Coding agents

Use typeship from Claude Code, Codex, VS Code, and the other agents: the one-line prompt, the runbook, the CLI's agent contract, the hosted MCP server, and the skills.

typeship is built to be driven by an agent as readily as by a person: the CLI prints JSON and stops with structured envelopes, the MCP server exposes every operation as a tool, every docs page has a markdown twin, and one hosted file tells an agent how to set everything up. This page is the map. Nothing on it is agent-only; the same commands work at a keyboard.

The one line

Paste this into your agent:

Read https://typeship.dev/agents.md and set typeship up for this repo.

/agents.md is a runbook written to the agent: detect the situation, install the CLI, pick a path (no account, has a key, needs a key, REST only, connect the MCP server), verify, report. It works with no account: the first path generates a package from any spec anonymously.

Set a machine up in one command

With a key in hand (create one under api keys in the console):

npx -y typeship-ax@latest init --all -k ak_...

init stores the key, installs the typeship skills, writes the hosted MCP server into every agent client found on the machine (with ${TYPESHIP_TOKEN} as an env reference, never the key itself), and adds a typeship block to AGENTS.md (or CLAUDE.md) with the auth rules, the discovery commands, and a compact index of every command. Run it again any time; the block is replaced in place.

What the CLI promises an agent

The full contract is on the CLI platform page, and typeship agent-guide --format json prints it. In short:

JSON on stdout. Every failure is one envelope on stderr: {status, issues: [{code, message}], docs_url, next_steps, detail}; branch on issues[].code.
Agent mode (--mode agent, TYPESHIP_MODE=agent, or no terminal) never prompts and never opens a browser.
Deletes need --force; without it the CLI returns CONFIRMATION_REQUIRED and the exact command.
typeship generate run ... --out sdk/ writes a generated package straight into a directory, no account needed.
typeship auth check, typeship doctor, typeship help --json, typeship docs search <term>.

Per harness

claude mcp add --transport http typeship https://typeship.dev/mcp-oauth   # signs you in; see the MCP page for the key form (CI)
/plugin marketplace add typeship-ax/skills
/plugin install typeship@typeship-skills

Or typeship init --all for both. In a session: "Read https://typeship.dev/agents.md and set typeship up for this repo."

Cursor is left out until Cursor speaks MCP 2026-07-28, the only protocol version typeship's servers serve; typeship mcp install --cursor writes the entry on request with that note.

Skills

The typeship skills are Agent Skills that wrap the CLI: a router (typeship), typeship-cli, typeship-api (REST without the CLI), typeship-spec-prep (make a spec generate well), typeship-mcp-clients, and typeship-ci.

npx skills add typeship-ax/skills

Docs for agents

/llms.txt: every page, one line each. /llms-full.txt: everything in one fetch.
Append .md to any docs URL, or send Accept: text/markdown, for that page as markdown with frontmatter and absolute links.
search_docs and read_docs on the MCP server need no key.
Every page has Copy page, View as Markdown, Open in Claude, and Open in ChatGPT under its title.

Without a key

An agent with a spec and no typeship account can still get a package: typeship generate run and POST /v1/generate work anonymously (the first 25 operations, rate limited per address), and the MCP server's generate_run does the same. The response's limits object says what was held back and where to sign up. Everything else needs a key from the console; the runbook tells the agent to ask for one rather than to guess.