warden-mcp

Programmatic Vaultwarden/Bitwarden vault management over MCP (Model Context Protocol), backed by the official Bitwarden CLI (bw).

This project exists to let agents and automation create/search/read/update/move vault items without re-implementing Bitwarden’s client-side crypto.

Published package: @icoretech/warden-mcp

Highlights

• MCP Streamable HTTP (SSE) endpoint at POST /sse + health check at GET /healthz
• Runtime guardrail metrics at GET /metricsz
• Item types: login, secure note, card, identity, plus an SSH key convention (secure note + standard fields)
• Attachments: create/delete/download
• Organization + collection helpers (list + org-collection CRUD)
• Safe-by-default: item reads are redacted unless explicitly revealed; secret helper tools return null unless reveal: true
• Strong fit for LLM automation: pair it with a browser-capable MCP host so an agent can fetch credentials, complete sign-in flows, read TOTP codes, and keep automating after login

Text-Only MCP Client Behavior

Some MCP hosts, including agent shells that optimize tool output for the model, only forward content[] text blocks and do not expose structuredContent to the agent. warden-mcp therefore mirrors the non-secret identifiers needed for follow-up calls into visible text by default.

For search/list tools, the text output includes concise rows with stable ids and safe metadata such as names, item type, username, URI values, organization id, folder id, and collection ids. Secret fields are not included in those summaries. Create/update/move/restore helpers that return a folder, collection, or item use the same concise visible summaries so text-only clients can immediately reuse the returned ids in follow-up calls. Delete helpers include the requested ids when no object is returned.

For scalar helper tools, the text output follows the reveal contract:

• get_username shows the username because Bitwarden treats it as non-secret
• get_password, get_totp, and get_notes show not revealed unless you pass reveal: true
• if NOREVEAL=true or KEYCHAIN_NOREVEAL=true, revealed values are still suppressed server-side

When a lookup term matches multiple login items, credential helpers return an AMBIGUOUS_LOOKUP error with visible candidate ids. Retry with term set to an exact candidate id, or call get_item with that id.

If KEYCHAIN_TEXT_COMPAT_MODE=structured_json is enabled, supported success and ambiguity/error results mirror their structuredContent into the visible text as serialized JSON instead of the human summaries.

LLM Automation Use Case

warden-mcp is not only useful for vault administration. A very practical use case is pairing it with an LLM that can also drive a browser.

That lets the agent do end-to-end authenticated workflows such as:

• open a site or backoffice in the browser
• read the right login from Vaultwarden or Bitwarden
• fill username and password without hardcoding secrets in prompts or config
• retrieve a current TOTP code with keychain_get_totp for TOTP-based MFA
• continue the real task after login, such as navigation, data entry, exports, or routine admin work

In practice, this is what makes the server useful for full automation, not just secret lookup. The same MCP session that gives the model browser control can also give it scoped access to the credentials and MFA material needed to finish the workflow.

Runtime Requirement

This package shells out to the official Bitwarden CLI, bw.

Runtime resolution order:

• BW_BIN if you set it explicitly
• bundled @bitwarden/cli optional dependency if it is present
• system bw from PATH

That means package installation can succeed even when the optional dependency is skipped by the environment. In that case you must install bw separately or point BW_BIN to it.

Explicit fallback install:

npm install -g @bitwarden/cli

Or run with an explicit binary path:

BW_BIN=/absolute/path/to/bw npx -y @icoretech/warden-mcp@latest

warden-mcp intentionally bundles a vetted @bitwarden/cli version instead of blindly following the newest upstream CLI on every release. New bw releases can change login and unlock behavior in ways that break automation, so bw upgrades should be smoke-tested against real Vaultwarden flows before bumping the bundled version. Official Bitwarden compatibility is intended, but it is not continuously proven in CI without a real Bitwarden tenant.

This repository's compose smoke now exercises both direct bw auth flows and the MCP/SDK layers with username/password auth plus user API-key auth against a real local Vaultwarden, so @bitwarden/cli bumps do not rely on unit coverage alone. The bundled postinstall compatibility shim rewrites the affected build/bw.js login strategies in place so the same Vaultwarden fallback can survive compatible CLI bumps without a version-stamped patch artifact.

Install And Run

Choose a transport

• Use --stdio when you want a local MCP host to spawn warden-mcp directly with one fixed Bitwarden profile
• Use default HTTP mode when you want one running warden-mcp service to serve multiple clients or multiple Bitwarden profiles via per-request X-BW-* headers

Local stdio mode

npx -y @icoretech/warden-mcp@latest --stdio

For stdio mode, you must provide Bitwarden credentials up front via env vars:

BW_HOST=https://vaultwarden.example.com \
BW_USER=user@example.com \
BW_PASSWORD='your-master-password' \
npx -y @icoretech/warden-mcp@latest --stdio

API key login works too:

BW_HOST=https://vaultwarden.example.com \
BW_CLIENTID=user.xxxxx \
BW_CLIENTSECRET=xxxxx \
BW_PASSWORD='your-master-password' \
npx -y @icoretech/warden-mcp@latest --stdio

Shared HTTP mode

Start one long-lived MCP server:

npx -y @icoretech/warden-mcp@latest

Verify it is up:

curl -fsS http://localhost:3005/healthz

This mode is what makes warden-mcp different from a simple local wrapper:

• the server stays stateless at the HTTP boundary
• Bitwarden/Vaultwarden credentials are sent per request via X-BW-* headers
• one running server can front different vault hosts or different identities without restarting
• it fits shared-agent and gateway setups much better than per-client local processes

Docker

docker run --rm -p 3005:3005 ghcr.io/icoretech/warden-mcp:latest

Global install

npm install -g @icoretech/warden-mcp@latest
warden-mcp

Connect From MCP Hosts

For local MCP hosts, stdio is the most portable option.

npx -y @icoretech/warden-mcp@latest --stdio

The examples below use Bitwarden API-key auth. If you prefer username/password login, replace BW_CLIENTID + BW_CLIENTSECRET with BW_USER.

CLI-based hosts

These hosts let you register warden-mcp directly from the command line:

# Codex
codex mcp add warden \
  --env BW_HOST=https://vaultwarden.example.com \
  --env BW_CLIENTID=user.xxxxx \
  --env BW_CLIENTSECRET=xxxxx \
  --env BW_PASSWORD='your-master-password' \
  -- npx -y @icoretech/warden-mcp@latest --stdio

# Claude Code
claude mcp add-json warden '{"command":"npx","args":["-y","@icoretech/warden-mcp@latest","--stdio"],"env":{"BW_HOST":"https://vaultwarden.example.com","BW_CLIENTID":"user.xxxxx","BW_CLIENTSECRET":"xxxxx","BW_PASSWORD":"your-master-password"}}'

# Qwen Code
qwen mcp add warden \
  -e BW_HOST=https://vaultwarden.example.com \
  -e BW_CLIENTID=user.xxxxx \
  -e BW_CLIENTSECRET=xxxxx \
  -e BW_PASSWORD=your-master-password \
  npx -y @icoretech/warden-mcp@latest --stdio

JSON config hosts

These hosts all use the same stdio payload shape. Only the config file location changes:

• Codex: ~/.codex/config.toml
• Cursor: ~/.cursor/mcp.json or .cursor/mcp.json
• Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json
• Qwen Code: ~/.qwen/settings.json or .qwen/settings.json

Shared JSON shape:

{
  "mcpServers": {
    "warden": {
      "command": "npx",
      "args": ["-y", "@icoretech/warden-mcp@latest", "--stdio"],
      "env": {
        "BW_HOST": "https://vaultwarden.example.com",
        "BW_CLIENTID": "user.xxxxx",
        "BW_CLIENTSECRET": "xxxxx",
        "BW_PASSWORD": "your-master-password"
      }
    }
  }
}

Codex uses TOML instead of JSON:

[mcp_servers.warden]
command = "npx"
args = ["-y", "@icoretech/warden-mcp@latest", "--stdio"]
startup_timeout_sec = 30

[mcp_servers.warden.env]
BW_HOST = "https://vaultwarden.example.com"
BW_CLIENTID = "user.xxxxx"
BW_CLIENTSECRET = "xxxxx"
BW_PASSWORD = "your-master-password"

startup_timeout_sec = 30 is a practical Codex default when using npx, because a cold first launch can spend several seconds downloading and unpacking the package before MCP initialization begins.

Windsurf

Windsurf uses the same stdio idea but stores it in ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "warden": {
      "command": "npx",
      "args": ["-y", "@icoretech/warden-mcp@latest", "--stdio"],
      "env": {
        "BW_HOST": "https://vaultwarden.example.com",
        "BW_CLIENTID": "user.xxxxx",
        "BW_CLIENTSECRET": "xxxxx",
        "BW_PASSWORD": "your-master-password"
      }
    }
  }
}

Shared HTTP connections

If your MCP host supports Streamable HTTP with custom headers, you can connect to one long-lived warden-mcp service instead of spawning a local stdio process.

Start the shared server:

npx -y @icoretech/warden-mcp@latest

Every MCP request must include:

• X-BW-Host
• X-BW-Password
• either X-BW-ClientId + X-BW-ClientSecret, or X-BW-User

Example health check:

curl -fsS \
  -H 'X-BW-Host: https://vaultwarden.example.com' \
  -H 'X-BW-ClientId: user.xxxxx' \
  -H 'X-BW-ClientSecret: xxxxx' \
  -H 'X-BW-Password: your-master-password' \
  http://localhost:3005/healthz

Example MCP endpoint:

http://localhost:3005/sse?v=2

This shared-server mode is useful when:

• one MCP gateway needs to front multiple Bitwarden profiles
• you want to rotate vault credentials per request instead of per process
• you are integrating from a custom client or agent host that can attach HTTP headers
• you want one always-on service instead of each editor spawning its own bw-backed subprocess

Headerless SSE clients (claude.ai web, similar hosts)

Some web MCP hosts can connect to a Streamable HTTP / SSE endpoint but cannot attach custom X-BW-* headers. In that case, run warden-mcp as a single-tenant shared server with fixed server-side credentials and explicit env fallback enabled:

BW_HOST=https://vaultwarden.example.com \
BW_CLIENTID=user.xxxxx \
BW_CLIENTSECRET=xxxxx \
BW_PASSWORD='your-master-password' \
KEYCHAIN_ALLOW_ENV_FALLBACK=true \
npx -y @icoretech/warden-mcp@latest

Some hosted or browser-based MCP clients expect the server URL to be available over HTTPS rather than plain http://localhost:3005/....

If you are running warden-mcp locally, that usually means putting one of these in front of it:

• a local HTTPS reverse proxy in front of warden-mcp
• a temporary tunnel such as Cloudflare Tunnel, ngrok, or Tailscale Funnel

Then point the client at the HTTPS endpoint, for example:

https://warden-mcp.example.com/sse?v=2

This works because headerless HTTP requests can inherit the server's own BW_* configuration when KEYCHAIN_ALLOW_ENV_FALLBACK=true. You can use BW_USER + BW_PASSWORD here as well; the example above shows the API-key variant because it is the default convention used elsewhere in this README.

Important limits:

• this is single-tenant only: every headerless client gets the same vault identity
• per-request profile switching does not work in this mode
• if a client can send X-BW-* headers, those headers still take priority over the server env
• anyone who can reach that endpoint inherits that vault access, so keep it behind trusted network access, VPN, tunnel access policies, IP allowlists, or another private boundary

Client examples for shared HTTP mode:

# Claude Code
claude mcp add-json warden '{"type":"http","url":"http://localhost:3005/sse?v=2","headers":{"X-BW-Host":"https://vaultwarden.example.com","X-BW-ClientId":"user.xxxxx","X-BW-ClientSecret":"xxxxx","X-BW-Password":"your-master-password"}}'

// Cursor (~/.cursor/mcp.json)
{
  "mcpServers": {
    "warden": {
      "url": "http://localhost:3005/sse?v=2",
      "headers": {
        "X-BW-Host": "https://vaultwarden.example.com",
        "X-BW-ClientId": "user.xxxxx",
        "X-BW-ClientSecret": "xxxxx",
        "X-BW-Password": "your-master-password"
      }
    }
  }
}

// Qwen Code (~/.qwen/settings.json)
{
  "mcpServers": {
    "warden": {
      "httpUrl": "http://localhost:3005/sse?v=2",
      "headers": {
        "X-BW-Host": "https://vaultwarden.example.com",
        "X-BW-ClientId": "user.xxxxx",
        "X-BW-ClientSecret": "xxxxx",
        "X-BW-Password": "your-master-password"
      }
    }
  }
}

// Windsurf (~/.codeium/windsurf/mcp_config.json)
{
  "mcpServers": {
    "warden": {
      "serverUrl": "http://localhost:3005/sse?v=2",
      "headers": {
        "X-BW-Host": "https://vaultwarden.example.com",
        "X-BW-ClientId": "user.xxxxx",
        "X-BW-ClientSecret": "xxxxx",
        "X-BW-Password": "your-master-password"
      }
    }
  }
}

Codex currently fits better with stdio here, because its MCP config supports a bearer token env var for remote servers but not arbitrary custom X-BW-* header injection.

Verify bw is available

bw --version

If that fails after install, your environment likely skipped the optional @bitwarden/cli dependency. Install it explicitly:

npm install -g @bitwarden/cli

How It Works

The server executes bw commands on your behalf:

• In HTTP mode, Bitwarden/Vaultwarden connection + credentials are provided via HTTP headers per request. Env-var fallback is disabled by default; set KEYCHAIN_ALLOW_ENV_FALLBACK=true to enable it for single-tenant HTTP deployments.
• In stdio mode, Bitwarden/Vaultwarden credentials are loaded once from BW_* env vars at startup.
• The server maintains per-profile bw state under KEYCHAIN_BW_HOME_ROOT and pins BITWARDENCLI_APPDATA_DIR inside that profile so the Bitwarden CLI keeps a stable local device/app identity across restarts instead of looking like a fresh client every time.
• Writes can optionally call bw sync (internal; not exposed as an MCP tool).

Timeout handling is also process-tree aware: if a bw command hangs, warden-mcp kills the full spawned process group rather than only the direct parent process. That prevents timed-out auth attempts from leaving orphaned bw/shell child processes behind.

Required Headers

• X-BW-Host (must be an HTTPS origin, for example https://vaultwarden.example.com)
• X-BW-Password (master password; required to unlock)
• Either:
  • X-BW-ClientId + X-BW-ClientSecret (API key login), or
  • X-BW-User (email for user/pass login; still uses X-BW-Password)
• Optional:
  • X-BW-Unlock-Interval (seconds)

Security Model

There is no built-in auth layer in v1. Run it only on a trusted network boundary (localhost, private subnet, VPN, etc.).

Credential resolution:

• HTTP mode requires X-BW-* headers on every request by default. Without them, tools return an error.
• Stdio mode reads BW_* env vars at startup (single-tenant).
• To allow HTTP mode to fall back to server env vars when headers are absent (single-tenant HTTP), set KEYCHAIN_ALLOW_ENV_FALLBACK=true. Security warning: this means any client that can reach the HTTP endpoint gets full vault access without providing credentials. Only use this behind network-level access control.

Mutation control:

• Set READONLY=true to block all write operations (create/edit/delete/move/restore/attachments).
• Set NOREVEAL=true to force all reveal parameters to false server-side. Clients can still request reveal: true, but the server will silently downgrade to redacted output. This prevents prompt injection from tricking an LLM agent into exfiltrating secrets.
• Set KEYCHAIN_TEXT_COMPAT_MODE=structured_json to mirror supported structured tool results into TextContent as serialized JSON. This is useful for text-only MCP clients that ignore structuredContent, but it also duplicates revealed secrets into the plain-text transcript.
• Tool names default to keychain_*. Override TOOL_PREFIX to change the namespace and TOOL_SEPARATOR to change the separator (default _, set . for legacy clients).
• Session guardrails:
  • KEYCHAIN_SESSION_MAX_COUNT (default 32)
  • KEYCHAIN_SESSION_TTL_MS (default 900000)
  • KEYCHAIN_SESSION_SWEEP_INTERVAL_MS (default 60000)
  • KEYCHAIN_MAX_HEAP_USED_MB (default 1536, set 0 to disable memory fuse)
  • KEYCHAIN_METRICS_LOG_INTERVAL_MS (default 0, disabled)
  • NOREVEAL / KEYCHAIN_NOREVEAL (default false; force all reveals to false)
  • KEYCHAIN_ALLOW_ENV_FALLBACK (default false; HTTP env-var credential fallback)
  • KEYCHAIN_TEXT_COMPAT_MODE (default unset; set to structured_json to copy structured results into TextContent)

Redaction defaults (item reads):

• Login: password, totp
• Card: number, code
• Identity: ssn, passportNumber, licenseNumber
• Custom fields: hidden fields (Bitwarden type: 1)
• Attachments: attachments[].url (signed download URL token)
• Password history: passwordHistory[].password

Reveal rules:

• Tools accept reveal: true where applicable (default is false).
• Secret helper tools (get_password, get_totp, get_notes, generate, get_password_history) return structuredContent.result = { kind, value, revealed }.
  • When reveal is omitted/false, value is null (or historic passwords are null) and revealed: false.

Production Deployment Checklist

If you run warden-mcp beyond local development, review these items:

1. TLS everywhere. Always terminate TLS in front of the HTTP endpoint. X-BW-* headers carry master passwords in cleartext — without TLS they are visible to anyone on the network.

2. Network isolation. Bind the server to 127.0.0.1 or place it behind an authenticated reverse proxy. The service has no built-in authentication; anyone who can reach /sse can issue vault operations.

3. Do not enable KEYCHAIN_ALLOW_ENV_FALLBACK on shared networks. This flag makes the server's own vault credentials available to any HTTP client that omits headers. Only use it in single-tenant setups where the network is fully trusted.

4. Enable READONLY=true when writes are not needed. This blocks all mutating tools at the MCP layer, limiting blast radius if an agent or client is compromised.

5. Restrict filesystem access to /data/bw-profiles. The bw CLI stores decrypted state under its HOME directory. Ensure the profile directory is not world-readable and is mounted with appropriate permissions (the Docker image runs as non-root by default).

6. Disable debug logging in production. KEYCHAIN_DEBUG_BW and KEYCHAIN_DEBUG_HTTP emit request details and CLI invocations to stdout. Debug logs may include session metadata and request structure. Keep them off unless actively troubleshooting.

7. Set NOREVEAL=true when secrets should never leave the server. This forces all reveal parameters to false server-side, regardless of what the client requests. Use this when the MCP host is an LLM agent that could be influenced by prompt injection — it prevents tricked agents from exfiltrating passwords or TOTP codes.

8. Monitor /metricsz. The endpoint is intentionally unauthenticated (for scraper compatibility) but exposes session counts, heap usage, and rejection counters. If this data is sensitive in your environment, restrict access at the network level.

9. Only enable KEYCHAIN_TEXT_COMPAT_MODE=structured_json for text-only clients you trust. It improves compatibility with clients that ignore structuredContent, but any revealed secret will also appear in plain-text TextContent, making transcript leakage easier.

Quick Start

Minimal local run

Run the published package in HTTP mode and verify the server is up:

npx -y @icoretech/warden-mcp@latest
curl -fsS http://localhost:3005/healthz

Local Development

Docker Compose

Starts a local Vaultwarden + HTTPS proxy (for bw), bootstraps a test user, and runs the MCP server.

cp .env.example .env
make up
curl -fsS http://localhost:3005/healthz

Run integration tests:

make test

make test now runs both compose-backed auth paths and verifies them at the raw CLI plus MCP/SDK layers:

• user/password login from .env.test
• api-key login from tmp/vaultwarden-bootstrap/apikey.env, generated by the bootstrap step and kept out of git via tmp/

The compose bootstrap step depends on the Playwright Docker image matching the playwright npm package version. If either moves, update the other in the same change.

Run session flood regression locally (guardrail sanity):

npm run test:session-regression

Local dev (host)

npm install
cp .env.example .env
npm run dev

Tool Reference (v1)

Choosing the right tool

• Start with keychain_search_items when you know a name, URI, username, folder, collection, or item type but not the exact item id. Then call keychain_get_item with the returned id for the full item shape.
• Personal folder tools manage one user's vault folders, for example Example Folder. Organization collection tools manage shared organization-scoped collections, for example Example Collection. organizationId is required for keychain_list_org_collections, keychain_create_org_collection, keychain_edit_org_collection, and keychain_delete_org_collection, and optional only for keychain_get_org_collection when you want to narrow a direct id lookup.
• keychain_move_item_to_organization moves an item out of the personal vault and into an organization, optionally assigning collection ids at the same time.
• keychain_send_create is the quick path for text or file Sends through the normal bw send flags. keychain_send_template, keychain_send_create_encoded, and keychain_send_edit are for the full Bitwarden Send JSON template or an encoded edit payload.
• Pass reveal: true only to tools that can return secrets, such as password, TOTP, notes, generated secrets, password history, or full item reads. By default these stay redacted. Text-only MCP clients follow the same reveal contract described in Text-Only MCP Client Behavior, and server-side NOREVEAL settings still win.

Vault/session:

• keychain_status
• keychain_sync (pull latest vault data from server via bw sync)
• keychain_sdk_version (returns the Bitwarden SDK version used by the CLI)
• keychain_encode (base64-encode a string via bw encode)
• keychain_generate (returns a generated secret only when reveal: true)

Items:

• keychain_search_items, keychain_get_item, keychain_update_item
• keychain_create_login, keychain_create_note, keychain_create_card, keychain_create_identity, keychain_create_ssh_key
• keychain_delete_item, keychain_restore_item

Folders:

• keychain_list_folders, keychain_create_folder, keychain_edit_folder, keychain_delete_folder

Orgs/collections:

• keychain_list_organizations, keychain_list_collections
• keychain_list_org_collections, keychain_create_org_collection, keychain_edit_org_collection, keychain_delete_org_collection
• keychain_move_item_to_organization

Attachments:

• keychain_create_attachment, keychain_delete_attachment, keychain_get_attachment
• keychain_get_item exposes safe attachment metadata (id, fileName, size) while redacting signed download URLs, so clients can discover the exact attachment id before downloading
• keychain_get_attachment accepts an attachment id or an unambiguous filename and returns { filename, bytes, contentBase64 }; decode contentBase64 locally when you need the original file bytes

Sends:

• keychain_send_list, keychain_send_template, keychain_send_get
• keychain_send_create (quick create via bw send)
• keychain_send_create_encoded, keychain_send_edit (advanced create/edit via bw send create|edit)
• keychain_send_remove_password, keychain_send_delete
• keychain_receive

Direct “bw get …” helpers:

• keychain_get_username (returns { kind:"username", value, revealed:true })
• keychain_get_password / keychain_get_totp / keychain_get_notes (only return real values when reveal: true)
• keychain_get_uri, keychain_get_exposed
• keychain_get_folder, keychain_get_collection, keychain_get_organization, keychain_get_org_collection
• keychain_get_password_history (only returns historic passwords when reveal: true)

Known Limitations

• bw list items --search (and thus keychain_search_items) does not reliably search inside custom field values.
• SSH keys are stored as secure notes in v1 (until bw supports native SSH key item creation).
• High-risk CLI features are intentionally not exposed yet (export/import).

Contributing

See AGENTS.md for repo guidelines, dev commands, and testing conventions.