Adopt MCP this quarter, leave A2A alone, and do not let a download chart make either decision for you.
The chart is the loud part. The Model Context Protocol reached 97 million monthly SDK downloads by March 2026, per figures the Linux Foundation and Atlan cite. A download count measures installs, and installs include build servers, dependency updates and developers trying a package once. It counts machines. Whatever it says about engineering teams, it says nothing about whether the protocol fits your business, so reason from the problem and not from the chart.
The problem is arithmetic. Run four agents against ten systems with no shared protocol and every pairing is a bespoke connector: up to forty pieces of glue code, each breaking when either side changes. With a shared protocol each agent speaks it once and each system exposes it once, so you maintain fourteen adapters, and the gap widens with every addition. Anthropic released MCP in late 2024 to standardize how an agent reaches tools and data. A system publishes an MCP server and any compatible agent can use it. That is a good case and it is the whole case.
A2A is a different tool for a different stage. Google's Agent2Agent protocol, now under the Linux Foundation, standardizes how separate agents built by different teams find each other through a machine-readable Agent Card and hand work over; the Linux Foundation reports more than 150 supporting organizations. You need discovery when agents from vendors who do not talk to each other must delegate without anyone coordinating them. Most firms run one or two agents from a single provider, where a direct handoff between two agents you control is simpler and easier to debug. Revisit A2A when a concrete handoff forces it.
What vendors underplay is that a protocol is a new door, not a new lock. An MCP server in front of your ERP is a new interface to your most important data, and it runs with whatever credential it was given. Wire it to an administrator's API key because that was quickest and every agent that connects inherits administrator power, including one that has just read a hostile email. The fix is old and dull: a dedicated user with the narrowest role the job allows. Odoo supports per-user API credentials, so an agent's reach can be limited to the models and documents its task needs. The careers page on this site reads job postings from Odoo that way, through a dedicated user with its own key and not an admin token. The prompt injection risk is why this is urgent: the more tools an agent can reach, the more one poisoned input can do.
So start read-only, add write actions one at a time with their own approval rule, and log every call. A standard interface makes consistent logging possible, which feeds the decision log described in our post on AI agent observability. If you want a second view on where MCP belongs in your stack, our AI services team can walk through it, or get in touch.
