Where procoder sits

Procoder is often compared against things it is not, so this page states the layer it occupies and what it is not trying to be. It is descriptive, not competitive: the tools named here are good at what they do.

Four layers, not one

A useful split, and one the ecosystem broadly agrees on:

Layer	What it is	Example
Connection	How an agent reaches a system at all	MCP servers
Actions	Individual capabilities an agent invokes	Tools, function calls
Behaviour	What an agent is told to do, in prose	Agent Skills, rule files
Governance	What is computed and refused regardless	Procoder's binary

Procoder spans the bottom two rows, and the distinction between them is the whole point.

Advice versus computation

A skill pack, a rules file, and an AGENTS.md are all advice. They change what an agent is likely to do. They cannot change what happens when the agent does something else, because nothing reads them at the moment of the commit.

Procoder ships both halves deliberately:

• The advisory half — AGENTS.md, skills/procoder/SKILL.md, and eleven generated host rule files. This is ordinary prose guidance, and it works exactly as well as prose guidance ever works.
• The computed half — a binary that reads the actual diff, resolves real tools, and exits non-zero. procoder check does not ask the agent whether the files are formatted; it formats them in memory and compares.

The second half is why procoder is a binary rather than a prompt. An agent under pressure talks itself past a paragraph. It does not talk itself past a non-zero exit code, because that conversation is not available.

What procoder does not do

• It does not write your code, and outside its own plugin cache it does not modify files at all. The binary prints; the agent writes. That is a hard rule with one recorded exception (procoder prune --apply, which touches only procoder's own cache).
• It does not replace a methodology. Procoder governs whether work is finished and honest. It has no opinion about how you decompose a problem, and it defers to BMad's artifacts when [planning] method = "bmad" is set. See Influences.
• It does not replace your linters. Each domain shells out to the field's canonical tool — gitleaks, semgrep, osv-scanner, ruff, prettier — rather than reimplementing them. Procoder decides what blocks; the tools decide what is true.
• It does not measure its own value. See Honest limits.

Compared to a skills pack

The closest comparison is a well-built Agent Skills collection. The difference is not quality; it is enforceability. A skills pack suggests a workflow; procoder computes a verdict. A pack can tell an agent to run the tests. Procoder can refuse to close the task when they did not run, because NOT run is a state it can observe and I ran them is a claim it cannot.

Procoder conforms to the Agent Skills spec and ships as a plugin, so the two are not alternatives — the skill is procoder's advisory half, and it is generated from the same source as every other host's rule file so they cannot drift.

For the specific projects solving adjacent problems, see Comparable projects. For the research behind the premises, see Research.