Use MCP as a governed boundary for coding-agent work.

The coding agent is good at exploring a repository, proposing a plan, editing code, and running commands. It should not also be the only authority deciding its own scope and whether its result is safe.

Kaplira sits beside MCP-capable executors such as Codex and Claude Code. The executor implements; Kaplira supplies the live project contract, grants bounded access, applies policy, and reviews the reported evidence.

• Context selected from current project knowledge
• Scope receipts issued before protected edits
• Execution events linked to semantic plan steps
• Changed-file and code-health review after implementation

A project should not lose its governance model when the team changes coding agents or model providers. Keeping the contract in a separate control layer lets multiple MCP-capable executors work against the same architecture, constraints, and evidence expectations.

Teams can use an API provider, an approved CLI, private infrastructure, or local inference while retaining the same project boundary.

Intent, plan steps, file operations, tests, and review outcomes form one run record. The result can be inspected without reconstructing what happened from chat history, and the next run can reuse what the project accepted or learned.