Ai MCP Server

Shippo AI

This repo is the one-stop shop for building AI-powered shipping integrations with Shippo.

It contains:

• 9 Agent Skills: Workflow knowledge for AI assistants covering rate shopping, address validation, label purchase (with customs), package tracking, batch shipping, shipping cost analysis, support-ticket drafting, integration best practices, and SDK/API upgrades. Authored once and distributed across multiple AI surfaces.
• Claude Code Plugin (providers/claude/plugin/): Install via --plugin-dir or the plugin marketplace (/plugin marketplace add goshippo/ai).
• OpenAI Codex Plugin (providers/codex/plugin/): Install via the Codex plugin marketplace; bundles the skills plus the OAuth MCP server.
• ClawHub Skill (providers/clawhub/skills/shippo/): Install via openclaw skills install @shippo/shippo.
• Claude apps (claude.ai / Desktop / Cowork): The whole plugin is packaged as a single upload-ready ZIP (shippo-plugin.zip), attached to every GitHub Release. One upload provisions all the skills.

What is a skill?

A skill is a folder containing a SKILL.md file, YAML frontmatter (at minimum: name and description) plus markdown instructions that tell an AI assistant how to perform a specific task. Skills can also bundle reference docs, scripts, and templates.

rate-shopping/
├── SKILL.md           # required: metadata + instructions
└── README.md          # optional: human-facing orientation

Agents load skills by progressive disclosure in three stages:

1. Discovery: at startup, the agent loads only each skill's name and description, just enough to know when it might be relevant.
2. Activation: when a user prompt matches a skill's description, the agent loads the full SKILL.md body into context.
3. Execution: the agent follows the instructions, optionally loading referenced files (shippo/references/*.md) as it works.

Agent Skills is an open standard originally developed by Anthropic. The same SKILL.md works in Claude Code, Cursor, OpenAI Codex, GitHub Copilot, VS Code, and 30+ other agents.

In this repo, the 9 skills under skills/ are the canonical source. They're propagated into providers/claude/plugin/skills/ and providers/codex/plugin/skills/ (1:1 mirrors) and providers/clawhub/skills/shippo/ (consolidated digest) automatically by the sync scripts.

Model Context Protocol (MCP)

Shippo hosts a remote MCP server with per-user OAuth. Each user authorizes once through Shippo, there is no API key to copy. The Claude Code and OpenAI Codex plugins point at this endpoint and trigger the sign-in on first use.

For per-tool semantics and usage, see the Shippo MCP server docs.

Building on OpenAI? See Using the Shippo MCP from the OpenAI Responses API / Agents SDK for the developer config (no submission required).

Capabilities

The 9 skills in this repo are organized by mode of engagement: what the user is doing, not by product surface. The AI assistant matches the user's intent to one of three modes, then loads the right skill.

Decide, "where do I start?"

Do, "execute this workflow"

Maintain, "upgrade or migrate"

A user who already knows the workflow they need ("buy a label", "track this package") jumps straight to a Do skill. A user starting fresh ("I'm building a checkout flow with shipping, where do I start?") hits the Decide skill, which routes them to the right Do skill. Maintenance gets its own skill so production-readiness questions don't compete with workflow content.

The 9 skills lean on 11 shared reference docs under skills/shippo/references/ (carriers, customs, CSV format, error reference, etc.). Skills load references on demand, the AI doesn't pull all 11 into context, just the ones a given workflow needs.

Install

Claude Code

git clone https://github.com/goshippo/ai.git
claude --plugin-dir ./ai/providers/claude/plugin

Or install from the plugin marketplace:

/plugin marketplace add goshippo/ai
/plugin install shippo

On first use, run /mcp, select the Shippo server, and sign in to authorize the MCP over OAuth (no API key to copy).

Skills are namespaced under /shippo:: invoke directly with /shippo:rate-shopping, /shippo:label-purchase, /shippo:tracking, etc., or just describe what you're doing in natural language.

OpenAI Codex

Codex installs the Shippo plugin (skills + OAuth MCP) from this repo's plugin marketplace:

codex plugin marketplace add goshippo/ai
codex plugin add shippo@shippo   # install the "shippo" plugin
codex mcp login shippo           # authorize the remote MCP over OAuth

See providers/codex/plugin/ for details. (To pull just the skill content without the plugin, Codex's skill-installer can also install a single providers/codex/plugin/skills/<name> directory.)

ClawHub

openclaw skills install @shippo/shippo

(Published as @shippo/shippo on the ClawHub registry.)

Claude apps (claude.ai / Desktop / Cowork)

The Claude apps load the plugin as a single ZIP. shippo-plugin.zip (the whole plugin: manifest, OAuth MCP config, and all skills) is attached to every GitHub Release. Download it and add it via the app's Plugins UI. A Team/Enterprise admin can provision it org-wide in one step: Organization settings → Plugins → upload shippo-plugin.zip → set "Installed by default" (or assign to a group), and all skills become available to members. (Code execution must be enabled in Organization settings.)

To build the ZIP locally: npm run build:app-plugin (output in dist/app-plugin/).

Shippo account

You'll need a Shippo account. Getting rates and validating addresses incur no charge; purchasing a label uses Shippo's discounted carrier rates and charges your account. The Claude Code and Codex plugins authorize per-user via OAuth on first use, so there's no API key to copy.

How it works

This plugin bundles two things, with a deliberate division of labor between them:

• Skills (this repo): Cross-tool workflow narrative: routing decisions (checkout vs single label vs batch), UX gates ("ask before purchasing a live-mode label"), CSV ingestion, validation sequencing, test/live mode discipline, response-handling rules. Loaded on activation when the user's request matches a skill's description.
• MCP server (docs): Per-tool semantics: tool name, parameters, return shape, single-call constraints. Each tool description is terse, one verb phrase, one tool. Workflow guidance is intentionally NOT duplicated here.

The skills teach the assistant how to ship across multiple API calls. The MCP server gives the assistant the per-call truth about each tool. The two surfaces are disjoint by design, same precedent Stripe uses (terse mcp.stripe.com tool descriptions, rich stripe/agents skills): so raw MCP users get accurate per-tool semantics and skill-installed users additionally get the workflow narrative, without contradiction.

Repo layout

• skills/: canonical skill content (9 skills + 11 shared references). Edit here; everything else flows from here.
• providers/claude/plugin/: Claude Code plugin distribution. 1:1 mirror of canonical via scripts/sync.js.
• providers/codex/plugin/: OpenAI Codex plugin. skills/ is a 1:1 mirror of canonical via scripts/sync.js; .codex-plugin/plugin.json + .mcp.json (hand-authored) carry the manifest and the OAuth MCP wiring. Cataloged from .agents/plugins/marketplace.json at the repo root.
• providers/clawhub/skills/shippo/: ClawHub bundle distribution. The SKILL.md is auto-generated from SKILL.md.template (hand-curated framing) + canonical skill bodies via scripts/compose-clawhub-digest.js. References are auto-synced via scripts/build-clawhub-bundle.js.
• dist/app-plugin/: the single shippo-plugin.zip for the Claude apps, built from providers/claude/plugin/ by scripts/build-app-plugin.js (not committed; produced on demand and on release).
• scripts/: sync, compose, and build helpers.

Authoring

# 1. Edit canonical content
vim skills/<skill-name>/SKILL.md
# (or skills/shippo/references/<name>.md, or providers/clawhub/skills/shippo/SKILL.md.template
#  if you're changing ClawHub-only framing)

# 2. Sync + verify (one command)
npm test

# 3. Commit canonical edits AND synced output together
git add -A && git commit -m "..."

npm test runs all the sync steps (Claude Code + Codex mirrors, ClawHub digest compose, ClawHub references sync) and verifies the result is internally consistent. CI runs the same command. No npm install needed, the repo has no third-party dependencies, just scripts.

Preview your edit

• Claude Code: run claude --plugin-dir ./providers/claude/plugin from the repo root to launch Claude Code with the local plugin loaded. Edits to skills/<name>/SKILL.md are reflected immediately. Skills are namespaced under /shippo: (e.g., /shippo:rate-shopping).
• ClawHub digest: after npm test runs, the rendered output lives at providers/clawhub/skills/shippo/SKILL.md: read it directly to see what ClawHub-installed users will get. There's no local-server preview today.

See CONTRIBUTING.md for the full authoring discipline, including the version-bump rules and the cross-reference phrasing rule for skill content.

License

MIT