Claude Code can split a big job across several AI agents working at once. Left alone, they step on each other: one edits another's files, shortcuts creep into your architecture, and nobody remembers why the code is the way it is. code-kit gives each agent its own part of the codebase and checks its work before it can call it done. Context Graph keeps a short note on why each file exists, and makes an agent read a file before it's allowed to change it.
You describe what you want built. Claude Code becomes the lead and works out a plan with you. Then it hands the work to specialist agents and keeps them in line. Six words cover most of it:
Your main Claude Code session. It plans the work, hands out tasks, reviews what comes back and merges it. It doesn't build the features itself.
Areas of the codebase, each with its own agent: say, the web pages, the database, the AI calls. An agent can only change files in its own lane.
The plan's tasks, small enough to build and test in one go: "a settings screen where you paste your API key". Each one names the requirements it delivers.
The parts of your architecture and which may depend on which. If the business logic must never talk to the database directly, an agent can't make it.
A few sentences per file, kept in git: what the file is for, what it relies on, and what it must keep true. Agents read it before they edit, and update it after.
The conventions your code follows ("money is always stored in cents") and the record of why each change was made, so the next agent doesn't undo it.
When the lead finishes one thing, code-kit hands it the next: start the tasks that are ready, review finished work, merge what passed. If an agent goes quiet, it's nudged, then restarted. You decide how much it does without asking. A short summary above your prompt always shows the goal, what's happening now, what comes next, and anything waiting on you.
Telling an AI agent the rules in a prompt works for a while, then it drifts. These plugins hook into every action an agent takes and stop the ones that break your rules, explaining why and what to do instead. Here are real examples from building one app.
An AI agent only knows what it read in this session, and forgets it when the session ends. Context Graph writes the important parts down, next to your code in git: a card for each file, the rules each part of the project follows, and the reasons behind past decisions. When an agent is about to change a file, it gets exactly that file's card, rules and history. When it's done, it leaves a note on what it changed and why.
Context Graph draws the parts of your project as a tree. Each box shows how many rules that part has and how many of its files have a card: “3r 21/21” means three rules, and all 21 files have one. In the desktop app a ring fills as the cards are written. A box pulses blue when an agent records a decision there. If agents keep working around one of its rules, it turns red, and Context Graph suggests rewording the rule for you to accept or reject.
Drawn by Context Graph's own code, from the book editor's real modules. Switch between the desktop app and the terminal to see both.
For any task, code-kit's code map lays out the files it touches, grouped by part of the architecture, with lines showing which file uses which. A file lights up in an agent's colour while that agent reads or changes it, and says whether its why is written, out of date, or missing. Step through the files to read each one's card.
Drawn by code-kit's own code, from the real files of one task in the book editor: ST-19, the settings screen.
We asked for a book editor: upload a manuscript, have AI suggest grammar and flow fixes for each paragraph without changing the writer's voice, see the original and the suggestion side by side, and accept the changes you like. Claude Code worked out the specification with us, split it into tasks, and built it with a team of agents, with both plugins switched on throughout.
The empty app and the shared definitions every other part builds on, written by the lead itself. The first merge into the main branch needed a person's sign-off.
Reading Word, Markdown, plain text, EPUB and PDF files, built by several agents at once. Finding the chapters, fixing where they split, and a library of your books.
Claude reads the whole chapter once, then suggests fixes paragraph by paragraph. The changes are highlighted word by word, and you pick the ones you want.
The edited book back out as Word or Markdown, with the writer's italics intact.
Added after trying the app: paste your own API key in settings, and see the original page next to how the app split it into paragraphs.
You need Claude Code, Node 22 or later, and git. Type these into Claude Code. Neither plugin enforces anything until you've read and approved the rules it suggests for your project, and you can use either one on its own.
Gives each agent its own part of the code, protects your architecture, makes agents check their work against the spec and prove it before they're done, and keeps the build moving.
Keeps a card for every file and the reasons behind your code in git, and makes agents read a file before they change it. Works in Claude Code and Codex.