MCP tools

MCP tool security starts with the action path.

MCP makes tools easier for agents to reach. That is useful, but it also means engineering teams need to know which tool can call which system, under which token, with what approval, and what proof remains afterward.

Last updated: May 16, 2026

MCP gives agents a standard way to connect to external tools and data sources. The practical security question is not “do we use MCP?” It is whether a coding session can call a tool that writes to GitHub, Jira, Slack, cloud, data stores, incident systems, customer systems, or internal APIs using a standing token.

What to map for each MCP tool

Tool and operation

List the declared tool, callable operations, read/write behavior, and whether the call can mutate external or internal state.

Credential

Identify OAuth grant, service token, PAT, app credential, local token, or delegated user identity.

Target system

Connect the tool call to GitHub, Slack, Jira, cloud APIs, incident tools, data stores, package systems, or internal APIs.

Approval and evidence

Record who approved the tool call, why it was allowed, which policy applied, and where the outcome is logged.

MCP tool risk tiers

Tier Examples Default review
Read-only Docs search, repo lookup, ticket search, read-only logs. Allow with owner and logging.
Write-capable GitHub issue writer, Jira transition, Slack sender, PR commenter. Approve based on target, audience, and reversibility.
Privileged AWS deploy helper, database query tool, customer lookup, incident system action. Require explicit approval, scoped credential, and outcome evidence.

Risks worth reviewing first

  • Tools that can write, delete, deploy, publish, change tickets, notify customers, or call production-adjacent APIs.
  • Token passthrough or broad standing tokens where the tool carries more authority than the task needs.
  • Tools reachable from repo instructions, agent settings, local config, or shared developer environments without clear owner.
  • Missing evidence tying human request, agent session, tool call, credential, target, approval, and outcome together.

Starter boundary

Allow read-only and low-impact tools by default. Require approval for write, deploy, publish, destructive, customer-facing, data-access, cloud, or privileged internal-system calls. Block unmanaged token passthrough, ownerless tools, and tools that bypass existing approval paths.

agent -> MCP tool -> service token -> internal API -> action target -> approval/evidence

Source notes

Map one MCP tool path.

Clyra maps tool reachability, credential scope, owner, approval point, and evidence gaps for selected workflows.

Map one workflow