Other coding agents
This page is for developers who use a coding agent other than Claude Code and Codex. It shows how each one follows an instruction from instruction.md, and when a new version reaches a session that is already running. Each tool has a page of its own with its exact setup.
At a glance
| Tool | How it follows | A new version arrives | Set up | Checked against | Systems |
|---|---|---|---|---|---|
| Gemini CLI | hooks | at your next prompt | one command | Gemini CLI v0.69.0, against its source | any |
| Antigravity | hooks | before the next model call | one command | Antigravity's documentation | any |
| GitHub Copilot CLI | hooks | after the next tool call | one command | Copilot CLI 1.0.114, from its documentation | macOS, Linux and Windows |
| Cursor | hooks, and a rule | after the next tool call | one command, and watch | Cursor 2026.11.13, from its documentation | any |
| OpenCode | a file it reads again | before the next model request | watch, and its settings | OpenCode 2.3.2 and 1.19.6, against their documentation and source | any |
| Kilo Code | a file it reads again | as a session runs | watch, and its settings | Kilo Code 7.8.27, from its documentation | any |
| Qwen Code | hooks | at your next prompt | one command | Qwen Code v0.28.1-preview.0, against its source | any |
| Factory Droid | hooks | at your next prompt | one command | Factory Droid 1.14.0, from its documentation | any |
| Devin | hooks | at your next prompt | one command | Devin's documentation | any |
| Kiro | hooks | at your next prompt | by hand | Kiro's documentation | any |
| Cline | hooks | at your next prompt | one command | Cline 4.1.26, against its hooks documentation | macOS and Linux |
| JetBrains Junie | a hook, and its AGENTS.md | after the next tool call | by hand | Junie CLI (early access), from its documentation | any |
| Crush | a hook, and its context files | after the next tool call | one command | Crush v0.101.0, against its hooks documentation | any |
| Goose | a file it reads again | with every message | watch, and a setting | Goose 1.60.1 | any |
| Aider | a file it reads again | with every message | watch, and a setting | Aider, against its source | any |
| pi | its own package | at your next prompt, in its system prompt | pi install | pi 1.5.2 | any |
| Amp | a plugin | at your next prompt | a file to copy | Amp's plugin interface (@ampcode/plugin); not yet run inside Amp | any |
| Zed, JetBrains and other ACP clients | a go-between before the agent | at your next prompt | the agent's command | ACP schema v1.33.0 | any |
"Checked against its source" means the hook's answer was read in the tool's own code. "From its documentation" means it follows what the tool's documentation says, and each tool's page lists what isn't verified yet.
Three ways a tool follows
- Hooks. The tool runs a command at some of its events and adds what the command prints to the agent's context. instruction.md's command is
instruction agents hook <tool>. When a session starts, it gives the instruction's whole text. At later events, it adds nothing while the version is current, and the new version, marked as replacing the old one in full, once one is released. Which events a tool sends decides when a new version arrives: at your next prompt, after the next tool call, or before the next model call. - A file the tool reads again. Some tools take no text from a hook, but read a file again as they work: before each model request (OpenCode) or with every message (Goose, Aider).
instruction watchkeeps that file current, and the tool's settings name it. - Code of instruction.md's in the tool. pi takes instruction.md's extension, Amp a plugin file, and an editor that speaks the Agent Client Protocol (ACP) a go-between that runs before the agent. Each reads the instruction as the hooks do and gives it at the moment the tool allows.
The hooks, pi's extension, Amp's plugin and the go-between read the way How following works describes: at most every refreshSeconds, within 1.5 seconds, else the last good copy, so a prompt is never held up. A session gets each version once, even when a tool runs two sets of hooks for one event (Cursor, Copilot CLI and Devin also run Claude Code's .claude/settings.json).
Set up a project
What you need: Node.js 20.19 or later. The instruction command is the instruction.md package (npm i -g instruction.md, 0.11.0 or later). instruction agents … hands over to @instruction.md/agents, 0.4.1 or later. Without installing anything, the same commands run with npx:
npx -y @instruction.md/agents follow instruction://acme/support-agent@stable --host gemini-cli --write Sign in once with instruction login; instruction agents login is the same sign-in. A public instruction can be followed without signing in.
For a tool with hooks, run this in your project's folder, with the tool's name after --host:
instruction agents follow instruction://acme/support-agent@stable --host <tool> --write --host takes gemini-cli, antigravity, copilot, cursor, qwen-code, droid, devin, cline and crush, and kiro and junie, whose hooks you add by hand. It writes, keeping anything already there:
.instruction/follow.json: which instruction the project follows. Commit it, so everyone on the project follows the same one.- The tool's hooks, in the tool's own file. Commit it.
.instruction/.gitignore, so the copy the hooks keep (.instruction/current.md),state.jsonandgiven/stay out of git.
{
"follow": "instruction://acme/support-agent@stable"
} - Without
--write, the command prints the same files for you to add yourself. - Either way, it also prints how to add instruction.md's MCP server to the tool.
instruction connect --host <tool> <uri>prints only that. --require-signatureapplies only versions whose signature checks out.--dir Dsets up another folder.--cli CMDchanges how the hook runs the agents' command, for example--cli instruction-agentsonce@instruction.md/agentsis installed globally.- Every field of the follow file is on How following works.
Any other --host is refused with --with-hooks and --write are for --host …: OpenCode, Kilo Code, Goose and Aider follow a file that watch keeps, and their pages say how.
The MCP server in tools that speak an older MCP
instruction.md's server speaks MCP 2026-07-28 only. Most coding agents speak an earlier version, so they reach it through the instruction command's own stdio server, instruction mcp --follow <uri>. It serves the followed instruction as the resource instruction://current and the prompt follow, and, signed in, instruction.md's tools on the same connection. Sign it in once with instruction login. Each tool's page shows the settings instruction connect --host <tool> <uri> prints. For VS Code, see VS Code.
The MCP server is how an agent reads or changes instructions. What keeps a running session on the current version is the hooks, the file or the extension.
Keep watch running
The tools that read a file need instruction watch running while you work. Run it in a terminal of its own, or as a service of your user. In a project's folder, instruction watch without an address follows .instruction/follow.json and writes .instruction/current.md. On Linux, as a systemd user service:
# ~/.config/systemd/user/instruction-watch.service
[Unit]
Description=instruction.md: keep .instruction/current.md current
After=network-online.target
[Service]
WorkingDirectory=<the project's folder>
# the absolute path that `command -v instruction` prints; without an address, watch follows .instruction/follow.json
ExecStart=<path to instruction> watch
Restart=always
RestartSec=30
[Install]
WantedBy=default.target systemctl --user daemon-reload
systemctl --user enable --now instruction-watch.service The launchd agent (macOS) and the scheduled task (Windows) for watch are on Codex. Use watch without an address there, and set the project's folder as the working folder: the plist's WorkingDirectory key, or New-ScheduledTaskAction -WorkingDirectory. A service needs node on its PATH; the Codex page says how.
Plans and live updates
The hooks, pi's extension, Amp's plugin and the ACP go-between read the instruction; they don't ask for live updates. Reading is never limited, on any plan. instruction watch asks for live updates while it runs. It keeps the file for OpenCode, Kilo Code, Goose, Aider, Cursor's rule and your own GEMINI.md or AGENTS.md. On the Community plan, one agent or app has live updates at a time:
Plans Community 1 · Pro unlimited · Team unlimited · Enterprise unlimited
When another one has them, watch says live updates are taken: … Still reading on schedule. and goes on reading every 60 seconds. The rule is on Live updates.
Your own instructions in every project
Several tools read a file of your own when a session starts. instruction watch <uri> --out <file> --block keeps a region of it current and leaves the rest as you wrote it. The files it knows, live ones first:
~/.codex/AGENTS.mdand~/.config/opencode/AGENTS.md, read again before each model request;~/.gemini/GEMINI.md(Gemini CLI, Antigravity),~/.copilot/copilot-instructions.md,~/.qwen/QWEN.md,~/.config/kilo/AGENTS.md,~/.junie/AGENTS.md,~/.config/amp/AGENTS.md,~/.agents/AGENTS.md(Cline),~/.pi/agent/AGENTS.mdand~/.config/zed/AGENTS.md, each read when a session starts.
Check it worked and stop following
Each tool's page says how to run its hook by hand, which prints what the tool receives, and how to remove what the setup wrote. Once a session has applied a version, the instruction's Agents card counts the follower under its tool. instruction agents check checks Claude Code and Codex setups only.
If it doesn't work
| What you see | Why | What to do |
|---|---|---|
--with-hooks and --write are for --host … |
The tool takes no hooks from the command | Follow its page: it reads a file that watch keeps |
npx -y @instruction.md/agents watch … fails |
The printed watch line names the wrong package |
Run instruction watch … or npx -y instruction.md watch … |
| A session starts without the instruction | The tool hasn't trusted the project's hooks, or the setting that turns hooks on is off | See the tool's page: the setup ends with what the tool asks |
| A private instruction isn't read | The hook or watch isn't signed in |
instruction login |
| Still stuck | Ask us |
Questions
Which tools get a new version in the middle of a task?
Antigravity gets it before the next model call, and OpenCode before the next model request. Copilot CLI, Cursor, Cline, Junie and Crush get it after the next tool call. The others get it at your next prompt or with your next message; the table above says which.
Do the hooks use my plan's live updates?
No. The hooks, pi's extension, Amp's plugin and the ACP go-between only read, and reading is never limited. Only instruction watch asks for live updates. See Plans and live updates.
Can one project be set up for several tools?
Yes. Run the setup once per tool, with each tool's --host. They share one .instruction/follow.json, and each tool's hooks go in its own file.
Is my tool missing?
A tool that runs hooks the way Claude Code does can use instruction agents hook claude-code. Otherwise, have it read a file that instruction watch keeps. Tell us which tool on Support.