Gini Agent is a personal agent whose runtime is the system of record. A single Bun process per instance owns state and performs the work (conversations, runs, tasks, approvals, memory, skills, jobs, tools, traces, audit events, and runtime health), and every surface is a client of the same authenticated API contract: a Next.js web app, a CLI, an Expo mobile app, MCP surfaces, and messaging bridges.
Chat is treated as a control surface rather than a transcript. When the agent needs something from you it renders an inline control and waits: a secure field for a key, password, one-time code, or payment detail that flows straight to the gateway without entering the model, the transcript, or the audit trail, a handoff to a live view of the agent's own browser when a step hits a sign-in wall, a short list of choices when more than one path is reasonable, and a confirmation before anything is sent, posted, or bought in your name. The same cards render on the web and iOS apps over one wire protocol and reach a phone off-LAN through a runtime tunnel.
Provider support covers Codex OAuth read from the local codex login, API keys for OpenAI, Azure OpenAI, DeepSeek, and OpenRouter, the first-party Anthropic Claude API, Amazon Bedrock through the model-agnostic Converse API signed with AWS SigV4 (Claude, Nova, Llama, Mistral, and DeepSeek models), and any OpenAI-compatible local server such as oMLX, vLLM, LM Studio, or llama.cpp. Credentials are read from environment variables or the setup form rather than written into configuration. Embeddings, reranking, and voice-message speech-to-text run locally by default from a shared model cache.
The runtime can autostart per user on macOS, and further instances get their own ports and isolated state and logs, so several can run at once. Remote access is available through Gini Relay, Tailscale, ngrok, or Cloudflare tunnels, a Docker deployment runs an instance headless with a real browser under Xvfb, and an import command brings an existing OpenClaw install across in a plan-then-apply pair that archives the source state first.
Features
- Gateway runtime: one Bun process per instance owns state, the agent loop, tools, memory, and jobs behind an authenticated API
- Multiple clients: Next.js web control plane, CLI, iOS app, MCP surfaces, and Telegram and Discord bridges on the same contract
- In-chat controls: secure credential fields, browser sign-in handoff, choice prompts, and confirm-before-send
- Approval-gated tools: file, terminal, and code tools require approval, reviewable from the CLI
- Memory and skills: retain and recall with embeddings and reranking, plus skill learning from task outcomes with a daily review and a human gate
- Provider catalog: Codex OAuth, OpenAI, Azure OpenAI, Anthropic, Amazon Bedrock, OpenRouter, DeepSeek, and any OpenAI-compatible local server
- Local models by default: embeddings, reranking, and speech-to-text run on-device from a shared model cache
- Parallel instances: isolated state, ports, and logs per instance, with an ephemeral smoke instance
- Remote access: Gini Relay, Tailscale, ngrok, and Cloudflare tunnel modes with in-app confirmation
- Migration and deployment: plan-and-apply import from OpenClaw, and a Docker image that runs headless with a browser under Xvfb
