Skip to main content
The Kelviq Agent Toolkit packages everything an AI coding agent needs to work with Kelviq — the full MCP server plus five Agent Skills that teach the agent Kelviq’s workflows — behind a single install. Skills are the missing half of tool access. The MCP server gives your agent 45 typed tools, but tools alone don’t tell it that a new plan sits in DRAFT until published, that checkout is the default way to charge a customer, or that a retried usage report can double-count. The skills carry exactly that working knowledge, and they load automatically whenever a task looks Kelviq-shaped.
The Agent Toolkit is new. The skills and their names are stable, but content will keep improving alongside the platform. The bundled MCP server is always the latest published @kelviq/mcp-server — no toolkit update needed when the server improves.

​
Install — Claude Code

Two commands, in any Claude Code session:
This installs the full bundle: the MCP server is configured for you (no claude mcp add, no -- separator gotchas) and all five skills are available immediately. The server reads your API keys from environment variables. Set them in your shell profile before starting Claude Code:
Get keys at app.kelviq.com/settings/api-keys. Verify: run /mcp — you should see kelviq ✓ Connected, 45 tools. Then ask something Kelviq-shaped (“set up a Pro plan for my product”) and watch the relevant skill activate.
No keys yet? Install anyway. The docs tools (docs_search, docs_read) work without any credentials, so your agent can answer Kelviq questions from the documentation immediately. API tools return a clear Configuration error until a key is set.

​
Install — any other agent

The five skills are also published through the open skills ecosystem, installable into 70+ agents (Cursor, Codex, GitHub Copilot, OpenCode, Zed, and more):
Add --skill <name> to install a single skill instead of the full set:
This path installs skills only — it doesn’t configure an MCP server for you. Each skill starts with a short Setup section showing the .mcp.json snippet to connect @kelviq/mcp-server on your agent; see the MCP server guide for client-specific instructions.

​
The five skills

​
kelviq — the hub

Loads for any Kelviq task that doesn’t match a more specific skill. It carries the knowledge every other skill assumes:
  • Identifiers vs UUIDs — when an endpoint wants the URL-safe identifier and when it wants the UUID, and how to resolve one from the other.
  • Feature types — BOOLEAN (flag), CUSTOMIZABLE (numeric quantity), METER (usage-tracked), and how entitlement shapes differ per type.
  • Plan lifecycle — plans are created in DRAFT and invisible until published; updating a published plan creates a new draft.
  • Environments — KELVIQ_ENV=sandbox for safe testing, and why “not found” errors often mean “right key, wrong environment”.
  • A customer-access diagnosis flow — comparing what a customer has (entitlements_get), what they bought (subscription_list), and what their plan grants (plan_entitlements_list).

​
pricing-plans

The golden path from empty account to published plan: create the product, create the features, create the plan (in DRAFT), attach entitlements, set prices, publish, verify. Includes the two decisions agents most often get wrong: whether to migrate existing customers on publish (updateFeatures / updatePricing — always a deliberate, confirmed choice), and where prices are managed (the dashboard or the CLI — the MCP server deliberately has no price-write tool).

​
setup-payments

Checkout-first selling. The skill teaches that checkout_create_session is the default way to charge a customer — direct subscription creation charges off-session and is only enabled for some organizations — plus trial handling (trialEnd), safe plan changes (paymentBehavior), cancellation types, and the customer portal (always the signed URL from portal_session_create; a bare portal URL doesn’t authenticate).

​
usage-based-billing

Runtime enforcement. Entitlement checks via the edge-served entitlements_get (built to be cheap enough to call on the request path), the difference between SET and DELTA usage reporting — and why a blindly retried DELTA double-counts — plus license key activation, validation, and deactivation with instanceId handling.

​
pricing-as-code

Wraps the Kelviq CLI loop for agents: kelviq pull → edit kelviq.config.ts in a branch → kelviq push --dry-run → review → apply → kelviq promote from sandbox to production. The skill has a built-in safety rule: an agent may run dry-runs and show you diffs freely, but applying a real change always requires your explicit approval — it will never chain a dry-run into an unattended --yes.

​
Skills, prompts, and tools — how they fit

The toolkit layers three things, each doing what it’s best at:
  1. Tools (from the MCP server) do the work — 45 typed API calls, each carrying safety annotations so clients can auto-approve reads and confirm destructive writes.
  2. Skills (from this toolkit) supply the judgment — which tool, in what order, with which confirmations. They activate automatically based on what you ask.
  3. Prompts (from the MCP server) are guided, step-by-step flows you invoke explicitly — setup_product, launch_plan, and diagnose_customer. The skills route to them rather than duplicating them.
Deep reference detail (field lists, price shapes, config schemas) intentionally lives in none of these — the skills direct the agent to docs_search / docs_read, so answers always come from the current documentation.

​
Testing safely with sandbox

Set KELVIQ_ENV=sandbox (with sandbox API keys) and every tool call targets your sandbox environment — separate data, separate keys, nothing crosses to production. Leave it unset for production. See the sandbox guide for details.

​
Updating

  • MCP server: nothing to do. The toolkit runs npx -y @kelviq/mcp-server, which always resolves to the latest published version.
  • Skills (Claude Code): /plugin marketplace update kelviq pulls the latest marketplace content.
  • Skills (other agents): re-run npx skills add kelviq/agent-toolkit.

​
Troubleshooting

kelviq: tools aren’t available — on Claude Code, check /mcp; if the server isn’t listed, reinstall the plugin. On other agents, the skills-only install doesn’t include the server — follow the Setup section in any installed skill or the MCP server guide. Configuration error from API tools — KELVIQ_SERVER_API_KEY isn’t visible to the process that launched your agent. Set it in your shell profile (not just the current terminal) and restart the agent. “Not found” on an identifier you can see in the dashboard — your key and KELVIQ_ENV are probably pointing at the other environment. Sandbox and production are fully separate; a key only works against its own environment.