Day 4 of MLH Global Hack Week: Builders Week. Yesterday I put Build Buddy's benchmark results in a Tiger Data hypertable. Today I connected an AI agent to that database through MCP (Model Context Protocol), and the most useful thing I did was lock it down first.
Read-only before anything else
Tiger's MCP server gives an agent real power. Its tools include listing services, reading logs and metrics, and also deleting services and resetting passwords. Before connecting any agent:
tiger config set read_only all
With read_only all, the MCP tools that write (delete, reset password, create, resize) aren't even registered: the agent can't see them, so it can't call them. Database sessions also open in Tiger Cloud's read-only mode, so the server rejects any writes or schema changes (per the Tiger CLI docs). The CLI also keeps the database password in the Windows credential store (password_storage: keyring), not in a plain-text file.
What the agent did
- Listed my services. Easy.
-
Tried to query the
build_resultstable and couldn't. My network blocks the database port, so direct SQL timed out. - Diagnosed the failure itself. It read the service's live logs through Tiger's API and saw that no client connection ever arrived.
- Was honest about it. It said plainly that it couldn't query the table live and was using my local results file instead.
It also checked my work
Analysing that file, the agent:
- caught a mistake in my Day 3 post: the two warm CS0029 runs were 7.7 s and 7.6 s, not 7.6 s twice
- found something I'd missed: with cold starts removed, CS0161 is the slowest error even when the model is warm (10.4–12.9 s every run)
What I'm taking away
- Lock permissions down before you connect an agent. Read-only first, widen later only if needed.
- The agent working around a blocked connection, and telling me so, is what I want. An agent that quietly made up "live" numbers would be far worse.
- Check the agent, too. It labelled Alberta as Central Time. Small, but it's the same lesson as every post this week.
Next
Builders Week ends Thursday. After that: behaviour checks in Build Buddy (issue #1), so a fix only counts if it builds and behaves the same.
Update (Nov 11): done. Live rerun: 6/6 compiled, 5/6 behaved (issue #1).
Top comments (2)
Read-only first is the right order of operations, and it is the part most people skip when they are excited to connect an agent. The moment that stuck with me was the agent diagnosing its own failure — reading the live logs, seeing no connection ever arrived, and saying plainly it was using the local file instead of inventing live numbers. That is the behaviour you actually want from an agent with database access. It catching the Day 3 numbers mistake (7.7 and 7.6, not 7.6 twice) is a nice bonus — it shows the checking works both ways. Also, the Alberta labelled Central Time slip is a good reminder that even small factual slips are worth noting in a build log. Looking forward to the behaviour checks in the next post.
Thanks! The honest "I couldn't connect, so I used your local file" was my favourite part too. An agent that quietly made up live numbers would have been far worse than one that failed out loud.
The behaviour checks are already in: a live rerun gave 6/6 compiled, 5/6 behaved. The CS0161 fix compiles but prints 3 lines instead of 1, and Gemma gives that same answer every run. Details: github.com/GurminderKaur/build-bud...