agentkai

An open-source personal AI agent that runs on your own computer.

agentkai is model-agnostic by design: bring your own model — Claude, Gemini, GPT, a local Ollama model, GLM, or anything LiteLLM speaks — and the agent keeps the same tools, memory, scheduler, and dashboard. Everything runs locally first; your data stays on your machine.

60-second quickstart

pip install agentkai

export ANTHROPIC_API_KEY=...   # or GEMINI_API_KEY / OPENAI_API_KEY / ZHIPUAI_API_KEY
agentkai run "summarize what's in ~/Documents" --model claude
agentkai dashboard             # local UI at http://127.0.0.1:8931/

No key? Install Ollama, ollama pull qwen3:32b, then agentkai run "hello" --model local — fully offline.

⭐ If agentkai helped you, a star means a lot — it helps other developers find the project.

Why agentkai

Cloud assistants rent you a brain. agentkai gives you the whole animal: one agent loop with real tools, a memory that persists in plain markdown, a scheduler that works while you sleep, and a local dashboard — all wired to whichever model you choose, on your own machine.

• Model-agnostic, BYOK. Aliases (claude, gemini, gpt, glm, local) with fallbacks and capability probing. Keys are yours, never logged, never leave your machine except to the provider you chose.
• Real tools, permission-gated. Shell/filesystem (root-scoped, blocklisted), a real Chromium browser, an MCP client (stdio + SSE), and 12 bundled skills: Gmail, Calendar, GitHub, Spotify, places, images, voice, media, payments, shopping, travel, health. Every tool call passes a risk-based allow/ask/deny gate, audited — unattended jobs can never run high-risk tools.
• Markdown memory. SOUL.md / USER.md / MEMORY.md, daily logs, people & groups, keyword search. Plain files you can read, edit, and git — no black-box vector store.
• Scheduler. Cron, one-shot, heartbeat check-ins, and nightly dreaming (memory consolidation), all in SQLite. Manage jobs from the CLI or the dashboard — create, delete, run now.
• Messaging gateway. Telegram, Discord, built-in WebChat, and an experimental WhatsApp bridge. Per-session isolation; owner approvals via APPROVE / DENY.
• Voice & media. Speech-to-text/TTS via faster-whisper or OpenAI; image generation via OpenAI or a local Stable Diffusion server. Honest errors when nothing is configured — nothing faked.
• Devices & goals. Pair your phone via the companion-app contract; durable goals and tracked items with a daily briefing prompt.

CLI reference

Run agentkai <command> --help for flags on any command.

Web dashboard

Prefer a UI? agentkai dashboard launches a local web app: chat with the agent, watch runs stream in with tool calls inline, browse and edit memory, inspect the tool registry, and manage scheduled jobs — create, delete, or trigger a run now.

agentkai dashboard
# → http://127.0.0.1:8931/?token=<fresh-token-every-launch>

It binds only to 127.0.0.1 (never exposed to the network), generates a fresh login token on every launch, and loads zero third-party resources — the entire UI is packaged with agentkai and works fully offline. Same agent, same tools underneath.

Docs

📚 Documentation: docs/

Quickstart · Models · Tools · Browser · Memory · Scheduler · Channels · Skills · Voice · Media · Devices · Goals · Dashboard · Privacy · Troubleshooting

Honest scope note: this is early-stage software (v0.3.0). Core flows — agent runs, tools, memory, scheduler, dashboard, channels — are implemented and tested. Areas marked experimental in the docs are real code with real limits; they are labeled, not hidden.

Development

git clone https://github.com/Bilal-Junaid-Jiwani/agentkai
cd agentkai
python -m venv .venv && .venv/bin/pip install -e ".[test,voice]"
.venv/bin/python -m pytest tests/   # full suite

License

Apache-2.0. See LICENSE.