Capability 02 of 04 · Infrastructure
MCP & Tool Integration
How agents connect to external tools, APIs, and internal systems via MCP (now universal standard) and plugins.
5
maturity levels
15
practices
15
guides
L1 · Stage 01Assisted
Practices - what it looks like3 guides
- Agent uses built-in tools onlyZero MCP is the baseline state: your AI agent has no programmatic connection to any tool, system, or data source outside its training data.guide→
- Agent relies on public / general knowledgeThe agent answers from its training data, so it knows what the Stripe API looks like in general but nothing about how your integration actually works.guide→
- Integrations done by copy-pasteThe universal first way of giving an agent context: paste the error, the schema, the ticket into the chat. When the conversation ends, so does the context.guide→
Criteria - what to measure
- 01Agents use their built-in tools
- 02Agents draw on their general / built-in knowledge
- 03Team is aware of MCP as a standard for agent-tool integration
- 04Integrations, if any, are manual (copy-paste between tools)
L2 · Stage 02Delegated
Practices - what it looks like3 guides
- 1-3 basic MCP servers (Git, Jira, docs)The first practical MCP deployment replaces the most expensive copy-paste operations with programmatic connections.guide→
- Manual MCP setup per developerManual MCP setup per developer is the phase where MCP servers exist and work, but each developer is responsible for installing and configuring them independently.guide→
- Basic tool authorization; the stateless MCP core is shipping, so servers drop sticky sessions and run serverless - and Roots, Sampling and Logging are on a 12-month clockBasic tool authorization is the first deliberate access control layer on MCP tool usage: a version-controlled ruleset saying which agents can call which tools, under what conditions, written down rather than clicked through.guide→
Criteria - what to measure
- 011-3 MCP servers are configured (e.g., Git, Jira, documentation)
- 02MCP setup is documented but configured manually per developer
- 03Agents authenticate before they can use a tool server
- 04MCP server configurations are shared via repository (not local-only)
- 05At least one MCP server provides internal documentation or codebase context
L3 · Stage 03Systematic
Practices - what it looks like3 guides
- MCP platform: centralized server managementA centralized MCP platform moves server configuration, deployment, and credential management from individual developer machines to organization-managed infrastructure.guide→
- Servers for architecture, ownership and SLA data are run as products: versioned, owned and monitoredThe servers exposing architecture, ownership and SLA data are operated as products rather than as scripts: versioned, owned by a named team, and monitored like anything else in production.guide→
- RBAC per MCP tool through an MCP gateway that shows each caller only the tools it may use (Kong MCP bundling); clients registered via Client ID Metadata Documents, tokens bound to one server; lazy tool-loading cuts tokens and live attack surfaceRole-Based Access Control per MCP tool means defining precisely which agents can call which tools, based on the agent's role, the task it's performing and the data it's operating on - enforced through an MCP gateway that shows each caller only the tools it may use, with clients registered via Client ID Metadata Documents and tokens bound to one server.guide→
Criteria - what to measure
- 01Centralized MCP platform manages server provisioning, configuration, and lifecycle
- 02Domain-specific MCP servers exist (Architecture MCP, Ownership MCP, SLA MCP)
- 03RBAC controls which agents can access which MCP tools
- 04MCP server health is monitored with alerting on downtime
- 05New MCP servers go through a standardized review and onboarding process
- 06Each caller sees only the MCP tools it is authorised to use (per-caller tool filtering at the gateway)
L4 · Stage 04GovernedMost teams aim here
Practices - what it looks like3 guides
- The organisation's tool surface reachable through one governed MCP gateway, with access granted centrally by the IdP (Okta Cross App Access / ID-JAG, MCP Enterprise-Managed Authorization) instead of per-user consent sprawl, and a kill switch that revokes an agent's live tokens at the gatewayThe Toolshed model, pioneered by Stripe, consolidates hundreds of distinct tools behind a single MCP endpoint.guide→
- Agent discovery: agent knows what tools are availableAgent discovery is the capability for an agent to dynamically enumerate what tools are available in its current environment and adapt its behavior accordingly.guide→
- MCP governance: every server on a lifecycle from intake to deprecation with a per-server pause switch; tool metadata watched for change at runtime, not only at install (Deadbugz rewrote its tool descriptions after three calls); plugins pinned by full commit SHA, never a branch name (Plugin4Shell); injection resistance tested in the IDE configuration you actually shipMCP servers stop being configuration and become production services: owners, changelogs, versioned APIs, a lifecycle from intake to deprecation with a per-server pause switch, tool metadata watched for change at runtime, plugins pinned by full commit SHA, and audit logs.guide→
Criteria - what to measure
- 01The organisation's tool surface is reachable through a unified, governed MCP gateway rather than per-team wiring
- 02Agent discovery: agents can query available tools and their capabilities at runtime
- 03MCP governance covers lifecycle management, versioning, and audit logging
- 04MCP tool usage analytics track which tools are used, by which agents, how often
- 05MCP server versioning allows rollback to previous versions without downtime
- 06Agent access to tools is granted centrally through the IdP rather than by per-user consent
L5 · Stage 05Self-improving
Practices - what it looks like3 guides
- MCP as nervous system: bidirectional context flowAt L5, MCP is no longer just a protocol for giving agents access to tools - it is the real-time information backbone that connects every part of the software delivery system.guide→
- Production → MCP → Agent → Code → Deploy → ProductionThe closed loop: production detects a condition, MCP carries it to an agent, the agent changes the code, CI/CD ships it and production verifies the effect.guide→
- Agent-to-Agent Protocol (A2A) + MCP combinedMCP connects agents to tools; A2A connects agents to each other. Together they are the infrastructure layer a multi-agent system runs on.guide→
Criteria - what to measure
- 01MCP operates as a bidirectional nervous system: production data flows to agents, agent actions flow to production
- 02Full production loop: Production -> MCP -> Agent -> Code -> Deploy -> Production
- 03Agent-to-Agent Protocol (A2A) and MCP are combined for multi-agent coordination
- 04MCP latency for context delivery is under 500ms P95
- 05A2A protocol enables agents to discover and delegate to other agents without human configuration
From the Field
Recent releases, projects and discussions the AI Radar classified into this capability.
release
mcp-use/mcp-use
discovered
An evolutionary multi-agent collaboration platform based on AgentScope Java.
release
cloudflare/mcp-server-cloudflare
discovered
A multi-agent travel assistant system built using the **Agent2Agent (A2A)** protocol and **MCP (Model Context Protocol)**
discovered
AI-powered travel hacking with points, miles, and award flights. Drop-in skills and MCP servers for OpenCode and Claude Code.
discovered
Howardzhangdqs/mcp-wechat-server
release
Skyvern-AI/skyvern
release
sourcebot-dev/sourcebot
Which level is your team at in MCP & Tool Integration?
The criteria above are what we check in a facilitated assessment. Run it yourself to place this capability, see which gates you have passed, and compare it against the other 3 in Infrastructure.
Start the assessment →