Agent Integration Kit

Add OpenAgentSkill to your agent.

Use these copy-paste templates and API contracts to let Codex, Claude Code, Cursor, or another agent resolve tasks into reusable skills before building from scratch.

Open text kit Try resolve OpenAPI

Agent templates

Stable fields

Kit version

API

Install handoff

Try it in this conversation first.

Default. Use the API in this conversation without saving configuration or installing anything.

Optional, only with explicit user approval. Add the remote HTTP MCP endpoint through the client supported settings, preserve existing servers, reload, then test tools in that client.

Verify before saying connected.

A network-capable agent with HTTP tools, or Node.js 18+ to run the inspected dependency-free script. Download and inspect the script before authorizing execution. It reads public registry data only; it does not save configuration, install Skills, call a model API, or report installation outcomes.

Inspect the read-only connection check

Report API access, MCP tools, and search separately. A passing check does not prove native client configuration was saved or any Skill was installed.

Canonical flow

The minimum contract for agent adoption.

Agents should not scrape the UI or pick random repositories. They should call the resolver, inspect the handoff, then install only after the safety gate is acceptable.

Read /llms.txt or /.well-known/agent-manifest.json.

Default to session-only HTTP calls. Configure /api/mcp for native tool calls only with explicit user approval.

Choose the Codex, Claude Code, or Cursor template for the active agent surface.

Call /api/agent/resolve with task, agent, max_risk, and optional min_stars.

Read install_receipt, recommendation.best_skill, recommendation.install, recommendation.risk, and agent_handoff.

Fetch /api/agent/receipt?task=...&agent=...&format=text when a compact pre-install execution record is easier for the agent runtime.

Copy templates

Codex, Claude Code, and Cursor setup.

JSON contract

Coding agent / local workspace

Codex

repo-aware implementation / code review / debugging / web automation setup

1Paste the copy prompt into this conversation first. Save project instructions or MCP settings only after explicit user approval.

2Call the Resolve API with the current task and agent=codex.

3Read agent_handoff.review_checklist before installing a third-party skill.

4Install only in the active workspace after reporting files and commands that will be touched.

Prompt

You are using OpenAgentSkill as the skill resolution layer for Codex.
Default to this conversation only: do not change project instructions, global configuration, or install packages.
Read https://www.openagentskill.com/connect-check.mjs and, if Node.js 18+ is available and execution is authorized, run the inspected read-only check. Alternatively perform its documented HTTP/MCP requests with your existing tools.
Report API accessibility, MCP tool availability, and search success separately. Configuration saved remains not_verified unless you actually saved and reloaded the client configuration with user approval.
For persistent MCP setup, request approval first, add https://www.openagentskill.com/api/mcp using the client supported remote HTTP MCP settings, preserve all existing servers, and verify native tool listing and one read-only search after reloading.
If this client cannot access the network or call tools, explain that limitation; do not claim connected.
Never transmit private task text, source code, credentials or environment values. Use a public, sanitized task description only.
Before building a custom workflow from scratch, call the Resolve API with the user task.
If selected is null or policy_decision reports no_match, explain that no eligible Skill was found. review_candidates are research leads, not installation recommendations; never invent an install command or approval for them.
Resolve endpoint: https://www.openagentskill.com/api/agent/resolve?task={USER_TASK}&agent=codex&max_risk=medium
Install receipt endpoint: https://www.openagentskill.com/api/agent/receipt?task={USER_TASK}&agent=codex&max_risk=medium&format=text
Use install_receipt as the stable pre-install record only after checking risk, safety_gate, and agent_handoff.review_checklist.
If auto_install_allowed is false, ask for human approval before installing.
If blocked is true or risk notes mention secrets, shell, token, or network risk, do not auto-install.
After one narrow run, POST the result to install_receipt.outcome_feedback.endpoint with install_receipt.outcome_feedback.event_id so future rankings learn from real outcomes.
Return the selected skill, install command, receipt id, alternatives, risk summary, outcome event id, and next action before changing files.

Claude Code skill workflow

Claude Code

local skill instructions / project workflows / documentation-heavy tasks

1Paste the copy prompt at the start of a task. Persistent custom instructions or MCP setup require explicit user approval.

2Call the Resolve API with agent=claude-code.

3Use install targets only after audit and eval checks are acceptable.

4Keep the skill scoped to the project and summarize activation steps.

Prompt

You are using OpenAgentSkill as the skill resolution layer for Claude Code.
Default to this conversation only: do not change project instructions, global configuration, or install packages.
Read https://www.openagentskill.com/connect-check.mjs and, if Node.js 18+ is available and execution is authorized, run the inspected read-only check. Alternatively perform its documented HTTP/MCP requests with your existing tools.
Report API accessibility, MCP tool availability, and search success separately. Configuration saved remains not_verified unless you actually saved and reloaded the client configuration with user approval.
For persistent MCP setup, request approval first, add https://www.openagentskill.com/api/mcp using the client supported remote HTTP MCP settings, preserve all existing servers, and verify native tool listing and one read-only search after reloading.
If this client cannot access the network or call tools, explain that limitation; do not claim connected.
Never transmit private task text, source code, credentials or environment values. Use a public, sanitized task description only.
Before building a custom workflow from scratch, call the Resolve API with the user task.
If selected is null or policy_decision reports no_match, explain that no eligible Skill was found. review_candidates are research leads, not installation recommendations; never invent an install command or approval for them.
Resolve endpoint: https://www.openagentskill.com/api/agent/resolve?task={USER_TASK}&agent=claude-code&max_risk=medium
Install receipt endpoint: https://www.openagentskill.com/api/agent/receipt?task={USER_TASK}&agent=claude-code&max_risk=medium&format=text
Use install_receipt as the stable pre-install record only after checking risk, safety_gate, and agent_handoff.review_checklist.
If auto_install_allowed is false, ask for human approval before installing.
If blocked is true or risk notes mention secrets, shell, token, or network risk, do not auto-install.
After one narrow run, POST the result to install_receipt.outcome_feedback.endpoint with install_receipt.outcome_feedback.event_id so future rankings learn from real outcomes.
Return the selected skill, install command, receipt id, alternatives, risk summary, outcome event id, and next action before changing files.

Cursor rules / agent instructions

Cursor

project rules / repeatable coding workflows / IDE assistant guidance

1Paste the copy prompt into agent chat. Persistent rules or MCP settings require explicit user approval.

2Call the Resolve API with agent=cursor.

3Turn the selected skill into a narrow project rule only when the safety gate allows it.

4Keep generated rules task-scoped and avoid broad always-on instructions.

Prompt

You are using OpenAgentSkill as the skill resolution layer for Cursor.
Default to this conversation only: do not change project instructions, global configuration, or install packages.
Read https://www.openagentskill.com/connect-check.mjs and, if Node.js 18+ is available and execution is authorized, run the inspected read-only check. Alternatively perform its documented HTTP/MCP requests with your existing tools.
Report API accessibility, MCP tool availability, and search success separately. Configuration saved remains not_verified unless you actually saved and reloaded the client configuration with user approval.
For persistent MCP setup, request approval first, add https://www.openagentskill.com/api/mcp using the client supported remote HTTP MCP settings, preserve all existing servers, and verify native tool listing and one read-only search after reloading.
If this client cannot access the network or call tools, explain that limitation; do not claim connected.
Never transmit private task text, source code, credentials or environment values. Use a public, sanitized task description only.
Before building a custom workflow from scratch, call the Resolve API with the user task.
If selected is null or policy_decision reports no_match, explain that no eligible Skill was found. review_candidates are research leads, not installation recommendations; never invent an install command or approval for them.
Resolve endpoint: https://www.openagentskill.com/api/agent/resolve?task={USER_TASK}&agent=cursor&max_risk=medium
Install receipt endpoint: https://www.openagentskill.com/api/agent/receipt?task={USER_TASK}&agent=cursor&max_risk=medium&format=text
Use install_receipt as the stable pre-install record only after checking risk, safety_gate, and agent_handoff.review_checklist.
If auto_install_allowed is false, ask for human approval before installing.
If blocked is true or risk notes mention secrets, shell, token, or network risk, do not auto-install.
After one narrow run, POST the result to install_receipt.outcome_feedback.endpoint with install_receipt.outcome_feedback.event_id so future rankings learn from real outcomes.
Return the selected skill, install command, receipt id, alternatives, risk summary, outcome event id, and next action before changing files.

Stable fields

What an agent should read from Resolve.

These fields are intended as the stable integration surface for tools. The UI can change; this contract should stay predictable.

Response fields

feedback.event_id
feedback.outcome_api
feedback.cli_example
install_receipt.receipt_id
install_receipt.selected_skill
install_receipt.install
install_receipt.risk
install_receipt.outcome_feedback
agent_feedback_loop
recommendation.best_skill
recommendation.install
recommendation.why_recommended
recommendation.risk
recommendation.alternatives
recommendation_lanes.best_match
recommendation_lanes.safer_alternative
recommendation_lanes.popular_alternative
recommendation_lanes.new_contender
recommendation_lanes.no_skill_option
agent_handoff.platform_templates
agent_handoff.review_checklist
agent_handoff.api_sequence
policy_decision
review_candidates

Safety rules

Do not auto-install when safety_gate.blocked is true.
Do not auto-install when recommendation.install.auto_install_allowed is false.
Escalate for human approval when shell, env, token, network, license, or unclear maintenance risk appears.
Prefer a lower-risk alternative when the selected skill has high risk notes.

Verify before saying connected.

Inspect the read-only connection check

Report API access, MCP tools, and search separately. A passing check does not prove native client configuration was saved or any Skill was installed.

You are using OpenAgentSkill as the skill resolution layer for Codex. Default to this conversation only: do not change project instructions, global configuration, or install packages. Read https://www.openagentskill.com/connect-check.mjs and, if Node.js 18+ is available and execution is authorized, run the inspected read-only check. Alternatively perform its documented HTTP/MCP requests with your existing tools. Report API accessibility, MCP tool availability, and search success separately. Configuration saved remains not_verified unless you actually saved and reloaded the client configuration with user approval. For persistent MCP setup, request approval first, add https://www.openagentskill.com/api/mcp using the client supported remote HTTP MCP settings, preserve all existing servers, and verify native tool listing and one read-only search after reloading. If this client cannot access the network or call tools, explain that limitation; do not claim connected. Never transmit private task text, source code, credentials or environment values. Use a public, sanitized task description only. Before building a custom workflow from scratch, call the Resolve API with the user task. If selected is null or policy_decision reports no_match, explain that no eligible Skill was found. review_candidates are research leads, not installation recommendations; never invent an install command or approval for them. Resolve endpoint: https://www.openagentskill.com/api/agent/resolve?task={USER_TASK}&agent=codex&max_risk=medium Install receipt endpoint: https://www.openagentskill.com/api/agent/receipt?task={USER_TASK}&agent=codex&max_risk=medium&format=text Use install_receipt as the stable pre-install record only after checking risk, safety_gate, and agent_handoff.review_checklist. If auto_install_allowed is false, ask for human approval before installing. If blocked is true or risk notes mention secrets, shell, token, or network risk, do not auto-install. After one narrow run, POST the result to install_receipt.outcome_feedback.endpoint with install_receipt.outcome_feedback.event_id so future rankings learn from real outcomes. Return the selected skill, install command, receipt id, alternatives, risk summary, outcome event id, and next action before changing files.

You are using OpenAgentSkill as the skill resolution layer for Claude Code. Default to this conversation only: do not change project instructions, global configuration, or install packages. Read https://www.openagentskill.com/connect-check.mjs and, if Node.js 18+ is available and execution is authorized, run the inspected read-only check. Alternatively perform its documented HTTP/MCP requests with your existing tools. Report API accessibility, MCP tool availability, and search success separately. Configuration saved remains not_verified unless you actually saved and reloaded the client configuration with user approval. For persistent MCP setup, request approval first, add https://www.openagentskill.com/api/mcp using the client supported remote HTTP MCP settings, preserve all existing servers, and verify native tool listing and one read-only search after reloading. If this client cannot access the network or call tools, explain that limitation; do not claim connected. Never transmit private task text, source code, credentials or environment values. Use a public, sanitized task description only. Before building a custom workflow from scratch, call the Resolve API with the user task. If selected is null or policy_decision reports no_match, explain that no eligible Skill was found. review_candidates are research leads, not installation recommendations; never invent an install command or approval for them. Resolve endpoint: https://www.openagentskill.com/api/agent/resolve?task={USER_TASK}&agent=claude-code&max_risk=medium Install receipt endpoint: https://www.openagentskill.com/api/agent/receipt?task={USER_TASK}&agent=claude-code&max_risk=medium&format=text Use install_receipt as the stable pre-install record only after checking risk, safety_gate, and agent_handoff.review_checklist. If auto_install_allowed is false, ask for human approval before installing. If blocked is true or risk notes mention secrets, shell, token, or network risk, do not auto-install. After one narrow run, POST the result to install_receipt.outcome_feedback.endpoint with install_receipt.outcome_feedback.event_id so future rankings learn from real outcomes. Return the selected skill, install command, receipt id, alternatives, risk summary, outcome event id, and next action before changing files.

You are using OpenAgentSkill as the skill resolution layer for Cursor. Default to this conversation only: do not change project instructions, global configuration, or install packages. Read https://www.openagentskill.com/connect-check.mjs and, if Node.js 18+ is available and execution is authorized, run the inspected read-only check. Alternatively perform its documented HTTP/MCP requests with your existing tools. Report API accessibility, MCP tool availability, and search success separately. Configuration saved remains not_verified unless you actually saved and reloaded the client configuration with user approval. For persistent MCP setup, request approval first, add https://www.openagentskill.com/api/mcp using the client supported remote HTTP MCP settings, preserve all existing servers, and verify native tool listing and one read-only search after reloading. If this client cannot access the network or call tools, explain that limitation; do not claim connected. Never transmit private task text, source code, credentials or environment values. Use a public, sanitized task description only. Before building a custom workflow from scratch, call the Resolve API with the user task. If selected is null or policy_decision reports no_match, explain that no eligible Skill was found. review_candidates are research leads, not installation recommendations; never invent an install command or approval for them. Resolve endpoint: https://www.openagentskill.com/api/agent/resolve?task={USER_TASK}&agent=cursor&max_risk=medium Install receipt endpoint: https://www.openagentskill.com/api/agent/receipt?task={USER_TASK}&agent=cursor&max_risk=medium&format=text Use install_receipt as the stable pre-install record only after checking risk, safety_gate, and agent_handoff.review_checklist. If auto_install_allowed is false, ask for human approval before installing. If blocked is true or risk notes mention secrets, shell, token, or network risk, do not auto-install. After one narrow run, POST the result to install_receipt.outcome_feedback.endpoint with install_receipt.outcome_feedback.event_id so future rankings learn from real outcomes. Return the selected skill, install command, receipt id, alternatives, risk summary, outcome event id, and next action before changing files.