In the AI era, something funny is happening: side projects that have been collecting dust in ~/code/wishful-thinking/
are getting dusted off. Suddenly that "someday I'll write this" repo on GitHub has a real README, a working CI, and three open issues you actually want to close.
Why? Because the AI doesn't complain. Doesn't get bored. Doesn't ask "but why do we need this when we have X?" at 11pm when all you want is to feel better with yourself, when your children are on the bed and you think "now is the moment i should use to realize some of my personal projects."
So while the discourse is busy replacing you with LLMs, you're quietly using LLMs to replace the procrastination that used to replace you.
This post is about one of those moments, one of those ideas born in the night, that probably, in pre-AI era, would have been forgotten behind "it's too late to open my laptop": a small CLI called
agent-contexts I wrote because I was sick of having 40
AGENTS.md
tabs open across 40 different repos.When you sit down with Claude Code, Cursor, Gemini CLI, or any of the other agentic tools, they don't read your mind. They need to be told things. Some of those things you tell them in chat. But the important stuff — the project conventions, the "don't do this here" rules, the "we use tabs not spaces, fight me" — should live in the repo itself, so every agent (and every teammate) picks them up automatically.
Different agents have different conventions for where to look:
CLAUDE.md
(and also AGENTS.md
, because Anthropic isn't petty).GEMINI.md
.AGENTS.md
..cursorrules
and similar.So if you're running a polyglot agent setup (and who isn't these days)you're maintaining multiple files per project. That's annoying before you've written a line of code.
The thing i learned the hard way: bigger context does not equal better context.
When you feed your agent a 4,000-line AGENTS.md
covering every
architectural decision from 2019 to today, two things happen:
The same project rarely wants the same context for every task:
Same repo. Four different AGENTS.md
files. (Plus nested ones for
subfolders, because your src/payments/
deserves its own context, thank you very much.)
The "obvious" solution at the beginning was one big monolithic AGENTS.md
at the root... but soon you understand the unefficency of that approach. The other "obvious" solution currently is a hand-maintained tree of nested AGENTS.md
files in every folder, it wins on quality but loses on maintenance. The day someone updates the root convention, they have to remember to propagate it. And nobody remembers.
The Vercel team had a similar itch a while back, and they scratched it with agent skills. They treated the bundle of "things an agent needs to do a job" (instructions, scripts, references) as a package. So they made it installable. You can npx skills add <source>
, the tool fetches the right files, drops them in the right places, and tracks what's installed in a lockfile. Deterministic, versioned, no fuss.
But skills are about capabilities: "here's how to run my CI", "here's how to deploy", "here's a script to seed the DB". They don't really cover the AGENTS.md
use case, where you want to distribute context, not behavior.
So I built something in the same vein, but scoped to context files.
agent-contexts
: npm, but for AGENTS.md
agent-contexts is a CLI. The way to think about it:
Curate agent context once in a central git repo. Install it into every
consumer project. Versioned, deterministic, reproducible.
You write a contexts.yml
at the root of your contexts repo:
version: "1"
name: engineering-contexts
mappings:
".":
context_source: ./agents/root/AGENTS.md
description: Repo-wide conventions
src/products:
context_source: ./agents/products/AGENTS.md
description: Products module
src/common/database:
context_source: ./agents/database/AGENTS.md
description: Sequelize + MySQL integration
Each mapping points a folder in the consumer project at a profile file in the contexts repo. agent-contexts add
then:
.contexts/cache/<slug>/
.contexts.lock
pinning each source to a commit SHA and each file to a SHA-256 hash.agent-contexts install
is the npm ci
of context. It only reads
contexts.lock
, never the manifest, so CI doesn't have to fetch the manifest or guess about upstream availability. Same lock, same files, every time.
agent-contexts status
recomputes truth from disk and exits non-zero if anything is broken, missing, or hand-edited. Easy CI gate.
Sometimes you need a different context for the same folder, depending on what you're doing.
Say your repo has a default tone — calm, professional, focused on the production codebase. But today you're refactoring and you want a more pedagogical tone with extra "before you change this, consider…" notes. Or you're onboarding a junior dev and you want a glossary of project-specific terms front and center.
You author those as a tag:
mappings:
".":
context_source: ./agents/root/AGENTS.md
src/products:
context_source: ./agents/products/AGENTS.md
tags:
onboarding:
mappings:
".":
context_source: ./agents/root/onboarding/AGENTS.md
src/products:
context_source: ./agents/products/onboarding/AGENTS.md
refactor:
mappings:
src/products:
context_source: ./agents/products/refactor/AGENTS.md
Switch with a flag:
npx agent-contexts add your-org/your-contexts --tag onboarding
npx agent-contexts update --tag refactor
Effective set is root ∪ tag
(tag wins on conflicts), so you only write the differences. The lock records which tag each entry is on, so install
reproduces it and update
keeps you there.
NestJS microservice. You want a Star-Wars-toned AGENTS.md
for
src/products
— because why not.
mkdir -p agents/star-wars
cat > agents/products/AGENTS.md <<'EOF'
Follow the Repository pattern for data access.
Split Read and Write services (CQS-lite).
Use `class-validator` DTOs for every input.
EOF
cat > agents/star-wars/products.md <<'EOF'
The domain heart, this is. Where artifacts of the holonet catalog,
manage you do.
- The Repository pattern, follow — data access through one gate
- Read and Write services, split — CQS-lite, the path of clarity
- DTOs with `class-validator`, use — `@nestjs/swagger` for the holomap
EOF
cat > contexts.yml <<'EOF'
version: "1"
mappings:
src/products:
context_source: ./agents/products/AGENTS.md
description: Products module
tags:
star-wars:
mappings:
src/products:
context_source: ./agents/star-wars/products.md
description: Products module — Yoda edition
EOF
In the consumer project:
npx agent-contexts add ../your-contexts --tag star-wars --force -y
npx agent-contexts status
npx agent-contexts update # keeps the recorded tag
npx agent-contexts add ../your-contexts -y # back to root mappings
That's the whole loop. Symlink in, versioned, switchable, reproducible.
Tags turn out to be the right tool for one specific scenario that I keep running into: a CI job that boots a code-review agent on every PR. You don't want that agent reading the same AGENTS.md
files as your dev-time sessions, you want it focused, skeptical, and pointing at your review checklist, not your "how to add a new endpoint" doc.
So you ship a ci-code-review
tag alongside the default one:
mappings:
".":
context_source: ./agents/root/AGENTS.md
src/products:
context_source: ./agents/products/AGENTS.md
tags:
ci-code-review:
mappings:
".":
context_source: ./agents/root/ci-review.md
description: Code-review checklist + tone for CI agents
src/products:
context_source: ./agents/products/ci-review.md
description: Review-time conventions for the products module
Then the CI job does the swap before invoking the agent:
npx agent-contexts install
npx agent-contexts update --tag ci-code-review
npx your-cr-agent
Same project, same lock, completely different AGENTS.md
content at every target. The dev never sees the CI variant; the CI never sees the dev variant. Both live in the same repo, both are versioned, and a PR against either one is just a normal git workflow.
| Command | What it does |
|---|---|
agent-contexts add <source> |
|
| Fetch, select mappings, link, write lock. | |
agent-contexts install |
|
Headless restore from contexts.lock . The npm ci . |
|
agent-contexts update |
|
| Pull upstream, diff, re-pin. | |
agent-contexts status |
|
| Per-entry truth from disk. Exit 5 on problems. | |
agent-contexts list |
|
| Pretty table straight from the lock. | |
agent-contexts reset |
|
Undo: remove links, restore .bak backups, delete lock + cache. |
All commands take --json
, -y
, --dry-run
, and --verbose
. None of them need a TTY, so they slot into CI without fuss.
The full docs and source live at
Pros
contexts.lock
, same files, anywhere — laptop,
CI, your colleague's machine.npm
, cargo
, or pip
already gets
it.AGENTS.md
, CLAUDE.md
, GEMINI.md
, whatever your
agent of the week wants.--json
everywhere, deterministic exit codes, no TTY
needed, for example you can wire it into a CI job that swaps to a ci-cr
tag before booting the reviewer agent, then walk away.Cons (fair's fair)
install
won't find anything
to fetch from. The reason is in the docs (TL;DR: only a git URL is the
same string everywhere). Use git for shared/CI; locals are for solo
experiments...
escapes, lowercase
hex hashes, that sort of thing. Read the spec once and you'll stop
fighting it.agent-contexts
version.If you've ever felt the pain of maintaining AGENTS.md
files across a fleet of repos — or worse, watching them drift out of sync — give agent-contexts a try. The README is short, the install is one command, and the worst case is that you
rm
a .contexts/
folder and a contexts.lock
.Found a bug? Open an issue. Got a feature idea? Open a PR. Want to yell at me about the name (it should be agent-context
and the binary should be contexts
, I know, I know)? There's an issue tracker for that too.
And a proper thank-you to the Vercel Skills folks. The shape of
That's all. Go ship something.