Teams that run Odoo increasingly want to ask an AI assistant plain-language questions ("which sale orders are stuck in sent past their commitment date?", "draft a purchase order for the low-stock parts", "summarize this quarter's helpdesk backlog") and have the assistant act against live company data instead of a stale export. Out of the box, Odoo exposes an XML-RPC/JSON-RPC API, but that API speaks in raw model names, field lists, and integer IDs. It has no notion of an AI tool contract, no permission-scoped tool catalog, and no structured way to describe "here is what you may read and what you may do." Bolting a general-purpose assistant onto that raw surface either leaks everything the connecting user can see, or forces every prompt to hand-roll RPC calls. That gap, a safe and described and action-aware bridge between an AI client and Odoo, is what this build closes.

We build a Model Context Protocol (MCP) server tailored to your Odoo deployment. On the Odoo side we ship a proper addon (__manifest__.py, an mcp.tool and mcp.audit.log model layer built on models.Model with fields, compute methods and @api.depends, plus ir.model.access.csv and record rules) that defines exactly which models, fields, and operations are exposed as MCP tools and resources. The MCP server process itself authenticates to Odoo over JSON-RPC/XML-RPC using a dedicated technical user whose access rights are the real enforcement boundary, so an assistant can never do more than that user is permitted to do. Read tools map to search_read with server-side domains and field allowlists; write tools (create quotation, log an activity, post an internal note) are wrapped so they validate inputs, respect record rules, and write an audit row before committing. Every tool is described with a typed schema the AI client discovers at connection time, so the assistant knows the arguments, defaults, and return shape without guessing.

The server runs alongside Odoo (stdio for a local desktop assistant, or an HTTP/SSE transport behind your reverse proxy for team use) and can drive any MCP-compatible AI client. We can expose curated resources (a company's product catalog, open pipeline, current stock) and curated actions (draft SO/PO, create CRM lead, schedule an activity, trigger an existing automated action or server action). Odoo QWeb reports can be surfaced as downloadable tool outputs, and Enterprise-only capabilities (for example studio-defined models or subscription objects) are wired only when your edition supports them, with a Community-safe fallback path documented where it differs.

Because this is build-to-order, we start with a short scoping call to agree which models, fields, and actions belong in your tool catalog, your Odoo edition and version (17.0, 18.0, or 19.0), and which AI client(s) will connect. From confirmed scope, typical delivery is two to four weeks, ending with UAT on a staging database and a supported go-live. Pricing starts from $249 (indicative, single-company base scope); the number of exposed models and write actions, multi-company record-rule complexity, custom integrations, and whether the transport must be hardened for external/team access increase the quoted scope. You receive a fixed quote after the scoping call, never a surprise charge.