Coding Agents

How to connect a coding agent to Appstrate (skills and plugin, CLI or MCP), and which configuration tasks you can delegate to it.

If you are the platform or internal-dev engineer rolling out Appstrate inside a company, you can connect a coding agent such as Claude Code or Codex and delegate configuration work to it: authoring agents, wiring integrations, rotating keys, testing end-user impersonation, publishing AFPS packages.

Instead of opening the dashboard, reading the docs and hand-typing curl or appstrate api, you describe what you want ("connect Gmail and create an agent that summarizes new emails every morning") and the coding agent executes it against your active profile.

This page covers when to use each connection path. For the plugin and skills sync (install, targets, flags), see Skills and coding agents.

Three ways to connect

Way	What the coding agent gets	Setup
Skills and plugin	The skills of your spaces, synced into the agent's skill directories with `appstrate code sync`. In Claude Code, the agents of your pinned space also become `/appstrate:run-<agent>` commands.	See Skills and coding agents
CLI	`appstrate api` for any REST call without handling a token, plus commands to run agents and edit packages (`appstrate packages pull`, `push`, `publish`). Uses your named profiles. The repository ships an operating manual for coding agents, `apps/cli/AGENTS.md`.	CLI
MCP	The platform as an MCP server: `search_operations`, `describe_operation` and `invoke_operation` over the whole API, plus `run_and_wait`. No CLI needed, and the key's permissions decide what is visible.	`claude mcp add --transport http ...`, see below

They combine: the skills carry your team's conventions, the CLI or MCP is the hands.

Connect over MCP

Each organization has one MCP endpoint, /api/mcp/o/<orgId>. With an API key that holds mcp:read and mcp:invoke:

claude mcp add --transport http appstrate-acme https://appstrate.yourcompany.com/api/mcp/o/<orgId> \
  --header "Authorization: Bearer apst_your_key"

Clients that support browser OAuth can skip the key and log in instead. The dashboard shows the exact command for each organization. Details: Connecting MCP clients.

Why this pattern

Faster config loops. Authoring an agent, connecting its integrations and adding a webhook normally takes many clicks or precise API calls. A coding agent can do it from one prompt.
Multi-instance workflows. The CLI uses named profiles (dev, staging, prod). The agent can diff a configuration between them or replay a production change on staging first.
End-user impersonation checks. Ask the agent to validate that a flow works for three end-users. With an API key, it loops through them with the Appstrate-User header.
Less context switching. Ship an internal feature that uses Appstrate from the same chat where you wrote the code.
AFPS packaging. The agent can write, import and publish AFPS packages through the CLI, including custom integrations.

Use cases that land well

"Create an agent called email-triage that reads my Gmail inbox and posts a summary to Slack. Use the @appstrate/gmail-mcp and @appstrate/slack integrations and select only the tools it needs."
"Rotate the API key named ci-deploy: create the new one, check it works, then revoke the old one."
"Write a custom integration for our internal API at https://api.acme.com, with an API key in the Authorization header, import it, and add it to the reporting agent."
"Pack my local ./my-skill folder as an AFPS and publish it under @acme/my-skill."
"List every agent whose runs failed more than 10% of the time this week and show the last failed run's logs."
"Diff my dev profile's agent list with prod and tell me what is missing on prod."

The agent picks the right call, runs it and reports back. If something fails, it should stop and show the error instead of retrying blindly. Review what it is about to do on a production profile, as you would a colleague's change.

Reference

Skills and coding agents: the Claude Code plugin and appstrate code sync.
CLI reference: the commands a coding agent drives.
Skills as a feature: how Appstrate itself consumes Agent Skills.
Design choices: why runs are isolated, multi-tenant and API-first.