Prompt

IntelliJ API Power User Guide

RECOMMENDED: Execute Kotlin code directly in IntelliJ IDEA's runtime with full access to IntelliJ Platform APIs.

MCP Steroid - IDE API Access for AI Agents

Execute Kotlin code directly in IntelliJ IDEA's runtime with full access to the IntelliJ Platform API.

Important Notes for AI Agents

Learning Curve: Writing working code for IntelliJ APIs may require several attempts. This is normal! The API is vast and powerful. Keep trying - each attempt teaches you more about the available APIs. Use printException() to see stack traces when errors occur.

Drop-in replacement for LSP: This MCP server replaces LSP (Language Server Protocol) tools with IntelliJ's native APIs — same operations, deeper understanding:

• PSI (Program Structure Interface) instead of LSP document symbols — full semantic analysis
• IntelliJ inspections, refactorings, intentions instead of LSP code actions
• Full project model with module dependencies instead of workspace folders
• Platform-specific indices for O(1) code search instead of filesystem scans

Common task → resource cheat sheet

Before reading further, if your task matches one of these, skip straight to the linked recipe:

The full index is in the "MCP Resources (Use Them)" section below.

Quickstart Flow

1. steroid_list_projects → get list of open projects
2. Empty on a clean machine? Run `devrig backend download --json`, install the required product, then
   call steroid_open_project. It auto-starts a sole installed backend; no separate start is required.
3. Poll steroid_list_projects until the path appears, then pick its `project_name` (the unique routing
   key — NOT the human-readable `name`). Do not reuse a key from before an IDE/project restart.
4. On a first Maven/Gradle open, fetch `mcp-steroid://skill/execute-code-maven` or
   `mcp-steroid://skill/execute-code-gradle`, then trigger and await external-system configuration exactly
   as that recipe shows.
5. steroid_execute_code → run Kotlin code with that project_name
6. steroid_execute_feedback → report success/failure for tracking

For unattended Java/JVM work on 2026.2, install IDEA Ultimate with a pinned 2026.2 version. devrig launches it as a frontendless Remote Development backend with MCP Steroid included, so a window or screenshot is not a readiness requirement. Full lifecycle details: mcp-steroid://open-project/managing-backends (fetch it after steroid_list_projects supplies a project_name).

Each steroid_list_projects entry has TWO name fields: project_name (the within-IDE-unique, opaque routing KEY you pass back to every project-scoped tool) and name (the human-readable folder name, informational/display only). Always route by project_name. To find the right project for a file or directory path, pick the project whose path is the longest prefix of your target path (this disambiguates nested checkouts and git worktrees).

Example session:

→ steroid_list_projects
← {"projects":[{"project_name":"my-app-9fk2a0xq","name":"my-app","path":"/path/to/my-app","backend_name":"iu-9fk2a0xq"}],
   "backends":[{"backend_name":"iu-9fk2a0xq","intellij":{"name":"IntelliJ IDEA 2026.1.3","version":"2026.1.3","build":"IU-261.25134.95"}}]}

→ steroid_execute_code(project_name="my-app-9fk2a0xq", code="println(project.name)", ...)
← "my-app"

→ steroid_execute_feedback(project_name="my-app-9fk2a0xq", task_id="...", execution_id="...", success_rating=1.0, explanation="Got project name")

When to Use This Skill

ALWAYS prefer IntelliJ APIs over file-based operations:

The IDE has indexed everything. It knows the code better than any file search.

Available Tools

steroid_list_projects

List all open projects. Each entry has project_name (the unique routing key to pass to steroid_execute_code and the other project-scoped tools), name (the human-readable folder name, informational only), and path. To map a file/dir path to a project, pick the project whose path is the longest prefix of your target path.

steroid_list_windows

List open IDE windows (and background tasks) and their associated projects. Some windows may not be tied to a project and a project can have multiple windows. Use this in multi-window setups to pick the correct window_id for screenshot/input tools. Each window and task entry references its project by project_name — the single routing key. Look up that project's human-readable name and path via steroid_list_projects by that key (they are not duplicated on window/task entries).

steroid_take_screenshot

Capture a screenshot of the IDE frame and return image content.

HEAVY ENDPOINT: Use only for debugging and tricky configuration. Prefer steroid_execute_code for regular automation.

Parameters:

• project_name (required): the project_name from steroid_list_projects (a unique routing key, NOT the raw folder name)
• task_id (required): Task identifier for logging
• reason (required): Why the screenshot is needed
• window_id (optional): Window id from steroid_list_windows to target a specific window

Artifacts (saved under the execution folder):

• screenshot.png
• screenshot-tree.md
• screenshot-meta.json

The response includes window_id (also stored in screenshot-meta.json); pass it to steroid_input to target the same window. window_id is also returned by steroid_list_windows.

steroid_input

Send input events (keyboard + mouse) using a sequence string.

HEAVY ENDPOINT: Use only for debugging and tricky configuration. Prefer steroid_execute_code for regular automation.

Parameters:

• project_name (required): the project_name from steroid_list_projects (a unique routing key, NOT the raw folder name)
• task_id (required): Task identifier for logging
• reason (required): Why the input is needed
• window_id (required): Window id from steroid_list_windows (also returned by steroid_take_screenshot)
• sequence (required): Comma-separated or newline-separated input sequence (commas inside values are allowed unless they look like , <step>:; commas are optional when using newlines)

Sequence examples:

• stick:ALT, delay:400, press:F4, type:hurra
• click:CTRL+Left@120,200
• click:Right@screen:400,300

Notes:

• Comma separators are detected by , <step>: patterns, so avoid typing , delay: etc in text.
• Trailing commas before a newline are ignored.
• Use # for comments until the end of the line.
• Targets default to screenshot coordinates; use screen: for absolute screen pixels.
• Input focuses the screenshot window before dispatching events.

steroid_execute_code

Execute code with IntelliJ's brain, not just text files.

Give your AI agent a senior developer's toolkit: semantic code understanding, automated refactorings, and IDE intelligence that LSP can't provide.

Parameters: project_name, code (Kotlin suspend function body), task_id, reason, timeout (optional)

Returns: execution_id plus ONLY what the script explicitly prints (println / printJson / printCsv / printToon). The last expression's value is ignored by the runtime — a script that computes but never prints returns no data, just a HINT: about the missing print.

Complete guide: mcp-steroid://skill/coding-with-intellij (API reference, patterns, examples, best practices)

steroid_execute_feedback

Rate execution results. Use after steroid_execute_code.

steroid_open_project

Open a project in the IDE. A managed backend cold start may block until MCP is reachable; only the project-open phase is asynchronous.

Through devrig, a sole installed managed backend is selected and started automatically; backend startup waits until MCP is reachable. The project-open request itself remains asynchronous.

Parameters:

• project_path (required): Absolute path to the project directory to open
• task_id (required): Task identifier for logging
• reason (required): Why you are opening the project
• trust_project (optional): If true, trust the project path before opening (skips trust dialog). Default: true

Workflow:

1. Call steroid_open_project with the project path
2. If trust_project=true, the project will be trusted automatically (no trust dialog)
3. Poll steroid_list_projects until the path appears and keep its fresh project_name
4. If a frontend exists, use steroid_list_windows; only use screenshot/input when it reports a modal
5. A frontendless Remote Development backend has no required window. For a first Maven/Gradle open, await the build-system model using its execute-code sync recipe before indexed semantic queries

MCP Resources (Use Them)

This server exposes built-in articles through steroid_fetch_resource. These are the fastest way to load full examples and guides without guessing or copy/pasting from the web.

How to access resources:

1. Call steroid_fetch_resource with a mcp-steroid:// URI to load the content.

The fetch tool is project-scoped because articles render for the target IDE. On a clean machine, use the initialize instructions and open-project description to bootstrap; fetch articles after a project appears.

Key resources provided by this server:

• mcp-steroid://prompt/skill - This guide as a resource.
• mcp-steroid://skill/coding-with-intellij - Comprehensive guide for writing IntelliJ API code (execution model, patterns, examples).
• mcp-steroid://prompt/debugger-skill - Debugger-focused skill guide (breakpoints, sessions, threads).
• mcp-steroid://lsp/overview - Overview of LSP-like examples and how to use them.
• mcp-steroid://lsp/<id> - Runnable Kotlin scripts (e.g., go-to-definition, find-references, rename, code-action, signature-help).
• mcp-steroid://ide/overview - Overview of IDE power operation examples (refactorings, inspections, generation).
• mcp-steroid://ide/<id> - Runnable Kotlin scripts (e.g., extract-method, introduce-variable, change-signature, safe-delete, optimize-imports, pull-up-members, push-down-members, extract-interface, move-class, generate-constructor, call-hierarchy, project-dependencies, inspect-and-fix, inspection-summary, find-duplicates, project-search, run-configuration).
• mcp-steroid://debugger/overview - Overview of debugger examples (breakpoints, sessions, threads).
• mcp-steroid://debugger/<id> - Runnable Kotlin scripts (e.g., set-line-breakpoint, debug-run-configuration, debug-session-control, debug-list-threads, debug-thread-dump).
• mcp-steroid://open-project/overview - Guide for opening projects via MCP.
• mcp-steroid://open-project/<id> - Project opening examples (e.g., open-trusted, open-with-dialogs, open-via-code).
• mcp-steroid://open-project/managing-backends - devrig download, auto-start, Remote Development, and readiness phases.

These resources are designed to be plugged directly into steroid_execute_code after you configure file paths/positions.

Critical Rules

These are the essential rules you must follow. For detailed examples and patterns, read mcp-steroid://skill/coding-with-intellij.

1. Script Body is a SUSPEND Function

// This is a coroutine - use suspend APIs!
// Under the default modal=smart_non_modal, waitForSmartMode() runs automatically before your script.
delay(1000)         // coroutine delay - works directly

NEVER use runBlocking - it causes deadlocks.

NEVER re-probe waitForSmartMode() before every operation. The automatic wait (under the default modal=smart_non_modal; skipped under non_modal / unleashed) before script start is only a point-in-time check; IntelliJ may enter dumb mode again before the next statement. For index-dependent PSI queries, wrap the whole query in smartReadAction { }. After project open/import/sync/configuration, first await Observation.awaitConfiguration(project), then use smartReadAction { }.

2. Imports Are Optional

Default imports are provided automatically. Add imports only when you need APIs outside the defaults. Imports must be at the top of the script, never after code.

3. Read/Write Actions for PSI/VFS

THREADING RULE — NEVER SKIP: Any PSI access MUST be inside readAction { }. Modifications require writeAction { }. Threading violations throw immediately.

Built-in helpers (no imports needed):

// Reading PSI/VFS/indices
val data = readAction { project.name }

// Modifying PSI/VFS/documents
writeAction { /* modifications here */ }

// Runs index-dependent PSI work under IntelliJ's smart-mode read constraint
val smart = smartReadAction { /* PSI operations */ }

For detailed threading patterns, see mcp-steroid://skill/coding-with-intellij-threading.

4. Context API

Built-in helpers available in every script (no imports needed):

The output methods are the only way to get data back to the agent — the script's last-expression value is ignored by the runtime, so print everything you need.

Tabular output cheat sheet — for find-references, call-hierarchy, project-search, document-symbols, or any flat array-of-records result. Signatures are different on purpose; printCsv takes parallel lists (positional), printToon takes a list of maps (keyed). Mixing them up is the #1 first-try compile error.

// CSV — printCsv(headers: List<String>, rows: Iterable<List<Any?>>, dictColumns: Set<String> = emptySet())
// Best when one column has repeated long values (absolute paths, FQNs).
// `dictColumns` emits a per-column @col: preamble and replaces each cell with a short ID (`p1`, `p2`, ...).
printCsv(
    headers = listOf("idx", "path", "line"),
    rows = emptyList<List<Any?>>(),
    dictColumns = setOf("path"),
)

// TOON — Token-Oriented Object Notation; https://github.com/toon-format/toon.
// printToon(value: Any?) — drop-in for printJson on any value; cheapest on uniform-shape lists.
// Pass List<Map<String, Any?>>. Same key set in every map -> compact [N]{cols}: form, column order
// from the FIRST map's keys. Different key sets are ACCEPTED too -> less compact per-element block
// ([N]: plus `key: value` lines), so never normalize naturally ragged records just to get the
// compact form. Do NOT pass headers / rows / dictColumns — that's printCsv.
printToon(listOf(mapOf("path" to "/abs/A.kt", "line" to 17), mapOf("path" to "/abs/B.kt", "line" to 42)))

Same records, both formats — most recipes finish by emitting one list of records twice:

val records: List<Triple<String, Int, String>> = emptyList()  // built once
printCsv(
    headers = listOf("idx", "path", "line", "snippet"),
    rows = records.mapIndexed { i, (p, l, s) -> listOf(i + 1, p, l, s) },
    dictColumns = setOf("path"),
)
printToon(records.map { (p, l, s) -> mapOf("path" to p, "line" to l, "snippet" to s) })

Full API reference with literal sample outputs and an end-to-end example: mcp-steroid://skill/coding-with-intellij-context-api → "Tabular Output".

5. Running Inspections

The IDE has hundreds of inspections — DuplicatedCode, RedundantCast, UnusedDeclaration, language-specific DFA, etc. Two paths from a script:

Pitfall — do not printJson(result) from runInspectionsDirectly. The result is Map-compatible for legacy callers, but each ProblemDescriptor carries live PSI/VFS objects and can recurse through Jackson. Snapshot the descriptors under a read action, include failedTools, and compute an explicit status:

val vf = findProjectFile("src/main/java/com/example/Foo.java") ?: error("file not found")
val result = runInspectionsDirectly(vf)

val findings = readAction {
    result.entries.flatMap { (toolId, descriptors) ->
        descriptors.map { descriptor ->
            mapOf(
                "toolId" to toolId,
                "message" to descriptor.descriptionTemplate,
                "elementText" to (descriptor.psiElement?.text ?: "")
            )
        }
    }
}

val status = when {
    result.failedTools.isNotEmpty() -> "check_failed"
    findings.isNotEmpty() -> "findings"
    else -> "clean"
}

printJson(mapOf("status" to status, "findings" to findings, "failedTools" to result.failedTools))

failedTools is the guard against false clean results: when the sweep cannot run for an input file, or an inspection tool crashes, the findings map may be empty but failedTools is non-empty.

6. Running Tests

Always prefer the IntelliJ IDE runner over ./mvnw test or ./gradlew test. The IDE runner returns a simple exit code (0 = all passed), shows structured results, and reuses the running JVM.

See mcp-steroid://skill/coding-with-intellij → "Run Tests via IntelliJ IDE Runner" for the complete pattern.

Only fall back to CLI test commands when the IDE runner cannot be used. Even then, never print the full output — always take(30) + takeLast(30) to avoid MCP token limit errors.

Error Handling

Use printException for errors - it includes the stack trace in the output:

try {
    // risky operation
} catch (e: Exception) {
    printException("Operation failed", e)
}

Troubleshooting

Check if Server is Running

The MCP server runs inside IntelliJ. To verify:

1. Open IntelliJ IDEA with the MCP Steroid plugin installed
2. Open any project
3. Check .idea/mcp-steroid.md in the project folder for the server URL
4. The server port is configurable via mcp.steroid.server.port; read .idea/mcp-steroid.md for the active URL

Endpoints

• / - Returns this SKILL.md content
• /skill.md - Same as above
• /mcp - MCP protocol endpoint for tool calls
• /.well-known/mcp.json - MCP server discovery

Fetching mcp-steroid:// articles (preferred)

Use the steroid_fetch_resource MCP tool (it requires project_name for correct IDE-conditional rendering) instead of HTTP fetching or ReadMcpResourceTool. The articles are NOT exposed via resources/list or prompts/list — the tool is the canonical discovery surface.

Common Issues

• "Project not found" - Run steroid_list_projects first to get exact project names
• No output from execute - Only printed values come back; the last expression's value is ignored by the runtime. End the script with println() / printJson() of everything you need
• Timeout - Increase timeout parameter (default 60 seconds)
• Script errors - Check Kotlin syntax; imports are optional

Detailed Guides

For API examples, patterns, and in-depth coverage, read the dedicated articles:

Other Resources

• Debugger Skill Guide - Debug workflows and stateful execution
• Test Runner Guide - Test execution patterns
• LSP Examples - LSP-like operations (navigation, code intelligence, refactoring)
• IDE Examples - IDE power operations (refactorings, inspections, generation)
• Debugger Examples - Debugger workflows and API usage
• Test Examples - Test execution and result inspection
• VCS Examples - Version control operations (git blame, history)
• Open Project Examples - Project opening workflows

This is like LSP, but more powerful. IntelliJ APIs offer deeper code understanding and more features than standard LSP. Don't settle for file-level operations when you have IDE-level access.