Integrations
Every way to connect an agent: Claude Code, Codex, Cursor, VS Code, the Claude app, CI, LangChain, the OpenAI Agents SDK, n8n and Zapier.
Every integration ends in the same place: a request in your inbox, a person’s yes or no, and an entry on the record. What differs is how the agent comes to ask. There are three ways, and they can be combined.
- MCP tools
- The agent gets a request_approval tool and decides when to call it, guided by its instructions or a skill. Quickest to set up; relies on the model choosing to ask.
- Hooks
- Claude Code and Codex run a hook before every matching tool call, such as every shell command. The model cannot skip it.
- Code
- Your own tool calls the HTTP API, or the TypeScript or Python client, before it acts. The action cannot happen without a yes. Best for agents that run with nobody watching.
Use one agent key per agent. Each can then be revoked or paused on its own, and rules can treat each differently. Create keys under Agents in your workspace.
Choose your agent
- Claude CodeMake Claude Code ask a person before it runs a shell command or calls a tool: the MCP server, the skills, and a PreToolUse hook.
- CodexMake OpenAI Codex ask a person before it acts, through the MCP server or a PreToolUse hook that approves shell commands.
- CursorGive Cursor's agent an approval tool through MCP, with the key kept out of the configuration file.
- VS CodeConnect GitHub Copilot's agent mode in VS Code to approvals through .vscode/mcp.json, with the key asked for once.
- Claude appAdd the approvals server to the Claude app as a custom connector, so Claude asks before acting on your behalf.
- GitHub Actions and CIHold a deploy or any CI step until a person approves it, from GitHub Actions or any pipeline that runs a shell.
- LangChainWrap a LangChain or LangGraph tool so it runs only after a person approves, using the HTTP API.
- OpenAI Agents SDKMake an OpenAI Agents SDK function tool ask a person first, using the HTTP API.
- n8n and ZapierPut a person's approval in the middle of an n8n or Zapier workflow with plain HTTP steps.
Any other agent
Anything that can make an HTTP call can ask. Try it from a terminal with a key in place of gl_YOUR_KEY; the request appears in your inbox within a few seconds.
curl -X POST https://api.variantgate.com/v1/requests \
-H 'authorization: Bearer gl_YOUR_KEY' \
-H 'content-type: application/json' \
-d '{"agent":"my-agent","action":"email.send","summary":"Send a test email to myself"}'The API reference lists every endpoint, and Requests explains each field.
Check that it works
Ask the agent to do something that needs approval, such as sending a test email to yourself. A request appears in your inbox. Approve it and the agent carries on; deny it once, on purpose, and check that it stops.
Decide from chat
Waiting requests can be announced in Slack, Microsoft Teams or Discord, with Approve and Deny buttons, and emailed to everyone who can decide. Notifications shows how to connect each one.
The skills
asking-for-approval teaches an agent when to ask, how to write a summary a person can decide on, and to act only on an explicit yes. adding-approvals-to-code helps a coding agent put approval checks into an agent you are building. Both are plain SKILL.md files you can read before installing: asking-for-approval, adding-approvals-to-code. Claude Code reads skills from ~/.claude/skills, Codex from ~/.agents/skills.