The shipit CLI

One command surfaces the whole runtime: a coding agent for your repo, a live one-shot runner, an interactive REPL, an OpenAI-compatible server, and catalogs for every role, model, MCP server, and tool.

shipit code "fix the failing test"     🛠 coding agent in your repo
shipit browse "cheapest SFO→JFK?"      🌐 computer use: drive a real browser
shipit run "prompt"                    one-shot with live ⏺/⎿ tool cards
shipit chat                            interactive REPL (bottom-pinned input)
shipit serve                           your agent as an OpenAI-compatible API
shipit roles | models | mcp | tools    catalogs
shipit doctor | version                health / version

Install gives you the shipit entry point: pip install shipit-agent.

Providers & models

Pick a provider with --provider or $SHIPIT_LLM_PROVIDER, a model with --model or $SHIPIT_MODEL. shipit models lists the curated latest catalog with defaults marked:

shipit code — the coding agent

Roots Agent.for_project() in your current directory: project memory (SHIPIT.md/AGENTS.md), .shipit/commands/ slash commands, the checked-in permission policy, and the full 50-tool builtin roster — driven by a senior-engineer playbook (read first → smallest patch → verify with tests → report path:line).

shipit code "add input validation to the signup endpoint"
shipit code --plan "how would you introduce rate limiting?"   # read-only
shipit code --yes "rename UserSvc to UserService"             # auto-accept edits
shipit code --guardrails strict --max-tool-calls 30 "..."     # hardened

Modes: default asks [y]es / [n]o / [a]lways before ask-gated tools; --plan is read-only research; --yes is acceptEdits. The toolbox includes git_ops (structured git — status/diff/log/blame/commit; push and reset disabled by default), notebook_edit (structural .ipynb cells), hardened edit_file (unified diffs + external-change detection), bash, glob_files/grep_files, run_code, deep_research, download_file, build_document, background sub_agent, and more.

Two runtime layers keep small open-weight models productive: self-healing tool calls (calls emitted as text — <tool_call> tags, fenced JSON — are promoted to real executions, declared-tools-only, span-exact) and nudge-on-stall (one capped re-prompt when the model narrates intent without acting).

shipit serve — your agent as an API

shipit serve --provider bedrock --role finance-analyst --api-key secret

Exposes POST /v1/chat/completions (with SSE streaming fed by the agent's live events), GET /v1/models, and GET /health. Any OpenAI SDK works:

from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8399/v1", api_key="secret")
reply = client.chat.completions.create(
    model="shipit", messages=[{"role": "user", "content": "Close Q2."}])

Requests are validated before any agent work runs; auth is optional Bearer (open on loopback by default); stream errors arrive as in-band SSE frames.

shipit browse — computer use

shipit browse --show "find the cheapest direct SFO→JFK flight on May 20"

Drives a real browser with the vision loop: screenshots → reasoning → clicks/typing, streamed as live cards. --show opens a visible Chrome window slowed enough to watch; consent walls and popups are handled by the agent itself, and accepted consent persists across runs via --storage-state. Prefer a vision-capable model.

Attach MCP servers — --mcp

Every agent command accepts comma-separated catalog servers:

shipit code  --mcp playwright "verify the signup flow on staging"
shipit run   --mcp playwright,filesystem "check example.com's title"
shipit serve --mcp playwright --api-key secret

The catalog now includes the official Playwright MCP (@playwright/mcp — accessibility-tree navigation/click/type/screenshot, no API key), so any tool-calling model gets browser control without vision. Missing env vars or launchers fail fast with one clear message.

Human-in-the-loop

ask-gated tools pause for a real decision — in shipit chat and shipit code automatically, and in any library agent via:

from shipit_agent import Agent, console_permission_prompt
from shipit_agent.permissions import PermissionEngine

agent = Agent.with_builtins(
    llm=llm,
    permissions=PermissionEngine(ask=["bash", "write_file", "git_ops"]),
    permission_callback=console_permission_prompt(),
)
# ⏸ allow bash(command='rm dist')? [y]es / [n]o / [a]lways:

[a]lways persists for the session (the allowed-set is shareable across agents); EOF/Ctrl-C denies safely; denials reach the model as readable reasons it can react to.

The UI layer

Every command shares one terminal kit: a named 256-color palette, NO_COLOR/FORCE_COLOR respected before TTY detection, encoding-safe printing that can't crash a launch, and audience-bucketed banners. Live runs render Claude-Code-style ⏺/⎿ tool cards with durations.

shipit chat adds the full Claude-Code layout on real TTYs: chat scrolls in a VT100 scroll region while the input stays pinned to the bottom row (stdlib-only; SHIPIT_NO_TUI=1 opts out, pipes/CI fall back to plain text automatically). Slash commands autocomplete via Tab.

Related

• The super agent — the capabilities behind the CLI
• Observability — Langfuse/LangSmith/OTel
• Streaming — the event model the cards are built on