An MCP server for coding agents

A second pair of eyes, when your agent is stuck.

TrueHeading-MCP is a server your coding assistant connects to. You describe the problem in one sentence, and it answers with one line: where to look, and what is there. Then you open that file.

What an answer looks like

you:  the retry loop keeps doubling work on the same row, and I cannot see where
back: the retry path in send(), where the second attempt writes before the first resolves

One line, about two seconds. It is a hypothesis: the value arrives when you go and read the place it names. Ten minutes of searching becomes one look, and the line either holds up or it tells you something by not holding up.

What it is good at

• Before you write. Name the design you are about to build and the tension inside it; it points at the part most likely to give way, while changing it is still cheap.
• Before you ship. Describe what you just finished and believed was tested; it finds the seam your own tests walked past.
• When a failure will not explain itself. Give it the shape of the failing path and it names the first place to read.
• On work nobody has tested. A shell script, a deploy step, a hand-kept copy of a config: the things with no suite to catch them.
• Before you call it finished. checkup asks seven questions in one pass and gives seven lines back: the assumptions nobody wrote down, and the parts that were never run together. This is the one to reach for when the thing feels done.

The tools

ask

You are planning and want the weak point named before you build.

review

You just finished something and want it looked at before it ships.

debug

A failure will not explain itself.

nudge

You have drafted an answer; it says where to look at your own draft again. No arguments: the question is built in.

checkup

You are about to call something finished. One pass over several areas at once — assumptions, dependencies, the boundary with someone else's part, what was never run with the rest, the written record, and what got dropped — and seven short labelled lines back. No arguments.

sharp

The same one-pass shape, aimed at the failure modes work that already runs actually produces: exits that skip the clean-up, errors that stopped explaining themselves, identifiers with nothing behind them, values kept in two places. No arguments.

One more, balance, is free: it says how many units are left and what a call costs, and asking it spends nothing.
The first three take your words and answer with one line — and they are one call described for a different moment, not three tools with three rules: any of the three takes any question, so pick by what you are doing rather than by what you think you are allowed to ask. nudge takes nothing: its one question is fixed here, because a question you compose is a question we have published — it answers with one line too. checkup and sharp take nothing and ask seven fixed questions in one go; you get seven labelled lines back, one per question. A line of X means that question found nothing to point at — which is an answer, not a failure — and a line that says no line came back is the one that failed.

In your assistant the tools are listed with the vendor and the product in front of the act — ripplenav_trueheading_ask — because a client shows one flat list of every server's tools, and a tool called plain ask would sit beside its own. Below, and everywhere else on this site, the short form is used: it is the same tool.

How you start

• Create an account — an address and a password, nothing else — then sign in to see your balance and buy credit.
• Create an API key on your account page and paste it into your assistant's MCP configuration.
• Ask. Two or three rounds on one problem is normal: each answer is its own moment, so phrase the next one with what the last look turned up.