testapilocally.com

SignOff Studio MCP vs curl and psql in the shell

Coding agents fall back to curl, psql, sqlcmd and mysql because those are always there. They work. This page is about what changes when the agent is verifying something rather than poking at it.

Shell tools SignOff Studio MCP
Install Usually present One MCP server (npx -y @signoff-studio/mcp init)
Output size Whatever the command prints; big JSON and wide tables fill the context Compact: JSON shape plus first items, 50 rows, 200 characters per cell, full result on request
Pass or fail The agent parses output itself Each check returns pass or fail with the actual value
Read-only SQL Only if you set up a read-only user or transaction yourself Enforced by the driver on every db_query / db_assert
Credentials Passwords and tokens in command lines, environment and transcript From the workspace .env or the app's vault; redacted by value in all output
Prompt injection from data Response bodies and rows read like any other text Marked as data, not instructions
Audit Shell history .signoff/agent-audit.jsonl (no bodies or secrets)
Reuse Gone after the session save_request / save_db_check write checks into the repo

Should I block curl and psql then?

No. Keep them for quick looks and for things the MCP server doesn't do. Point the agent at the MCP tools for verification, either in CLAUDE.md / AGENTS.md (init --agents-md writes a short section) or by naming the server in your prompt.

Is psql with a read-only user just as safe?

For reads, a SELECT-only user is the strongest control, and we recommend it for shared databases. The server adds to it: single statements, row limits, redaction, an audit log, and one tool (verify_api_effect) that ties an API call to the database check.

See also: curl vs MCP for API testing · Safe read-only SQL access for AI agents

Updated 2026-11-19 · This page as Markdown · llms.txt