LocalTry MCP

Use LocalTry from ChatGPT and other MCP clients while keeping every business strictly isolated.

LocalTry MCP is a remote Model Context Protocol server for operating a LocalTry CRM with plain-language instructions. It runs on Cloudflare Workers, uses OAuth 2.1, and binds every connection to one verified LocalTry user, business, role, and permission set.

What it enables

Search customers, companies, contacts, leads, jobs, estimates, and invoices.
Create and update validated CRM records.
Create tenant-owned AI agents that become available in Flow Studio.
Inspect the complete tenant workspace architecture.
Search every registered page, action, agent, integration, and tenant addition.
Submit tenant-only workspace customizations through LocalTry Builder.
Run saved workflows.
Ask LocalTry Command to carry out multi-step business operations.
Review customization versions and restore an earlier version.
Read a tenant-scoped activity and audit history.

Available tools

Tool	Purpose
`get_workspace_overview`	Inspect the connected business's pages, modules, fields, workflows, integrations, and customization history.
`search_localtry_features`	Find the exact registered LocalTry page, action, agent, integration, or tenant addition before acting.
`search_crm`	Search tenant-scoped CRM records without exposing raw SQL.
`create_or_update_crm_record`	Create or update a validated CRM record.
`create_workflow_agent`	Describe and save a tenant-owned AI agent for use in Flow Studio workflows.
`request_workspace_customization`	Submit the same tenant-scoped engineering request as the Customize Workspace prompt.
`get_workspace_customization_status`	Read the latest customization conversation, progress, questions, and result.
`list_workspace_versions`	Review the workspace's customization history.
`restore_workspace_version`	Restore an approved earlier version for the connected workspace.
`run_workflow`	Run one of the business's saved workflows.
`run_social_research`	Start one public social or web research search and save it in the connected workspace.
`run_localtry_command`	Use LocalTry Command for verified, multi-step CRM and business operations.
`get_recent_activity`	Review recent CRM, workflow, and customization activity.

The tenant boundary

There is one shared MCP service, but every authorization is logically private. The client cannot supply or change a tenant identifier.

text

MCP client
  -> LocalTry OAuth consent
  -> token bound to user + business + role + scopes
  -> LocalTry MCP Worker
  -> private Cloudflare service binding
  -> LocalTry domain services
  -> business-scoped data and workspace runtime

The MCP Worker does not receive a database binding. It can reach LocalTry only through the private service contract in docs/localtry-bridge-contract.md.

Remote endpoint

The production endpoint is live:

text

https://mcp.localtry.com/mcp

Health check:

text

https://mcp.localtry.com/health

Clients discover OAuth 2.1 metadata automatically. Authorization supports PKCE, dynamic client registration, short-lived access tokens, and refresh tokens. Users sign in to LocalTry and approve one visible workspace; no LocalTry API key is copied into the MCP client.

Local development

Requirements: Node.js 20 or newer and a Cloudflare account.

Terminal

npm install
npm run types
npm run check
npm run dev

Create an OAuth KV namespace and replace the placeholder namespace ID in wrangler.jsonc:

Terminal

npx wrangler kv namespace create OAUTH_KV

The LOCALTRY_API service binding must point to the named LocalTryMcpBridge WorkerEntrypoint exported by the LocalTry CRM deployment.

See the bridge contract for the exact trust boundary.

ChatGPT

ChatGPT supports remote MCP apps through developer mode. See docs/chatgpt.md for the connection flow and current plan requirements. Creating a private custom app connects this server to a ChatGPT account or workspace; publishing in the public ChatGPT app directory is a separate OpenAI review process.

The non-secret public-review listing, starter prompts, test cases, and release notes are maintained in docs/openai-submission.md.

Design principles

No raw SQL tool.
No tenant ID tool argument.
No secret-reading tool.
Goal-oriented tools instead of mirroring the entire API.
Explicit approval for consequential workspace changes.
Versioning, audit events, and restore support.
Permission checks in both the MCP Worker and the CRM domain layer.

Status

Production is deployed at mcp.localtry.com. OAuth authorization, refresh tokens, MCP initialization, tool discovery, tenant binding, and a live workspace read have been verified end to end. The public Worker reaches the CRM only through a private Cloudflare service binding and has no database binding.

License

MIT