The situation
The Model Context Protocol has made it easy to connect an AI agent or assistant to tools: a ticketing system, a CRM, a database, an internal API. That ease is the problem. In many organisations, MCP servers are appearing with a developer’s personal access token, no distinction between reading and writing, no rate limits and no log of what the agent actually did.
When the question reaches a platform lead or security sponsor, it is usually framed like this: we need agents to use internal tools through MCP; how do we design identity, authorisation and safe failure handling so we can say yes?
This page is for commissioning that integration as a bounded delivery, with controls you can test and a team that can run it.
What the work involves
Inventory before integration. I work with each system owner to list the capabilities an agent might need, then classify each: what data it exposes, what side effects it has, whether those side effects can be reversed, and who owns the system. Read-only lookups and irreversible actions should not share a credential or an approval rule.
Identity and scopes per tool. For each tool we decide whose authority the call runs under. Delegated user access means the agent can never see more than the person it is acting for. A service identity is narrower and used where delegation is not possible. Scopes are defined per tool, default-deny, so an agent cannot discover and call something it was not granted.
Controls outside the model. This follows the Substrate Pattern I have published: the model can propose any call, but a layer it cannot bypass decides whether the call happens. Concretely that means schema validation on every argument, rate limits per tool and per identity, approval steps for consequential actions, and timeouts and clear error results so agents fail safely rather than retrying blindly.
Audit by design. Every call emits a structured event: user, agent, tool, arguments (with sensitive fields masked), result and latency. Security can answer “what did the agent do, for whom” without reconstructing it from application logs.
The signature deliverable
You receive the tool inventory, permissioned integration, negative tests and operational handover. Illustrative example of a tool inventory extract:
| Tool | System | Side effect | Runs as | Control |
|---|---|---|---|---|
| search_tickets | Ticketing | None | Delegated user | Scope: read; rate limit |
| add_ticket_comment | Ticketing | Reversible | Delegated user | Scope: comment; audit |
| issue_refund | Billing | Irreversible, financial | Service identity | Human approval; amount ceiling; audit |
| run_sql | Warehouse | None | Read-only role | Allow-listed views only; row limit |
Illustrative example showing the format, not a client deliverable.
How acceptance is judged
Acceptance criteria are agreed after the access design: every tool in scope is implemented with its agreed identity, scope and controls; the negative test suite passes, meaning every out-of-scope, injected or malformed call is refused and logged; approval steps cannot be skipped; and audit events are complete for a sample of real calls. The security sponsor and platform owner sign off.
Ownership and handover
The servers live in your repositories and run in your infrastructure. The platform owner receives deployment configuration, the credential rotation procedure, the process for adding a new tool through the same inventory and test steps, and guidance on reading the audit trail. I do the work personally; any specialist help is disclosed and approved by you first. Independent practitioner; not affiliated with or endorsed by Anthropic or the MCP maintainers.
Boundaries
If you need approval tiers, stop conditions and adversarial policy testing for agents that already take real actions, see governed AI agent implementation. If agent workflows fail halfway and retries repeat actions, see agent reliability and recovery engineering. If you need the whole agent built, start at production AI agents and agent infrastructure. If the tools exist to serve one business workflow, AI workflow automation may be the better frame.