Most teams that adopt AI inside Odoo end up with a mess of point integrations: one module calls a chat model directly from a Python method, another hardcodes an API key in a server action, a third talks to some external tool over HTTP with no logging. Keys leak into ir.config_parameter in plaintext, there is no rate limiting, no per-user quota, and no single place to see what was sent to which provider. Odoo core gives you ir.actions.server, automated actions, and the ORM, but it has no concept of an AI provider abstraction or a tool-calling protocol broker — so every AI feature reinvents credential handling, retries, and cost tracking, and none of it is auditable. When finance asks "how much did we spend on AI last month and who spent it," nobody can answer.

MCP Gateway is a build-to-order Odoo module that puts one broker in front of everything. We model AI providers and MCP tool servers as first-class Odoo records (models.Model with fields for endpoint, auth scheme, model catalog, and cost-per-token), so a single mcp.gateway.provider and mcp.gateway.tool registry becomes the only path any AI call takes. Credentials are encrypted at rest and never rendered back to the UI; access to each provider and tool is governed by ir.model.access.csv plus record rules, so a warehouse user can be allowed to call a document-extraction tool but not an expensive reasoning model. Every request flows through a broker service that handles provider selection, streaming, tool-call routing (the Model Context Protocol handshake, tool discovery, and invocation), retries with backoff, and a full mcp.gateway.log audit trail with token counts, latency, cost, create_uid, and the originating record. Downstream, your other modules — or automated actions, or scheduled jobs — call one clean internal API instead of a dozen SDKs, and you can swap or add a provider without touching business logic.

Technically, the gateway is exposed three ways so it fits however your team works. Inside Odoo, a Python service method lets your custom code and @api.depends computes request completions or tool runs. For no-code users, we ship server actions and automated-action hooks so a stage change on a record can trigger a governed AI call. And over the wire, the gateway is reachable through Odoo's XML-RPC/JSON-RPC API and an optional authenticated controller, so external agents and MCP-speaking clients reach your Odoo tools through the same policy and audit layer rather than going around it. OWL-based configuration screens, QWeb-rendered usage reports, and a cost dashboard round it out. Everything is built cleanly against the version you run — Odoo 17.0, 18.0, or 19.0 — and we account for Community vs Enterprise differences (for example, where Enterprise studio-built automations or web features exist, the gateway integrates with them; on Community we implement the equivalent with standard framework primitives).

Because this is made to order, we start from your actual providers, tools, security model, and cost-control rules, then design the record schema, policy matrix, and dashboards around them before writing code. You get installable source for your version, a staging UAT pass, and a clean handover. Typical delivery is one working week from confirmed scope. Pricing starts from $999 (indicative, single-company base scope); the number of AI providers and MCP tool servers integrated, multi-company routing and per-company quotas, localization depth, and the volume of existing AI call sites to be migrated onto the gateway increase the quoted scope.