MCP tools and workflows

Once MCP is set up, your editor's assistant can operate Guidal. This page covers what it can reach, how it finds capabilities, and what it deliberately cannot do.

How the agent discovers capabilities

The server presents a small set of core operations plus a way to search for everything else. The agent starts from the core set and looks up additional capabilities when a task needs them.

This matters for cost and reliability. Presenting every operation at once would consume a large share of the agent's context before it read a line of your code, and a long undifferentiated list makes agents pick wrong. A compact core plus search keeps the common path cheap and the long tail reachable.

You do not need to manage any of this — it is why the agent stays responsive on large tasks.

What agents can do

Inspect — list organizations, projects, services, databases and buckets; read status and configuration.

Deploy — create services, deploy new image tags, scale replicas.

Configure — read and set environment variables, manage domains.

Diagnose — read logs, read metrics, inspect deployment history.

Together these cover the normal loop: build something, ship it, look at what it did, fix it.

What agents cannot do

Destructive operations are not exposed. Deleting a service, database or bucket is not available through MCP. That is a human action via the CLI or the dashboard.

The reasoning is asymmetry of consequences. An agent that misreads "clean up the old test service" and deletes the production database causes damage no amount of apologising undoes. Withholding delete costs a few seconds of human effort in the rare case; granting it risks unrecoverable data loss in the ambiguous case.

Billing is read-only. An agent can report your plan and usage. It cannot change your subscription.

Prompts that work well

Describe the outcome, and give the agent enough to identify the target:

Deploy this repo to Guidal as a web service on port 3000 in a new project called checkout.

The api service is throwing 500s. Read the last 200 log lines and tell me what is failing.

Create a Postgres database for this project and wire its connection string into the api service.

Scale emailer to 3 replicas, then show me its memory usage.

Vague requests produce vague action. "Fix the deployment" gives the agent no way to know which service or what "fixed" means; "the api service fails health checks after the v2 deploy — find out why" gives it a starting point.

Keeping the agent accurate

Point it at the documentation when it needs platform concepts rather than your code:

https://guidal.ai/llms.txt

That index lists every page with a summary and a direct markdown URL, so the agent can fetch only what it needs. See AI access.

When something goes wrong

If the assistant reports a capability it cannot find, confirm the server is connected with guidal mcp status. If it reports authentication failures, run guidal login again. If an operation fails with a plan restriction, the underlying limit applies to agents exactly as it does to you — see Plans, billing and limits.