agentlink

One brain for every AI coding agent — with zero file duplication.

Write your rules and skills once. Claude Code, Cursor, Codex, Copilot, Antigravity and OpenCode all read the same files — the same inode, not a copy.

---

The problem everyone has

You write a skill for Claude Code. Now you want it in Cursor. And Copilot. And Codex. So you copy it. Then you fix a typo in one of them, and the others drift. Rename a folder and you have an orphan. Add a seventh agent and you copy again.

Existing tools solve this by generating copies from a source directory. That trades one duplication problem for another: a generate step you must remember to run, files that go stale the moment anyone edits the wrong one, and renames that leave orphans behind.

The insight

In 2026 the ecosystem already converged. For the two resources that matter most, the format is identical and only the location differs:

• Instructions → AGENTS.md, an open spec donated to the Linux Foundation in December 2025, read by 20+ tools.
• Skills → Agent Skills / SKILL.md, an open specification read by 30+ tools.

	Claude Code	Cursor	Copilot	Antigravity	Codex
Skills format	SKILL.md	SKILL.md	SKILL.md	SKILL.md	SKILL.md
Skills location	.claude/skills	.cursor/skills	.github/skills	.agents/skills	.agents/skills

Converting an identical format to an identical format is invented work. The real problem is placement — and placement is what filesystems solve.

What agentlink does instead

Every pairing of an agent with a resource gets one of four verdicts:

Verdict	When	Action	Cost
native	the tool already reads the canonical path	nothing	zero
link	same format, different path	symlink or junction	one reparse point
import	linking impossible; the tool has an include syntax	a one-line stub	~11 bytes
blocked	needs a human decision	nothing is written	—

The canonical layout is AGENTS.md + .agents/, which is not an agentlink invention — it is what Antigravity and Codex already read natively. Choosing the layout the ecosystem converged on is what keeps most verdicts at native, and every native verdict is work that never happens.

Why this makes bidirectional sync disappear

For native and link verdicts there is exactly one inode. So this is not a feature anyone implemented — it is arithmetic:

$ mkdir .cursor/skills/deploy && echo "..." > .cursor/skills/deploy/SKILL.md

$ cat .claude/skills/deploy/SKILL.md    # instantly there
$ cat .github/skills/deploy/SKILL.md    # instantly there
$ mv .github/skills/deploy .github/skills/release   # renamed for everyone
$ rm -rf .opencode/skills/release                   # gone for everyone

No daemon. No watcher. No generate step. No drift, because there is nothing to drift from.

Quick start

$ agentlink init      # pick your agents, then link them to the canonical layout
$ agentlink adopt     # or: absorb the .claude/ you already have
$ agentlink status    # see exactly who reads what

init asks which agents this repository actually uses, preselecting the ones it already finds:

Which agents does this repository use? (space toggles, enter confirms)
  [ ] antigravity     Google Antigravity — reads the canonical layout, costs nothing
> [x] claude-code     Claude Code — CLAUDE.md, .claude/skills
  [ ] codex           OpenAI Codex CLI — reads the canonical layout, costs nothing
  [ ] cursor          Cursor — .cursor/skills
  [ ] github-copilot  GitHub Copilot — .github/skills
  [ ] opencode        OpenCode — .opencode/skills

Nothing is created for an agent you did not pick, and dropping one later — via agentlink providers --select — retires the links it no longer needs. Scripts are never prompted: without a terminal, or with --providers claude-code,cursor, init just gets on with it.

Onboarding an existing repository that has only ever used Claude Code, with every agent selected:

$ agentlink adopt

  AGENTS.md (instructions)
    antigravity      native  reads AGENTS.md directly — nothing to do
    claude-code      ok      CLAUDE.md (import stub)
    codex            native  reads AGENTS.md directly — nothing to do
    cursor           native  reads AGENTS.md directly — nothing to do
    github-copilot   native  reads AGENTS.md directly — nothing to do
    opencode         native  reads AGENTS.md directly — nothing to do

  .agents/skills (skills)
    antigravity      native  reads .agents/skills directly — nothing to do
    claude-code      ok      .claude/skills (junction)
    codex            native  reads .agents/skills directly — nothing to do
    cursor           ok      .cursor/skills (junction)
    github-copilot   ok      .github/skills (junction)
    opencode         ok      .opencode/skills (junction)

  12 capabilities · 12 need nothing · 0 files copied

  3 created · 2 adopted

Seven of those twelve capabilities cost nothing at all, because those tools already read the canonical paths. 0 files copied is structural, not a coincidence.

Installation

# npm — ships a prebuilt binary per platform, no Rust needed
npm install -g @fialhosoft/agentlink

# Homebrew
brew install fialhosoft/tap/agentlink

# Cargo
cargo install agentlink-cli

Platform support

agentlink works on Windows without administrator rights or Developer Mode, which is the case most tools in this space get wrong.

	Directories (skills)	Files (CLAUDE.md)
Linux / macOS	symlink	symlink
Windows, Developer Mode on	symlink	symlink
Windows, out of the box	junction — never needed elevation	@AGENTS.md import stub

Link support is probed at runtime, not inferred from the platform: two machines running the same Windows build can differ, so agentlink tries and adapts.

Adding a new agent

A provider is data, not code. Adding one is a single file:

# crates/agentlink-domain/providers/my-agent.toml
schema = 1
id = "my-agent"
name = "My Agent"

[[capability]]
resource = "skills"
strategy = "link"
path = ".my-agent/skills"

That is the entire change. The build embeds every manifest in crates/agentlink-domain/providers/ automatically — no list to update, no Rust to write, no core to touch.

Manifests are validated at load time: a native claim is checked against the canonical path, unknown keys are rejected, and paths that could escape the workspace are refused. To try one before sending a pull request, drop it in .agentlink/providers/ and it takes effect immediately.

Commands

Command	Purpose
agentlink init	Pick the agents this repository uses, then materialise their capabilities
agentlink adopt	Move an agent's existing content into the canonical layout, then link it back
agentlink apply	Materialise anything missing or out of date (idempotent)
agentlink status --check	Exit 2 if anything is pending — for CI
agentlink doctor	Diagnose privileges, drift and stale links
agentlink providers	Show the capability matrix
agentlink providers --select	Change which agents this repository serves
agentlink clean	Remove everything agentlink created, keeping the canonical layout

How it treats your repository

agentlink commits only the canonical layout and keeps every agent-specific path local, in a managed .gitignore block:

# >>> agentlink >>>
/.claude/skills
/.cursor/skills
/.github/skills
/CLAUDE.md
# <<< agentlink <<<

Committed symlinks silently degrade into plain text files for teammates on Windows without core.symlinks, and junctions cannot be represented in git at all. So each developer materialises their own — an idempotent command that takes about 20 ms, ideal for a post-checkout hook. See ADR 0005.

Safety

The rule the whole design enforces: agentlink never destroys content it did not create.

• A lock file records everything it materialised. Nothing else is ever removed.
• No removal is recursive — it unlinks reparse points, never directory trees.
• A path holding real content is reported as blocked, never overwritten.
• Every blocked outcome prints the exact command that resolves it.
• status and apply render the same plan, so nothing happens unannounced.

What this release does not do yet

Being explicit about the boundary, because trust matters more than surface area:

• v0.0.1 covers instructions and skills — the two resources where the format already converged, so they are fully solved by linking.
• MCP, subagents, commands, hooks and settings are not covered yet. These genuinely differ in format (mcpServers vs servers, env vs environment, JSON vs TOML), so they need transformation and a reconciliation model rather than a link. That is the v0.1 milestone.
• Project scope only. User-level (~/.claude, ~/.codex) comes next.

Documentation

• Architecture — how the pieces fit
• Architecture Decision Records — why, and what was rejected
• Contributing — including how to add an agent
• Roadmap

License

Apache-2.0. Chosen over MIT for its explicit patent grant, which matters for a project that aims to be adopted as shared infrastructure by companies building commercial agents.