Privacy

Local-first

AgentKai runs on your computer. The agent loop, tools, memory files, scheduler database, dashboard, and messaging gateway are all local processes reading and writing ~/.agentkai/. There is no AgentKai cloud, no account, no telemetry, no analytics.

What leaves the machine

Exactly what you choose to send, and nothing else:

• Your model provider. Hosted models (Claude, Gemini, GPT, GLM) receive your prompts — including memory context — as API requests. That's the service working as intended. With --model local (Ollama), nothing leaves the box.
• Skill APIs you configured. Gmail/Calendar/GitHub/Spotify tools call those services' APIs with your tokens, because you asked them to.
• Nothing else. No crash reports, no usage stats, no update checks.

How keys are handled

• Keys come from environment variables (recommended) or ~/.agentkai/*_token.json files. Environment always wins.
• Keys are never printed, never logged. agentkai models shows only set/missing. Error messages carry status codes, not key material. Dashboard redaction uses the same rule.
• Device pairing tokens are stored as SHA-256 hashes, not the tokens themselves.
• Never commit ~/.agentkai/ to git. Skill setup docs say this per skill; it bears repeating: the directory holds your keys.

Memory and logs

Memory is markdown on your disk — back it up like any personal files. Event logs (~/.agentkai/runs/*/events.jsonl) record full tool inputs and outputs; they may contain sensitive text you worked with. They're yours, local, and deletable: remove a run's directory and it's gone.

Boundaries the agent holds

• Public actions (sending email, posting, publishing, charging money) always go through the permission gate — the agent asks first.
• It never invents recipients, never handles raw card numbers, never identifies people from images.
• Unattended scheduler jobs can't run high-risk tools at all.