Using OverSkill from AI agents (Claude, Cursor, Codex)

OverSkill is built for AI agents to use. Here's how to wire up Claude, Cursor, Codex, or your own agent.

OverSkill was designed with AI agents as first-class users. Claude, Cursor, Codex, and other agents can build apps on OverSkill directly — no humans in the loop required for most tasks.

Why this matters

You can ask Claude or another agent to:

• Build a new app on your OverSkill account
• Iterate on an existing app based on user feedback
• Generate templates for sale on the marketplace
• Maintain a portfolio of apps for clients

The agent talks to OverSkill the same way you would — describing what to build in plain English — just programmatically, through OverSkill's agent API and MCP tools.

Connecting your agent

OverSkill ships with an MCP server (Model Context Protocol) described at https://www.overskill.com/.well-known/mcp/server-card.json. Most modern AI agents support MCP natively. To wire it up:

1. In your agent's config, add OverSkill as an MCP server
2. Use the URL above (or https://www.overskill.com/SKILL.md for skill-style agents)
3. Authenticate with an API key from your OverSkill account (Dashboard → Team Settings → API Keys)
4. The agent can now create, edit, and publish apps on your behalf

Tip: use the www.overskill.com host in agent configs — the bare overskill.com address redirects there, and some agent HTTP clients don't follow redirects.

No account yet? Agents can build a first app with no API key at all via the build now, claim later flow (POST /api/v1/anonymous_builds) — the response includes a claim link the person can use to take ownership within 48 hours. Details are in SKILL.md.

Agent-friendly surfaces

Beyond MCP, OverSkill exposes everything agents need:

• Every help article has a .md URL (e.g. /help/articles/what-is-overskill.md) for clean markdown
• /llms.txt — short site map for LLMs
• /llms-full.txt — full API reference
• /SKILL.md — primary skill manifest (the complete build/deploy/manage guide)
• /agents — brand-pillar page with discovery details

Built for autonomous self-service

The API itself is designed so an agent never dead-ends:

• GET /api/v1/capabilities returns what the account can and cannot do right now. Every cannot entry carries the reason, the required_tier, and a signed unlock_url — a one-click upgrade link the agent can hand to its human operator instead of failing.
• next_steps[] on every response — success and error — tells the agent what to do next in plain language plus concrete {action, url, method} entries. An agent that just follows next_steps can navigate the whole platform without docs.
• notify_operator — when a task needs a human (for example, DNS records for a custom domain), endpoints like POST /api/v1/managed_apps/{id}/custom_domains/{id}/notify_operator generate the escalation for the agent so its operator gets a direct, actionable link.
• Errors are machine-readable — see OverSkill API errors and how to handle them. Notably, app.generation.blocked (out of credits) is distinct from failed: agents should surface the block + top-up link, never retry-loop.

The practical agent loop: check /capabilities → act → follow next_steps → on a cannot or credit block, escalate to the operator with the provided unlock_url.

What agents can and can't do

With proper authentication, agents can do anything you can do — including spend credits, publish apps, and process payments. So the credentials matter. Use scoped tokens where possible.

Agents CANNOT:

• Bypass the credit meter (they pay just like you do)
• Skip identity verification for payouts
• Take over an account they aren't authorized for

Pricing for agent-driven usage

Same as humans. Credits are credits. There's no separate agent tier — your agents draw from the same balance as your manual builds.

What to read next

• What are credits?
• Working with multiple apps