Livespace CRM MCP Server

Unofficial MCP server for Livespace CRM. It exposes 11 intent-shaped tools instead of mirroring the raw API and targets the stateless Streamable HTTP transport in MCP spec 2026-07-28.

Version 0.1.1 is available on npm, in the official MCP Registry, and as a GitHub Release.

The v1 implementation is complete in this repository. It has six read tools and five optional write tools, read-only defaults, bounded API access, sanitized errors and an elicitation-first confirmation flow. Use a test Livespace account while evaluating it.

Why this shape?

Livespace's RPC API has about 80 methods, but it does not provide sorting, aggregation or a direct operation for setting a deal stage. The server groups those lower-level calls into tasks an MCP client can use safely:

• discovery before IDs are used;
• bounded search, record reads and analysis;
• batch previews before writes;
• per-item outcomes and post-write checks;
• explicit handling of Livespace-specific stage moves and notifications.

Tools

move_deals_to_stage works by checking and unchecking process steps because Livespace has no "set stage" call. A deal stands at its furthest checked step. A backward move therefore unchecks completed steps and can change the historical meaning of those checkboxes.

notify_user validates the recipient and can add a deep link to a record. Livespace provides no notification read-back, so the tool reports a successful request as dispatched, never as delivered.

Requirements

• A Livespace plan with API access, currently Automation or higher.
• One Livespace API key and secret from Account settings -> API -> Users. Use a dedicated Livespace API user with only the permissions this MCP needs. Every API operation inherits that user's permissions.
• Bun 1.3.14 or newer (bun.sh). The current server uses Bun.serve and has no Node.js or Cloudflare Workers adapter.

The v1 deployment model is single-user: one Livespace credential pair and, if enabled, one MCP bearer token protects the server. There is no OAuth or multi-user credential routing.

Install and run from npm

The package is distributed through npm's public registry, but Bun is its runtime. You do not need Node.js or the npm CLI to run it.

For a first evaluation, create a private working directory outside a Git repository. Add a .env file there with your own Livespace credentials:

LIVESPACE_SUBDOMAIN=
LIVESPACE_API_KEY=
LIVESPACE_API_SECRET=

Fill the three empty values, protect the file, then start the published package:

chmod 600 .env
bunx livespace-crm-mcp

The default endpoint is http://127.0.0.1:3020/mcp. Keep the process running while your MCP client is connected. Start in read-only mode, call health, then use crm_metadata before any operation that needs a user, process, stage, group or dictionary ID.

bunx downloads the package from npm and caches it locally. To pin this security release, run bunx livespace-crm-mcp@0.1.1.

Connect an MCP client

Configure a client that supports Streamable HTTP with this server URL:

{
  "url": "http://127.0.0.1:3020/mcp"
}

The exact configuration field differs between clients. Set MCP_AUTH_TOKEN and configure the client to send Authorization: Bearer <your-token>, even on loopback. Authentication is mandatory when write tools are enabled.

This package exposes Streamable HTTP, not stdio. Some MCP clients can connect to the local URL but cannot launch bunx for you, so start the command in a separate terminal. Clients that accept only stdio are not supported yet. Each HTTP request must contain one JSON-RPC message; top-level batch arrays are rejected before dispatch.

Configuration

The server refuses a non-loopback bind unless MCP_AUTH_TOKEN and MCP_ALLOWED_HOSTS are set. It also refuses to enable writes without both MCP_AUTH_TOKEN and MCP_REQUEST_STATE_KEY. Once MCP_AUTH_TOKEN is configured, every /mcp request needs that Bearer token, including on loopback. Host and Origin checks run before the MCP handler.

Generate independent values for MCP_AUTH_TOKEN and MCP_REQUEST_STATE_KEY by running openssl rand -hex 32 twice. Do not reuse a Livespace credential.

The Bun process does not terminate TLS. Put a TLS-capable reverse proxy in front of every network-exposed deployment. The server reads credentials from environment variables, commonly through Bun's .env loading; protecting .env at rest is the operator's responsibility. Never commit it. The smoke script can also read the sandbox credentials from the macOS Keychain.

Write safety

The normal flow is elicitation-first:

1. A plain write-tool call or dryRun: true builds a plan and writes nothing.
2. On an elicitation-capable client, the server asks a human to approve the plan. A confirm: true argument cannot bypass this prompt.
3. The confirmation exchange uses a signed requestState that is valid for five minutes and consumed after an accepted or declined response. It is bound to the authenticated principal when present, the tool, arguments and preview.
4. If the relevant records changed before approval, the tool returns recordsChanged: true, writes nothing and presents a fresh plan.

Write execution is disabled by default on clients without form elicitation. Those clients can preview, but confirm: true is refused. An operator can set MCP_ALLOW_UNBOUND_WRITE_CONFIRMATION=true for compatibility. In that mode, confirm: true executes without signed proof that a human saw the preview.

Results distinguish these cases:

• verification: verified means the server re-read comparable fields after the write.
• A successful write can still report verification: unavailable when a follow-up read failed, no sent field had an independent comparator or the upstream API exposes no read-back. Inspect that item's error and re-read the record where possible; do not retry the write.
• unknown_outcome means the request may or may not have landed. Never retry it blindly.
• not_attempted means that item was not sent, usually because a time budget expired or an earlier item hit the upstream rate limit. Those unsent items can be submitted later; follow the error hint for the rate-limited item.

Notification success is reported as dispatched, never as delivered. If delivery matters, confirm it by another channel instead of sending a duplicate notification.

Security and privacy