Skip to content
Manybases
esc

Search pages and sections. Try a tool name, an error code, a route.

↑↓ navigate↵ openesc close
  1. Guides
  2. Start
  3. Start with an agent
Docs

Guides · Start

Start with an agent

The shortest path: an account, a connector, one sentence to your agent, and one visit to the dashboard.

1 · Create your account#

Sign up at dash.manybases.com with an e-mail and a password, confirm the e-mail, sign in. It starts free: two apps, no card.

2 · Connect your agent#

In Claude, add a custom connector with the URL https://mcp.manybases.com. A consent window opens on your dashboard, says what the connector will and will not be able to do, and issues the key for you — nothing to copy. From Claude Code it is the same URL: claude mcp add --transport http manybases https://mcp.manybases.com, then /mcp → Authenticate. A client that cannot reach a remote server can run the local server with your account key instead. Both are described in Connect an agent.

One agent at a time: an account has a single live agent key, and installing the connector revokes the key a local server was using.

3 · Ask for a backend#

create a backend for my habit tracker

Here is what the agent does, with which tools, and what it gets back:

  1. create_project — a project with its own database and file bucket, in about six seconds. It gets the project id.
  2. get_project — the project card, including urls.api, the address of the project's data plane. Every data tool reads that address itself.
  3. create_table — habits and checkins, each with an access preset declared at creation: prive so that each user only sees their own rows, or partage, collaboratif, journal… The eight presets are in Access presets.
  4. insert_row, query_rows, run_sql — to seed and check data. Rows an agent writes have no owner: they are reference data, not a person's.
  5. create_key with kind: "pk" — a publishable key for the front end. Its value is shown once, in the tool result. It signs users up right away: since 2026-09-07 no human step stands between the agent and the first user.
  6. get_logs — to debug, without a browser: every request to the project, its status, its error code and hint.

Then it builds the front end against the SDK (Start from code), deploys it from a repository (Deploy from GitHub), and writes server-side code as edge functions if the app needs any.

4 · The one dashboard visit#

Until 2026-09-07 a publishable key created by an agent was inert — it did everything except let people sign up, and you had to issue the active one yourself. That step is gone. What still needs you: if the app invites people, set its app URL in Settings, or let the first site deployment fill it in.

That is the boundary: the agent builds everything, including the key that lets people sign up. The full list of what needs a human is in The human boundary.

What the agent sees when something is refused#

Every refusal comes with a code and a hint written for the agent to correct itself: which key or session to provide, the available columns, the tool that helps. Guarded gestures answer 428 with a summary and a token to re-send. See Errors for agents.