devrig CLI

Independent open-source project. MCP Steroid and devrig are independent open-source projects. They are not made by, endorsed by, or affiliated with JetBrains s.r.o. IntelliJ IDEA, PyCharm, GoLand, WebStorm, Rider, CLion, and JetBrains are trademarks of JetBrains s.r.o.

What is devrig?

devrig is a small standalone command-line tool that connects your AI Agent to a JetBrains IDE running MCP Steroid — with no manual MCP configuration.

It does three jobs:

  1. Registers MCP Steroid with your coding agent. devrig install <agent> adds devrig as an mcp-steroid stdio MCP server in Claude, Codex, or Gemini.
  2. Bridges your agent to every running IDE at once. When your agent launches it as a stdio MCP server (devrig mcp), devrig discovers all the IntelliJ-based IDEs running on your machine — across projects — and routes the agent’s MCP Steroid calls to any of them through a single connection. One bridge, every IDE.
  3. Provisions an IDE. devrig backend download, devrig backend start, and devrig backend stop download and run a managed IntelliJ backend under devrig’s home directory, so an agent can spin up an IDE with no manual setup.

One bridge, every IDE

A single devrig mcp process connects your AI Agent to all the IntelliJ-family IDEs running on your machine at once — each open on a different project — and can download and start more on demand:

One devrig bridge connects your AI Agent to all running IDEs at once — and can start more
One devrig process bridges your agent to every IntelliJ-family IDE running on the machine — and can start more.

devrig is a Java application. The one-command installer downloads devrig together with a matching Java runtime into ~/.mcp-steroid, so there is nothing to set up by hand. To run devrig under a Java you manage instead, point DEVRIG_JAVA_HOME (or JAVA_HOME) at it.

Install

Register devrig as the mcp-steroid stdio MCP server in your coding agent:

devrig install claude
devrig install codex
devrig install gemini

The agent must be one of claude, codex, or gemini. After a successful install, devrig prints the agent it registered, the JAVA_HOME it recorded, and the exact stdio command (devrig mcp) the agent will run.

Commands

Run devrig --help (or devrig -h) for the authoritative usage, and devrig --version (or devrig -v) for the version. Help is generated from the same command tree that performs the work, so every nested command has focused help such as devrig install --help and devrig backend download --help. The equivalent discoverable route is devrig help <command>, including nested paths such as devrig help backend download. devrig tools prints the full MCP-tools-as-CLI reference — every tool command’s usage line and flags on one page, written for coding agents.

If required values are missing, devrig prints that focused help and identifies all missing values in one pass, including allowed enum values and alternatives such as --code / --code-file. Human output may use terminal color. Commands that advertise --json emit one ANSI-free JSON document for agents and scripts.

devrig
├── mcp
├── list_projects [--json]  (aliases: projects, project)
├── list_windows [--json]
├── execute_code ... [--json] [--out <path>]
├── execute_feedback ... [--json]
├── take_screenshot ... [--json] [--out <path>]
├── input ... [--json]
├── fetch_resource ... [--json]  (alias: prompt)
├── open_project ... [--json]
├── backend [--json]
│   ├── download [<id>] [--version <v>] [--json]
│   ├── start [<id>] [--version <v>] [--json]
│   ├── stop [<id>] [--version <v>] [--json]
│   └── provision [<id>] [--json]
├── install [--json]
│   ├── claude [--check]
│   ├── codex [--check]
│   ├── gemini [--check]
│   ├── config [--json]
│   ├── devrig
│   └── plugin [--check]
├── tools
├── help [<command>...]
└── version [--json]

Human output is formatted for the terminal and uses color when supported. Commands that advertise --json emit exactly one ANSI-free JSON document on stdout, suitable for an agent or a pipeline:

$ devrig version --json
{"version":"<version>"}

$ devrig install --json | jq -c '.targets[] | {name, kind}'
{"name":"claude","kind":"agent"}
{"name":"codex","kind":"agent"}
...

Generated MCP-tool commands suppress their live progress stream in --json mode because agent shell tools commonly merge stderr into their command result. Without --json, the same commands keep live progress on stderr and readable results on stdout.

Commands that do not have a coherent single-document response, including mcp, agent registration, and plugin installation, reject --json instead of mixing progress text with structured output.

devrig mcp

Runs devrig as an MCP stdio server. This is the command your coding agent launches after devrig install — you normally don’t run it by hand. While running, it discovers IDEs and bridges the agent’s MCP Steroid calls to them.

The legacy spelling devrig mpc is still accepted as a hidden alias, so older agent registrations keep working. Use devrig mcp for new setups.

MCP tools as direct commands

The same MCP tools exposed through devrig mcp are also regular shell commands. Every one supports --json; required inputs, types, enum choices, aliases, and help are generated from the tool schema rather than maintained in a second command definition.

CommandPurpose and important inputs
devrig list_projectsLists open projects and the routing keys accepted by --project_name.
devrig list_windowsLists attended IDE windows, modal/indexing/initialization state, and background tasks; it is not a project-routing or Maven/Gradle readiness gate.
devrig execute_codeRuns Kotlin in an IDE. Requires --project_name, --task_id, --reason, and either --code or --code-file; accepts --modal, --timeout, and --out. Quote inline Kotlin for the shell, for example --code='println("hello")', or prefer --code-file; use --code-file=- for stdin.
devrig execute_feedbackRates an execution. Requires --project_name, --task_id, --success_rating, and --explanation; accepts the exact optional --execution_id, and code can also come from --code-file.
devrig take_screenshotCaptures an IDE image. Requires --project_name, --task_id, and --reason; accepts --window_id and --out.
devrig inputSends keyboard and mouse steps to a window. Requires --project_name, --task_id, --reason, --window_id, and --sequence.
devrig fetch_resourceFetches an mcp-steroid:// guide given as a positional URI plus --project_name. devrig prompt is an alias; the former --uri spelling remains accepted for existing scripts.
devrig open_projectOpens --project_path. When one backend already owns that path, devrig reuses it automatically; use --backend_name to choose among candidates for a new or multiply-open path. --trust_project is on by default and --no-trust_project disables it. --wait polls for up to 300 seconds and returns the opened project’s opaque project_name, backend_name, and canonical path. This proves routing, not window or Maven/Gradle readiness.

Use command-scoped help for the authoritative grammar, for example:

$ devrig execute_code --help
$ devrig list_projects --json
$ devrig open_project --project_path="$PWD" --task_id=demo-open --reason="open current project from CLI" --wait --json
$ devrig execute_code --project_name="PROJECT_NAME_FROM_LIST_PROJECTS" --code='println("hello")' --task_id demo --reason 'verify IDE access' --json
$ devrig prompt mcp-steroid://prompt/skill --project_name="PROJECT_NAME_FROM_LIST_PROJECTS" --json

Generated tool commands do not repair the launcher or write to PATH; they only perform the requested tool call. Progress and diagnostics go to stderr, while stdout remains human output or one clean JSON envelope.

devrig install [--json]

With no target, lists every install target and detects which agent CLIs are on PATH. --json emits the same inventory as a targets array.

devrig install <agent> [--check]

Registers this devrig binary as the mcp-steroid stdio MCP server in the selected agent. --check is read-only: it reports registration drift and IDE reachability without changing configuration.

devrig install config [--json]

Prints the manual mcpServers configuration plus the equivalent Claude, Codex, and Gemini add commands. --json exposes serverName, mcpServers, and tokenized agentCommands fields.

devrig install devrig

Re-registers devrig’s stable launcher and user PATH entry.

devrig install plugin [--check]

Installs MCP Steroid into locally running JetBrains IDEs. --check lists the IDEs that would be asked without showing installation dialogs.

devrig backend [--json]

Lists discovered backends (with versions), grouped as MCP Steroid backends, other/incompatible IDEs, and installed-but-not-running (startable) backends. Per-backend open projects are listed by devrig list_projects. --json emits a single machine-readable object on stdout (pipe through jq); the default is human-readable text.

devrig list_projects [--json]

Lists open projects across all discovered backends. The default is a readable project/backend table. --json emits the standard generated-command envelope; project rows are under .data.content[].json.projects[]. devrig projects is the compatibility alias with identical help, behavior, and output; the singular devrig project spelling remains accepted, but its former top-level --json shape is replaced by the generated envelope described above.

devrig backend download [<id>] [--version <v>] [--json]

With no id, lists the IDEs available for download. With an id, downloads and installs a managed backend under devrig’s home directory. The id accepts <product>, <product>:<version>, or <product>-<version> — for example idea-community, idea-community:2026.1, or idea-community-2026.1.

Known product ids: idea-ultimate, idea-community, pycharm-pro, pycharm-community, goland, webstorm, rider, clion, rustrover, phpstorm, rubymine, datagrip, mps, android-studio.

MPS is published without a Linux ARM64 or Windows ARM64 distribution, so those two hosts report the platform as unavailable rather than downloading a mismatched archive.

devrig backend start [<id>] [--version <v>] [--json]

With no id, lists installed backends. With an id, starts an installed managed backend in detached mode and prints its pid, log, and config paths. A product-only id prefers the highest locally installed version.

devrig backend stop [<id>] [--version <v>] [--json]

With no id, lists currently running backends. With an id, stops a managed backend by its pid file. A product-only id prefers the highest locally installed version.

devrig backend provision [<id>] [--json]

With no id, lists port-discovered IDEs that can be provisioned. With an id (for example port-63342), prints manual MCP Steroid plugin install instructions for that IDE.

Options and environment

Options:

  • --debug enables verbose stderr logging (also enabled by DEVRIG_DEBUG).
  • --json emits one ANSI-free JSON document where advertised: backend, list_projects (also projects and legacy project), install, install config, version, backend lifecycle commands, and every schema-generated MCP tool command.
  • --help, -h prints command-scoped help and exits.
  • Root devrig --version / devrig -v prints the devrig version and exits. Backend lifecycle --version <v> options select an IDE version instead.

Environment variables:

  • DEVRIG_JAVA_HOME selects the JDK/JRE used to launch devrig instead of the bundled runtime. It overrides JAVA_HOME for the devrig process only.
  • DEVRIG_JVM_OPTS adds JVM options to the devrig launch, for example -Xmx512m.

Example: an agent provisions an IDE

This is the typical end-to-end flow where an agent gets a working IDE with no manual setup:

# 1. Register devrig with your agent (once)
devrig install claude

# 2. Download a managed IntelliJ IDEA Community backend (if not yet installed)
devrig backend download idea-community

Once downloaded, the agent can open a project immediately — steroid_open_project detects the installed (not-yet-running) backend as a startable candidate, starts it automatically (blocking until reachable), and opens the project in a single call. No separate devrig backend start step is needed.

To stop the backend when done:

devrig backend stop idea-community

devrig backend start <id> / devrig backend stop <id> still exist for explicit lifecycle control when you prefer it.

Frontendless IntelliJ IDEA Ultimate 2026.2

For unattended Java semantic work, devrig can run IntelliJ IDEA Ultimate 2026.2 (baseline 262) as a native Remote Development backend with MCP Steroid installed:

devrig backend download idea-ultimate --version 2026.2.0.1

After the download, call steroid_open_project normally. It starts the backend on demand, waits until MCP Steroid is reachable, and routes the requested project. No separate devrig backend start command and no Remote Development client window are required.

This mode is frontendless. The validated native IU-262 run logged Remote Development backend mode with headless=false; more generally, backend product mode takes precedence over the raw AWT-headless flag, and only plain non-backend headless mode is unsupported. A window or screenshot is therefore not a readiness requirement: wait until the requested path appears in steroid_list_projects, keep its returned project_name, then trigger and await Maven/Gradle import before indexed semantic work. If an attended frontend exists, steroid_list_windows remains useful for dialogs and indexing progress.

The native frontendless launcher is intentionally limited to IDEA Ultimate baseline 262. Other managed products/build lines currently use their standard launcher and may still need a display/Xvfb.

Next Steps