Three agents · one harness · your keys

Works with the agent you already use.

CodeMySpec runs as a plugin inside Claude Code, OpenAI Codex, or Google Antigravity. Claude Code has the deepest integration and is the flagship. Codex and Antigravity are early. All three run on your own API keys, with no markup on model tokens.

why this is possible

The harness isn't the plugin.

The harness is a local server on your machine, and each plugin is a thin adapter that points your agent at it. The requirement graph, the specs, the verification, and the QA are written once and reused by every agent. That is why supporting a new one is an adapter rather than a rewrite.

Three coding agents converging on one local harness Claude Code, OpenAI Codex, and Google Antigravity each connect through their own thin plugin adapter, which registers the harness tools and forwards lifecycle hooks. All three adapters converge on a single cms server running on your machine, which holds the requirement graph, the specifications, verification, and QA. Claude Code flagship thin adapter OpenAI Codex early thin adapter Google Antigravity early thin adapter cms · one local server on your machine requirement graph · specifications · verification · QA
Each adapter does three things: it registers the harness tools, forwards the agent's lifecycle hooks so verification can run, and ships the skills you drive the loop with. Everything else lives in the server.

the three, honestly

One is the flagship. Two are early. Here's the difference.

Every agent gets the same harness, but the agents themselves expose different hooks, so the integrations aren't identical. Rather than claim parity, here is what actually differs.

Claude Code

flagship

The original and the deepest integration. The full lifecycle, all eight skills, and the five specialized subagents that write specs, tests, code, BDD specs, and run QA.

Caveats: none. Everything downstream is ported from here.

OpenAI Codex

early

A thin port. Codex adopted a nearly identical hook protocol, so the same skills and the same server-side lifecycle work with very little translation.

Caveats: verification runs, but its file scoping doesn't recognize Codex's edit tool yet, so change-scoped checks act as though nothing changed. No subagent format.

Google Antigravity

early

Skills become slash commands automatically, and the five agents ship in Antigravity's own format. The server speaks its hook dialect.

Caveats: the surface emits no subagent events, so per-agent attribution doesn't apply, and the test-command rewrite arrives as a redirect rather than an edit.

The local server ships for macOS on Apple Silicon and Windows on x64. There's no Linux binary in the packaged install path today.

the same eight skills

The commands don't change when the agent does.

All three plugins ship the same eight skills, so the workflow you learn transfers.

/init

link your project

/product

define what to build

/design

plan the architecture

/next

start the next requirement

/develop

build one component

/implement

loop the whole graph

/qa

test the running app

/sync

refresh project state

what this buys you

You aren't betting your workflow on one vendor.

Your executable specifications and tests are real Elixir test files in your own repository, and the harness is a local server rather than a part of the agent. Switching agents means installing a different adapter. Every artifact comes with you, and when the next, better agent ships, the harness moves with you instead of trapping you.

It doesn't sit between you and the model. It sits around it.

questions

Straight answers.

What AI agents does CodeMySpec work with?

CodeMySpec works with Claude Code, OpenAI Codex, and Google Antigravity. Claude Code has the deepest integration and is the flagship; Codex and Antigravity are early. All three run on your own API keys with no markup on model tokens.

How can one tool support several different agents?

Because the harness is not the plugin. The harness is a local server on your machine, and each plugin is a thin adapter that registers the server's tools and forwards the agent's lifecycle events to it. The requirement graph, the specs, the verification, and the QA are written once and reused by every agent, so supporting a new one is an adapter rather than a rewrite.

Is CodeMySpec only for Claude Code?

No. Claude Code is the original and deepest integration, but CodeMySpec also ships plugins for OpenAI Codex and Google Antigravity. The same specs, verification, and browser QA work regardless of which agent drives them.

What is different about the early agents?

Claude Code gets the full lifecycle including the five specialized subagents. On Codex, verification runs but its file scoping does not yet recognize Codex's edit tool, so change-scoped checks behave as though nothing changed. On Antigravity the five agents ship, but the surface emits no subagent events, so per-agent attribution does not apply and the test-command rewrite is delivered as a redirect instead of an edit.

Which operating systems does the harness run on?

The local server ships for macOS on Apple Silicon and Windows on x64. There is no Linux binary in the packaged install path today.

Does using a plugin lock me into that agent?

No. Your executable specifications and tests are real Elixir test files in your own repository, and the harness is a local server rather than part of the agent. Switching agents means installing a different adapter; every artifact comes with you and still runs.

Install it in the agent you already have.