An AI agent that treats chat as a side effect, not the product

What it does

Gini is a personal AI agent that runs as a local Bun gateway. A single process per instance owns all state—conversations, tasks, tool approvals, memory, skills, and audit traces—while Next.js, CLI, mobile, and MCP clients merely talk to its authenticated API. It is designed to keep working across reboots without forcing you to tail a log file to know what happened.

The interesting bit

The architecture makes the Bun runtime the gateway and treats the Next.js UI as one of several authenticated clients, not the center of the system. Gini also implements a skill-learning loop with two-tier rewards, attribution, and a daily human-gated review, treating skill improvement as a first-class runtime concern.

Key highlights

• Local-first by default: embeddings, reranking, and speech-to-text run locally, and it supports a wide provider menu from OpenAI to Amazon Bedrock to local OpenAI-compatible servers.
• Approval-gated tools: file, terminal, and code execution wait for human sign-off.
• Parallel isolated instances: each agent gets its own state, ports, and logs, so smoke tests don’t collide.
• Built-in migration path from the openclaw project.
• Self-improving skills via a two-tier reward system and daily review with human approval.

Caveats

• Messaging bridges for Telegram and Discord exist but are explicitly not actively developed; the team recommends the web and iOS apps instead.
• On macOS 26 (Tahoe), a SIGKILL can leave the LaunchAgent in a pended nondemand spawn state where launchd refuses to auto-respawn; the docs note a manual workaround but no permanent fix.

Verdict

Developers who want a local, stateful agent with real architecture docs and gated tool use should look here; if you just need a quick ChatGPT wrapper, this is over-engineered and opinionated.