Using mybase with AI agents

mybase gives agents four ways in. Use as many as the agent supports.

1. MCP serverCopy Anchor Link

Streamable HTTP at https://mybase.example.com/mcp, authenticated with an API key. The tool list is filtered to what the key may do.

mybase token create agent --project shop --perm database:read,storage:read,queue:write,cron:write
export MYBASE_API_KEY=mb_…
claude mcp add --transport http mybase https://mybase.example.com/mcp --header "Authorization: Bearer $MYBASE_API_KEY"

Or in .mcp.json (Cursor: .cursor/mcp.json); the key stays in the environment:

{
  "mcpServers": {
    "mybase": { "type": "http", "url": "https://mybase.example.com/mcp", "headers": { "Authorization": "Bearer ${MYBASE_API_KEY}" } }
  }
}

Tools: search_docs, read_doc, whoami, list_projects, get_project_context, create_project, enable_service, describe_schema, query_rows, run_sql (admin), list_files, list_cron_jobs, create_cron_job, run_cron_job, list_cron_runs, list_queues, create_queue, send_message, peek_messages, cache_command, list_ratelimit_rules, create_ratelimit_rule, analytics_stats, list_secret_envs, list_secret_keys, set_secrets, create_secret_env (names only; no tool returns a secret value). Docs pages are MCP resources (mybase://docs/<slug>), and prompts setup_project, add_queue_worker, add_cron_job, add_analytics, use_secrets walk through common tasks.

Deleting projects, turning services off, purging or deleting queues, and managing keys are not exposed over MCP. Those stay with the dashboard and CLI.

2. Docs for agentsCopy Anchor Link

• https://mybase.example.com/llms.txt: index with a short "how to act" primer (llmstxt.org format)
• https://mybase.example.com/llms-full.txt: every page in one file
• https://mybase.example.com/docs/<slug>.md: one page as markdown, with this instance's URL filled in
• https://mybase.example.com/docs/search?q=<words>: best matching sections as JSON
• mybase docs, mybase docs <slug>, mybase docs --search "<words>"

3. CLI for scripts and agentsCopy Anchor Link

• --json on any command prints the API's answer as JSON; errors are JSON on stderr: {"error", "status", "exitCode"}.
• Exit codes: 0 ok, 1 failure, 2 usage (including a missing --yes), 3 not signed in / bad key, 4 not found.
• Without a terminal, destructive commands refuse to run unless --yes is given; they never hang on a prompt.
• MYBASE_API_URL + MYBASE_TOKEN configure it without mybase login.
• mybase context <project> prints the project as markdown (or JSON): services, tables and columns, jobs, queues, rules, env var names, docs to read.

4. mybase agent initCopy Anchor Link

Run in an app repo:

mybase agent init --project shop          # add --key to also make a project-limited key in .env.local

It writes (and on re-runs, replaces) a marked mybase section in AGENTS.md (and CLAUDE.md if the repo has one), a Claude Code skill at .claude/skills/mybase/SKILL.md, and a mybase entry in .mcp.json. --key writes the key to .env.local only if git ignores that file.

Agent notesCopy Anchor Link

• Give agents the narrowest key that works. run_sql and the SQL route need admin.
• Secrets in tool results (from create_project, enable_service for the cache) are shown once: save them to the user's env right away.