MCP for coding agents

The MCP server speaks Streamable HTTP at https://mcp.flarely.app/mcp. It holds no data; every tool call is a request to the API with your own token, so scopes, plan limits and the audit trail apply to agents exactly as to people.

Connect

Claude Code

claude mcp add --transport http flarely https://mcp.flarely.app/mcp \
  --header "Authorization: Bearer fl_api_..."

Cursor

// .cursor/mcp.json
{
  "mcpServers": {
    "flarely": {
      "url": "https://mcp.flarely.app/mcp",
      "headers": { "Authorization": "Bearer fl_api_..." }
    }
  }
}

Codex CLI

# ~/.codex/config.toml
[mcp_servers.flarely]
url = "https://mcp.flarely.app/mcp"
http_headers = { Authorization = "Bearer fl_api_..." }

Tools

Tool	What it does
`list_projects`, `list_checks`, `get_check`	Find where to put a check; read one with its code and last runs.
`create_check`, `update_check`	Create a check from a bundle of files, or replace the bundle as a new version.
`run_check`	Run it, waiting for the result by default. Pass `env` with `BASE_URL`.
`get_failure_context`	The failing test, step, locator, expected and received, last steps, artifact links.
`get_run`, `get_check_history`	Full run detail; the version history.
`configure_check`	Schedule, default env, name, pause and resume.
`delete_check`	Delete a check and everything it produced. No undo.
`list_incidents`	What is broken right now, with the run to look at.

The loop the server suggests

get_check before editing, so the agent edits the current code.
create_check or update_check, then run_check.
On failure, get_failure_context first, then a targeted edit.
Rerun until it passes. Say what changed in the version message.

Agent runs never open incidents or page anyone, so the agent can iterate freely. A skill file that teaches these conventions is in the repository under skills/flarely.