Terminal agent host

Switch models,
not context.

milk routes each prompt between a cheap primary agent and a deep escalation agent — keeping the full conversation in sync across both — then goes further: tools, MCP, multi-agent workflows, persistent memory, and an eval harness. Start cheap. Go deep when you need it.

Get started Documentation View on GitHub 
curl -fsSL https://raw.githubusercontent.com/scoutme/milk/main/install.sh | sh
milk
ROUTING
Routing Local Escalation Waiting
milk TUI demo
Why milk
01

Automatic routing

Each prompt is classified and sent to the right agent without you changing tools — sticky escalation keeps hard turns on the deep agent.

02

Context handoff

When escalation fires, the primary conversation is reformatted as context — the escalation agent orients itself without a separate setup step.

03

Agent tools & MCP

Built-in bash/file/HTTP tools, MCP servers (stdio or HTTP, OAuth-aware), and agent-as-tool — expose any configured agent as a callable peer.

04

Multi-agent workflows

Background sub-agents for async research, plus a native /workflow engine — dev, pair, swarm, or your own YAML — with typed verdicts and checkpoint/resume.

05

Memory & safety

A Percept store survives across sessions; loop detection warns and can auto-interrupt when an agent gets stuck.

Any backend, either role Streaming TUI milk eval