Shiprr

Shiprr is a minimal cargo agent with a smart routing layer designed to cut costs on top of LiteLLM Auto Router.

The visual identity uses a blue cargo ship and container motif. The workspace is dark, sparse, and terminal-native, inspired by the restraint of tools like Ghostty.

cd shiprr
./scripts/install.sh

shipr

shipr opens a full-screen agentic terminal UI. The conversation and work log scroll above a persistent composer at the bottom of the screen. Shiprr operates on the directory where it was started.

┌ SHIPRR · cargo agent ─────────────────────────────────┐
│        ██          S H I P R R                        │
│     ██ ██ ██       ━━━ cargo agent · v0.1.0          │
│     ██ ██ ██                                           │
│    ▀█████████▀                                         │
│      ▀█████▀                                           │
│       ▀  ▀                                             │
│      ● ● ●         ───── underway                    │
│ › fix retry typo in docs                              │
│                                                      │
│ ✦ Processing… Reading context                        │
│                                                      │
│ ● Response streams here token by token               │
├ Ask Shiprr ──────────────────────────────────────────┤
│ › type your next task                                │
└──────────────────────────────────────────────────────┘

While a task runs, one animated Processing… row shows the current model or tool action. Transient activity is not retained in the transcript. Answers stream into the conversation as they are generated, while completed tool calls remain in the feed. Press Esc to cancel or Ctrl+C to exit.

Shiprr sends OpenAI-compatible tool definitions to LiteLLM and repeats the model → tool → result loop until the task is complete. The minimal tool set is:

• list_files — discover the workspace
• read_file — inspect text with line numbers
• search — search with ripgrep
• write_file — create or overwrite a file
• replace_in_file — make one exact targeted edit
• run_command — execute one program directly in the workspace

Reads run immediately. File changes and commands pause on a visible y/n approval prompt. Paths are restricted to the current workspace and .git cannot be edited by file tools.

Responses and tool calls stream from LiteLLM using the OpenAI-compatible /v1/chat/completions endpoint.

• /help
• /clear
• /exit

Run shipr login to replace saved LiteLLM credentials or change the model. Existing configs without a model use auto_router1. SHIPR_MODEL overrides that default. Optional SHIPR_FAST_MODEL, SHIPR_BALANCED_MODEL, and SHIPR_HIGH_MODEL environment variables map harness routing tiers to separate gateway model aliases.

• Minimal by design: just an agentic loop
• Harness-level smart routing chooses cheaper viable policy first
• Manual overrides available in direct run mode (shipr run ... --quality ... --budget ...)
• IDE-like terminal surface with a fixed composer and streaming work feed
• Workspace-scoped tools with explicit approval for writes and commands

• Binary CLI: src/main.rs
• Agent loop: src/agent.rs
• Coding tools: src/tools.rs
• Smart routing crate: crates/shipr-smart-routing/src/lib.rs