Denkeeper
NEW Evals: A/B test a model before you switch →

Your AI agent.
Your rules.
Your hardware.

One binary, one TOML file. Chat with it on Telegram or Discord, approve anything risky with a tap, and keep every token on a budget. Runs happily on a Raspberry Pi.

curl -fsSL https://get.denkeeper.io | sh
Read the docs Homebrew, .deb, .rpm, Docker, Helm →
Talks to you on Telegram · Discord · Web · REST · MCP
Thinks with Anthropic · OpenAI · OpenRouter · Ollama
Runs on Linux · macOS · Raspberry Pi · Docker · Helm
Security first

Every tool call goes through the door.

How approvals work →
  1. 01 · message “Brief me every weekday at 7”
  2. 02 · agent plans schedule_create
  3. 03 · tier check Supervised tier: it writes, so it waits
  4. 04 · the door You tap Approve, or your supervisor agent does
  5. 05 · on record Runs, then lands in the audit log with its cost
Autonomous
walks straight through the door.
Supervised
knocks and waits. The default.
Restricted
only gets read-only tools.
Cost aware

Know what every conversation costs.

Per-session soft and hard limits, live pricing for every provider, and fallback rules that downshift to a cheaper model instead of just stopping.

Cost tracking →
$3.18 this week across 3 agents
Thu: soft limit hit, switched den to a cheaper model for the rest of the session