Teams that run Odoo increasingly want an AI agent to read a customer's open orders, draft a quotation, check stock, or summarize a project — but wiring a general-purpose AI assistant into a live ERP safely is where the road runs out. Odoo already exposes everything over XML-RPC and JSON-RPC, but those APIs speak in raw model names, field lists, and integer IDs; they have no notion of "tools," no self-describing schema an agent can discover, and no natural way to say "this agent may read sale.order and create crm.lead, but may never touch account.move or delete anything." Hand an autonomous agent a full API key and you have handed it your whole database. Odoo core gives you the door but not the doorman.

We build a dedicated MCP Server module that turns your Odoo instance into a first-class, permission-scoped tool provider for AI agents. The module implements the Model Context Protocol handshake and exposes a curated catalog of tools — each one a controlled operation like search_partners, read_sale_order, create_lead, or run_report — declared with a JSON schema the agent can discover at connect time. Under the hood every tool is a thin, audited wrapper over the standard ORM (models.Model, search, read, create, write), never a raw SQL bypass. Crucially, each tool call executes as a specific Odoo user, so it inherits that user's ir.model.access.csv rights and every ir.rule record rule automatically: an agent scoped to a sales user physically cannot read accounting entries or another company's records, because Odoo's own security kernel refuses. We add a per-agent credential model, a configurable allow-list of tools, optional field-level masking, rate limits, and a full request/response audit log (mcp.call.log) with an accompanying QWeb report so you can see exactly what every agent did and when.

Technically the server ships as a proper module: an __manifest__.py declaring the dependency graph, new models for agent registrations and tool scopes with @api.depends computed status fields, security defined in ir.model.access.csv plus record rules that isolate agent credentials per company and per team, OWL/XML backend views for administrators to register agents and toggle tool scopes, and controller endpoints that terminate the MCP transport (HTTP with streaming, plus JSON-RPC fallback for constrained environments). Write-class tools can be gated behind an approval step routed through Odoo automated actions and activities, so a "create purchase order" call becomes a draft awaiting a human click rather than a silent commit. Everything is built and tested against Odoo 17.0, 18.0, and 19.0, and works on both Community and Enterprise — where Enterprise-only apps are in scope (for example Studio-generated models or specific accounting flows), we expose them through the same scoped tool layer.

Because this is a made-to-order build, nothing ships until we agree on scope. We start with a short scoping call to map which models, actions, and reports your agents actually need, which Odoo users they should impersonate, and where human approval gates belong; then we build your version, run UAT on a staging clone of your database, and deploy to production with a rollback plan. Typical delivery is 2 to 4 weeks from confirmed scope. Pricing starts from $249 (indicative, single-company base scope); the number of exposed tools, the depth of write/approval workflows, multi-company isolation, and integration with more than one AI agent platform increase the quoted scope. You receive a fixed quote after the scoping call — never a surprise invoice.