I Built KIRA: A Local-First AI Agent That Has to Prove Its Work

Description and use

KIRA is a local‑first AI agent that verifies task completion through real tool outputs, offering transparent and permission‑aware interactions on Apple‑silicon Macs.

Problem

AI assistants often claim to have completed tasks without verifiable evidence, making it hard to trust their actions.

Outcome

KIRA demonstrates a reproducible, permission‑aware local AI agent that can be tested on real workflows, providing transparent evidence of task completion.

Workflow

• 1. User states a goal
• 2. Orchestrator V1 selects next tool
• 3. KIRA runs the tool with permission gates
• 4. Tool returns real result
• 5. Orchestrator evaluates evidence
• 6. Repeat until task genuinely complete

Evidence

KIRA should not say a task is complete unless the runtime has evidence from the tool that performed it.

Source

The installer downloads Orchestrator V1 into a local model directory.

Source

KIRA keeps mutating file, command, IDE, MCP, and automation actions behind explicit permission gates.

Source