1password-connect when you want a path where Connect is the only possible 1Password backend. This binary forces --auth-mode=connect, rejects Desktop/service-account auth, rejects non-Connect op CLI auth modes, and never instantiates the Desktop SDK service.

# From the project root, create/update the trust file and manifest first.
mcp-1password-connect trust-workspace

OP_CONNECT_TOKEN="<connect-token>" \
mcp-1password-connect \
  --connect-host=http://127.0.0.1:8090 \
  --enable-script-runner=true \
  --script-runner-allowlist-manifest="$HOME/.onepassword-mcp/workspace-trust.json"

For each new project that should be trusted for Connect workspace commands:

mcp-1password-connect trust-workspace

Then call workspace_trust_reload in the running MCP session.

Claude Desktop (stdio Transport)

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "1password": {
      "command": "npx",
      "args": [
        "-y", "mcp-1password@beta",
        "--auth-mode=desktop",
        "--account=1Password account name or UUID"
      ]
    }
  }
}

Service Account (CI / Headless)

{
  "mcpServers": {
    "1password": {
      "command": "npx",
      "args": ["-y", "mcp-1password@beta", "--auth-mode=service-account"],
      "env": {
        "OP_SERVICE_ACCOUNT_TOKEN": "<service-account-token>"
      }
    }
  }
}

Local 1Password Connect POC

Run the local Connect containers from docker-compose.connect.example.yml, then start the MCP without Desktop auth:

OP_CONNECT_TOKEN="<connect-token>" \
mcp-1password \
  --auth-mode=connect \
  --connect-host=http://127.0.0.1:8090

Connect mode supports vault/item reads, tracked placeholder creation through item_request_create, managed-item review through item_request_list, item update/delete, password_update, password_read, and secret_reveal. It does not expose generic password_create or item_create, vault mutation, group permissions, 1Password Environments, files, or item archive. See docs/connect-local-poc.md.

HTTP Transport (Remote Agents)

OP_MCP_HTTP_BEARER_TOKEN="$(openssl rand -base64 32)" \
mcp-1password \
  --auth-mode=desktop \
  --account="My Account" \
  --transport=http

HTTP security: The HTTP transport is designed for local/single-user use. The bearer token must be at least 16 characters, and --http-require-bearer=false is only allowed on localhost. If you bind the server to any interface other than 127.0.0.1, put it behind a reverse proxy with TLS termination (nginx, Caddy, Traefik). For multi-user or public deployments, add a real upstream authorization layer such as OIDC/OAuth with client identity, scopes, and expiry.

Configuration Reference

Every flag can also be set through an environment variable.

Script Runner

The script runner lets agents invoke pre-approved shell commands with 1Password-backed environment injection. In --auth-mode=connect, the MCP exposes workspace_trust_list, workspace_command_run, and workspace_trust_reload; envSecretRefs are resolved through Connect and injected directly into the child process without the op binary, OP_SESSION, or Desktop SDK auth. With Desktop or service-account auth, the MCP exposes the older op_script_list, op_script_run, and op_script_reload_allowlists tools, and the runner can also inject CLI authentication for commands that call op. By default, only startup-configured commandId entries can run. In Connect mode, workspace_trust_list resolves the requested workspaceRoot against startup-configured workspace trust entries and returns workspaceCommandResolution; when that resolution reports freeformCommands: true, workspace_command_run may accept a free-form command rooted in that workspace or its subdirectories. The startup command catalog and workspace trust files can be reloaded on demand with the reload tool.

For a new Connect-mode project, run this from the project root:

mcp-1password trust-workspace

This creates or updates .onepassword-mcp.json in the project, enables workspace commands for that project, and adds the file to ~/.onepassword-mcp/workspace-trust.json. Start the MCP with that manifest configured once:

mcp-1password \
  --auth-mode=connect \
  --enable-script-runner=true \
  --script-runner-allowlist-manifest="$HOME/.onepassword-mcp/workspace-trust.json"

If the MCP is already running with that manifest, call workspace_trust_reload after trusting a new project.

For local single-user sessions where preapproving every command is too expensive, start the server with:

mcp-1password \
  --auth-mode=desktop \
  --account="My Account" \
  --enable-unrestricted-script-runner=true \
  --op-cli-path=/absolute/path/to/op

With --enable-unrestricted-script-runner=true, op_script_run ignores --script-runner-root, --script-runner-allowlist, and --script-runner-allowlist-manifest. The first free-form op_script_run call returns an approvalUrl; open it locally, tick the checkbox, and type:

I_UNDERSTAND_THIS_ALLOWS_UNRESTRICTED_LOCAL_COMMAND_EXECUTION

By default, that approval is in memory only and applies once per MCP server process. If you tick Remember this approval for 24 hours on this machine, the server writes an encrypted local grant containing only the approval scope and expiration timestamp. After approval, op_script_run accepts command instead of commandId, runs it through a non-login /bin/sh -c shell in the requested workspace, and still supports envSecretRefs so secrets are injected into the child process without being returned to the model. When the remembered grant expires, the approval page is required again.

Command Catalog And Workspace Trust Format

Create a .onepassword-mcp.json file at the root of your project:

{
  "version": 1,
  "workspaceRoot": ".",
  "allowWorkspaceCommands": false,
  "commands": {
    "deploy-staging": {
      "description": "Deploy to staging",
      "command": "/usr/local/bin/deploy.sh",
      "args": ["--env", "staging"],
      "cwd": ".",
      "timeoutMs": 120000,
      "sensitiveOutput": false
    }
  }
}

• command must be an absolute path to an executable.

• The directory containing command is not automatically prepended to PATH. Use absolute paths in scripts, or configure --op-cli-path so the directory containing op can be injected.

• In --auth-mode=connect, allowWorkspaceCommands: true marks the matching workspaceRoot, workspaceRoots, or workspaceRootPrefixes as a trusted workspace command scope. workspace_trust_list exposes the decision as workspaceCommandResolution.freeformCommands; when it is true, workspace_command_run accepts a free-form command for that resolved workspace or its subdirectories. Secrets passed with envSecretRefs are resolved through Connect only, injected only into the child process, and are not returned to the model. This is an explicit trust decision for local worktrees you control.

• sensitiveOutput: true withholds stdout/stderr from the agent unless returnOutput=true is explicitly requested with reveal acknowledgement.

• workspace_command_run in Connect mode and op_script_run in Desktop/service-account mode accept an optional envSecretRefs object that maps environment variable names to op:// references. The server resolves those references, injects only the values into the child process environment, and never returns or audits the plaintext values.

• returnOutput=true does not require startup secret reveal for ordinary output. When envSecretRefs is provided or the command has sensitiveOutput: true, stdout/stderr/error messages are returned only with acknowledgePlaintext: "I_UNDERSTAND_THIS_RETURNS_SECRET_PLAINTEXT"; without that acknowledgement, execution is skipped with executionSkipped: true and outputState: "skipped_ack_missing". Returned stdout/stderr/error messages are redacted by exact secret value.

• After editing a startup-configured command catalog or workspace trust file, call workspace_trust_reload in Connect mode or op_script_reload_allowlists in Desktop/service-account mode with a reason. If the edited file is invalid, the reload fails and the previous in-memory configuration remains active.

Startup Manifest Format

Use --script-runner-allowlist-manifest=/absolute/path/to/workspace-trust.json when you want to add or remove workspace trust or command catalog files without restarting the MCP process. Manifest entries may be absolute paths or paths relative to the manifest file:

{
  "version": 1,
  "allowlists": [
    "/absolute/path/to/project-a/.onepassword-mcp.json",
    "../project-b/.onepassword-mcp.json"
  ]
}

After editing the manifest, call workspace_trust_reload in Connect mode or op_script_reload_allowlists in Desktop/service-account mode with a reason. Any new workspace roots are still checked against startup-configured --script-runner-root values when roots are provided.

Agent Routing Guidance

When an agent needs a secret only to run a local command, it should not call password_read with reveal=true or secret_reveal first. Prefer this flow:

1. In Connect mode, call workspace_trust_list for the current workspace; in Desktop/service-account mode, call op_script_list.

2. In Connect mode, inspect workspaceCommandResolution: if freeformCommands=true, call workspace_command_run with command; otherwise pick a listed commandId that performs the operation.

3. Call workspace_command_run in Connect mode or op_script_run in Desktop/service-account mode with envSecretRefs, mapping environment variable names to op:// references.

4. Leave returnOutput=false unless command output is required.

This keeps the plaintext secret out of the model transcript while still letting the command receive it.

Tracked Secret Requests

When an agent needs a new credential to exist in 1Password, prefer item_request_create over asking the user to paste the value. The tool creates a managed item with:

• credential fields initialized to the __FILL_ME__ placeholder;

• provenance fields for project, justification, and optional Linear ticket details;

• managed tags such as mcp-managed, awaiting-fill, and project:<slug>;

• op:// references that the agent can wire into configs or scripts.

The user fills the returned references in 1Password. Later, item_request_list reports which managed items are still awaiting fill, without returning the credential values.

Unrestricted Runner

The unrestricted runner is a separate, dangerous escape hatch for trusted local worktrees where allowlisting every command is too expensive. Enable it only for roots you are willing to approve for arbitrary command execution:

mcp-1password \
  --auth-mode=desktop \
  --account="My Account" \
  --enable-unrestricted-runner=true \
  --unrestricted-runner-root=/absolute/path/to/trusted/worktree

When an MCP client first calls op_unrestricted_run for a configured root, the tool returns authorizationRequired: true and a local approvalUrl. Open that URL on the same machine, tick the risk checkbox, and type:

I_UNDERSTAND_THIS_ALLOWS_UNRESTRICTED_LOCAL_COMMAND_EXECUTION

That approval is in memory only and expires after --unrestricted-runner-approval-ttl-ms. It is not written to config. The configured root is an approval scope, not an operating-system sandbox: approved commands run with your normal OS permissions and can still cd, read, write, or execute outside that path if the OS allows it.

op_unrestricted_run starts the command in the requested workspace root through /bin/sh -c on Unix-like systems, with a minimal inherited environment and no 1Password secret injection. The shell is intentionally non-login so local profile hooks such as RVM do not run before the requested command. Use op_script_run when a command needs 1Password values injected safely. As with script output, returnOutput=true without acknowledgePlaintext: "I_UNDERSTAND_THIS_RETURNS_SECRET_PLAINTEXT" skips execution and returns the required acknowledgement.

You can disable the browser approval page only with an explicit startup acknowledgement:

mcp-1password \
  --auth-mode=desktop \
  --account="My Account" \
  --enable-unrestricted-runner=true \
  --unrestricted-runner-root=/absolute/path/to/trusted/worktree \
  --unrestricted-runner-require-session-approval=false \
  --acknowledge-unrestricted-runner=I_UNDERSTAND_THIS_ALLOWS_UNRESTRICTED_LOCAL_COMMAND_EXECUTION

Security Model

• Secrets are opaque by default. Item fields are returned with valueState: "redacted" unless --enable-secret-reveal=true is passed.

• Plaintext reveal requires explicit consent. Tools that return secrets require acknowledgePlaintext: "I_UNDERSTAND_THIS_RETURNS_SECRET_PLAINTEXT".

• Password generators return a new plaintext secret. They require reason and acknowledgePlaintext: "I_UNDERSTAND_THIS_RETURNS_GENERATED_SECRET_PLAINTEXT", and audit the action without logging the secret.

• Destructive actions and permission mutations require per-call acknowledgement. Use acknowledgeDestructive: "I_UNDERSTAND_THIS_CAN_DELETE_1PASSWORD_DATA" for archive/delete operations and acknowledgePermissionMutation: "I_UNDERSTAND_THIS_CAN_CHANGE_1PASSWORD_PERMISSIONS" for permissions.

• Dangerous capabilities are opt-in and disabled by default, including writes, destructive actions, permission mutation, secret reveal, the script runner, the unrestricted script runner, and the unrestricted runner.

• Every sensitive action is audited to a JSONL file. Secret references and auth tokens are automatically redacted from logs.

• The default script runner uses spawn with shell: false, so shell injection is not available for commandId entries. Connect-mode workspace command resolution is intentionally broader: when workspaceCommandResolution.freeformCommands=true, the requested command runs through a non-login shell scoped by the resolved trusted workspace.

• The unrestricted script runner is session-approved and intentionally broad. When enabled, op_script_run ignores startup command catalogs and workspace trust files, then runs free-form shell commands after a local browser approval once per MCP process. Use it only for single-user local sessions you already trust.

• Remembered approvals are local, encrypted, and expiry-bound. The approval page can remember a grant for 24 hours by writing an AES-256-GCM encrypted file under ~/.onepassword-mcp; the file contains approval scope names and expiration timestamps, not 1Password secrets.

• Configuration reloads are bounded and audited. workspace_trust_reload in Connect mode and op_script_reload_allowlists in Desktop/service-account mode only reload direct startup configuration paths and manifest trust anchors configured at startup, record the reload reason, and keep the previous in-memory configuration if validation fails.

• Script secret injection is run-only. envSecretRefs values are resolved in memory, injected into the child process, redacted from returned output, and audited only by env var name, reference scheme, and reference hash.

• Unrestricted runner approval is local and in-memory. op_unrestricted_run requires a configured root plus browser approval by default. Audit entries store command hashes and lengths rather than the raw free-form command.

• Unrestricted runner roots are not a sandbox. The root limits which worktrees can request approval; it does not prevent an approved command from touching other paths allowed by the operating system.

• Bearer token comparison uses crypto.timingSafeEqual to reduce timing attack risk.

• HTTP binds to localhost (127.0.0.1) by default. It validates the Origin header, caps active sessions, expires idle sessions, and returns generic messages for server errors.

• Resources and capabilities avoid sensitive local metadata. Local paths, 1Password account names, HTTP host/port, and the op binary path are not exposed to MCP clients.

• errorMessage for scripts with sensitiveOutput: true is withheld unless output is explicitly requested with plaintext acknowledgement.

MCP Resource Notes

Resource URIs use the onepassword:// scheme instead of 1password://. Node.js URL parsing rejects schemes that start with a number, which breaks resource reads in practice.

Development

npm run lint   # TypeScript type checking
npm test       # Test suite
npm run build  # Compile to dist/

Commits must follow Conventional Commits. This project uses release-please to automate CHANGELOG generation and version bumps.

npm Publication

The publish.yml workflow uses npm Trusted Publishing through OIDC. It does not use a long-lived npm token (NPM_TOKEN, NODE_AUTH_TOKEN, or a 1Password secret) for publication.

npm-side prerequisites:

• package mcp-1password exists on npm and is associated with kefapps/onepassword-mcp-codex

• a GitHub Actions trusted publisher is configured for repository kefapps/onepassword-mcp-codex

• workflow filename: publish.yml

The GitHub workflow uses Node.js 24 and has the id-token: write permission required by npm to exchange the job OIDC identity for a short-lived publishing token. npm automatically generates provenance attestations when the package and GitHub repository are public.

Changelog

See CHANGELOG.md.

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md before opening a pull request.

Please report vulnerabilities privately instead of opening a public issue. See SECURITY.md.

License

MIT