AgentBridge

A JetBrains IDE plugin that bridges AI coding agents to IntelliJ platform APIs through 120+ native MCP tools. Agents work through inspections, refactorings, the test runner, the build system, and Git — the same tools you use — instead of operating through a terminal or generating diffs in isolation.

Key highlights:

• 7 agents — Copilot, Claude Code, Codex, Kiro, Junie, OpenCode, Hermes Agent. Switch with one click.
• 120+ MCP tools — code navigation, refactoring, testing, debugging, git, project management, and more — all through native IntelliJ APIs.
• Cross-client session resume — switch agents without losing conversation history.
• PWA web access — use the chat from any device on your local network via HTTPS.
• Human in the loop — every action is visible in the IDE. Undo, review, or redirect at any point.

Why this approach?

AgentBridge is built around a simple idea: if a full JetBrains IDE helps humans work in a fail-fast way, agents should be able to use that same IDE feedback loop too. Instead of asking an LLM to approximate refactorings, code intelligence, inspections, Git state, and project context from text alone, AgentBridge lets the agent call deterministic IDE tools directly.

See The Case for IDE-Native Coding Agents for the longer reasoning behind this IDE-native approach and why I think it is worth exploring alongside lighter agent wrappers and harnesses.

Project status, collaborators, and reviews

AgentBridge is starting to feel feature-ready. Most current work is tuning behavior, polishing workflows, fixing rough edges, and following what agent CLIs expose through ACP so new capabilities can be integrated as they become available.

I would be happy to have collaborators, especially for code review. I think a healthy practice for agent-authored changes is to have at least two pairs of human eyes on important changes; right now I mostly review the changes I asked Copilot to make. Just ping me if you would like to help. I am also glad there are now actionable issues reported on GitHub. That feedback helps me prioritize work and confirms that the effort I spend on this project is worthwhile.

If you use the plugin and find it useful, an honest rating or review on the JetBrains Marketplace would help other users decide whether to try it. The plugin is approaching 2,000 downloads, and Marketplace update stats suggest hundreds of active installs, but I understand why someone might hesitate to install a plugin with no reviews.

Supported Agents

Agent	Protocol	Authentication
GitHub Copilot	ACP (stdin/stdout)	OAuth sign-in
Claude Code (claude CLI)	Stream JSON (stdin/stdout)	Anthropic auth
Codex (codex app-server)	ACP (stdin/stdout)	Device-auth sign-in
Kiro	ACP (stdin/stdout)	Amazon auth
Junie	ACP (stdin/stdout)	JetBrains auth
OpenCode	ACP (stdin/stdout)	Configurable (multi-provider)
Hermes Agent	ACP (stdin/stdout)	Configurable (multi-provider)

Switch between agents with one click. Each agent has its own connection settings, tool permissions, and custom instructions.

What Agents Can Do

Every action goes through IntelliJ platform APIs — nothing happens behind your back.

Code intelligence — Search symbols, find references, go to declaration, type hierarchy, call hierarchy, class outlines, documentation lookup. Agents navigate code using the IDE's semantic index, not text grep.

File operations — Read, write, create, delete, rename, move files through IntelliJ's Document API. Every write supports undo/redo and triggers auto-format + import optimization.

Code quality — Run inspections, apply quick-fixes, format code, optimize imports, suppress findings. Agents see the same warnings you see in the editor and can apply IDE-suggested fixes. Optional Qodana and SonarQube integration.

Refactoring — Rename, safe-delete, and PSI-aware symbol replacement through IntelliJ's refactoring engine. Also: replace_symbol_body, insert_before_symbol, insert_after_symbol for structural edits that survive reformatting.

Testing — Discover tests, run them in the IDE's test runner, collect coverage data. Results appear in the Run panel so you can watch them live.

Debugging — Set breakpoints (line and exception), step through code, inspect variables, evaluate expressions, read console output. Full interactive debugging through MCP tools.

Git — Status, diff, log, blame, commit, stage, branch, stash, push, pull, merge, rebase, cherry-pick, tag, reset, revert — all through the IDE's VCS layer with editor buffer awareness.

Project management — Build, run configurations, project structure, module dependencies, SDK management, indexing status, source directory marking.

Infrastructure — Shell commands, terminal I/O, HTTP requests, IDE log reading, modal dialog interaction, build/run output reading, user prompts.

Editor — Open files, show diffs, scratch files, conversation history search, theme management.

Diff Review — Always-on tracker for every agent edit. Persistent green/amber/red diff highlights, a Review panel listing each file with +N / −N line counts and last-edited timestamps, optional Auto-Approve toggle, structured revert nudges (with diff body) sent back to the agent, and automatic gating of destructive git operations (commit, merge, rebase, reset --hard, …) until pending changes are resolved. The list survives IDE restarts. See docs/INLINE-DIFF-REVIEW.md.

Nudge system — Send mid-turn guidance while the agent is working (Enter), force-stop and redirect (Ctrl+Enter), or queue follow-up messages (Ctrl+Shift+Enter).

PWA web access — Access the chat panel from any device on your local network via HTTPS with Web Push notifications. Self-signed certificate with a trust endpoint for easy device setup.

Database — List data sources, browse tables, inspect schemas. Query execution available as an experimental tool.

MCP Tool Hooks — Intercept any tool call with shell scripts at four lifecycle points: permission (gate), pre (modify arguments), success (transform output), and failure (recover or augment errors). Hooks can chain, override error state, and enforce policies like bot identity for commits and GitHub API calls. Static prependString/appendString fields allow simple text modifications without scripts. Hot-reloads within 2 seconds. See docs/MCP-TOOL-HOOKS.md.

See FEATURES.md for the complete tool reference.

IDE Compatibility

All 120+ tools work across every JetBrains IDE, with a few exceptions for IDE-specific capabilities:

Java-only tools (require com.intellij.modules.java)

Available in IntelliJ IDEA (Ultimate and Community). Not available in WebStorm, PyCharm, GoLand, PhpStorm, RubyMine, CLion, RustRover, or Rider.

Tool	Why Java-only
build_project	Triggers JPS incremental build
get_class_outline	Resolves fully-qualified Java/Kotlin class names

get_call_hierarchy, find_implementations, edit_project_structure, and get_type_hierarchy (subtypes with file+line) all work across every IDE using platform-level PSI APIs. get_type_hierarchy supertypes and symbol-only lookup still require Java.

Rider-disabled tools

Disabled in Rider because C#/C++ PSI lives in the ReSharper backend, not the IntelliJ frontend. These tools depend on fine-grained PSI or test framework infrastructure that Rider doesn't expose to the IntelliJ layer.

Thanks to Reddit user VirusPanin for discovering these limitations by testing AgentBridge in Rider with C++ code.

Tool	Why disabled
search_symbols	classifyElement() fails on Rider's coarse PSI stubs
replace_symbol_body	PSI symbol resolution too coarse for structural edits
insert_before_symbol	PSI symbol resolution too coarse for structural edits
insert_after_symbol	PSI symbol resolution too coarse for structural edits
list_tests	IntelliJ TestFramework extensions don't cover Rider's NUnit/xUnit backend
run_tests	ConfigurationContext can't resolve Rider test runners from the frontend PSI

Conditional tools (all IDEs)

Tool(s)	Condition
database_list_sources, database_list_tables, database_get_schema	Database plugin installed (bundled with Ultimate editions)
run_sonarqube_analysis, get_sonar_rule_description	SonarQube for IDE plugin installed
run_qodana	Qodana plugin installed
memory_* (9 tools)	Memory feature enabled in AgentBridge settings

Summary

IDE	Tools available
IntelliJ IDEA	All 120+
WebStorm, PyCharm, GoLand, PhpStorm, RubyMine, CLion, RustRover	~118 (no Java-only tools)
Rider	~112 (no Java-only + no Rider-disabled tools)

Architecture

graph TD
    subgraph IDE["JetBrains IDE"]
        subgraph ui["UI Layer"]
            TW["Tool Window (JCEF Chat)"]
        end
        subgraph services["Services"]
            AAM["ActiveAgentManager"]
            APM["AgentProfileManager"]
        end
        subgraph clients["Agent Clients"]
            AC["AcpClient — Copilot, Junie, Kiro, OpenCode, Hermes Agent"]
            CC["ClaudeCliClient — Claude Code"]
            XC["CodexAppServerClient — Codex"]
        end
        subgraph mcp["MCP Layer"]
            PSI["PsiBridgeService (HTTP ↔ 120+ tools)"]
        end
        TW --> AAM
        AAM --> clients
    end

    AC <-->|" stdin/stdout — JSON-RPC 2.0 "| CLI["Agent CLI"]
    CC <-->|" stdin/stdout — stream-json "| CCLI["Claude Code CLI"]
    XC <-->|" stdin/stdout — JSON-RPC 2.0 "| CDEX["Codex app-server"]
    CLI -->|" stdio "| MCP["MCP Server (bundled JAR)"]
    MCP -->|" HTTP "| PSI

Three layers:

• UI — JCEF-based chat panel, model selector, context management. Agent-agnostic.
• Clients — AcpClient (shared by Copilot, Junie, Kiro, OpenCode, Hermes Agent), ClaudeCliClient, CodexAppServerClient. Each wraps its agent CLI's stdin/stdout protocol.
• MCP — 120+ tools implemented against IntelliJ's PSI, VFS, and platform APIs. Exposed via an HTTP bridge to a bundled MCP stdio server (JAR).

Cross-Client Session Resume

Conversations are stored in a canonical V2 format (JSONL). When you switch agents, the session is exported to the new agent's native format — preserving full conversation history across all six supported agents. See docs/SESSION-RESUME.md.

IntelliJ-Native File Operations

Built-in agent file tools (like bash, edit, create) are denied at the permission level. The agent retries using AgentBridge MCP tools, which write through IntelliJ's Document API — supporting undo/redo, triggering auto-format, and keeping the editor in sync.

Requirements

• JetBrains IDE 2025.3+ (IntelliJ IDEA, WebStorm, PyCharm, GoLand, Rider, CLion, RubyMine, PhpStorm, or any other JetBrains IDE)
• A supported agent CLI — at least one of: GitHub Copilot CLI, claude, codex, Kiro, Junie, or OpenCode

For development: JDK 21 and Gradle 8.x (wrapper included).

Quick Start

Install from Disk

Build the plugin, then install it:

./gradlew :plugin-core:buildPlugin

In your IDE: Settings → Plugins → ⚙ → Install Plugin from Disk → select plugin-core/build/distributions/plugin-core-*.zip.

Restart the IDE and open the AgentBridge tool window.

Development Sandbox

For development, run the plugin in a sandboxed IDE instance:

./restart-sandbox.sh

See DEVELOPMENT.md for the full development guide.

Running Tests

./gradlew test

Releases & Updates

How versioning works

Every merge to master triggers an automated release. The CI pipeline analyzes conventional commit messages to determine the version bump:

Commit prefix	Bump	Example
feat:	Minor (1.52.0 → 1.53.0)	New tool or capability
fix:, refactor:, docs:, etc.	Patch (1.53.0 → 1.53.1)	Bug fix, cleanup
feat!: or BREAKING CHANGE	Major (1.53.1 → 2.0.0)	Breaking change

Each release produces a signed ZIP artifact on the GitHub Releases page with provenance attestation.

Standard vs Experimental builds

Each release builds two plugin ZIPs:

• agentbridge-<version>.zip — the standard build, suitable for the JetBrains Marketplace.
• agentbridge-experimental-<version>.zip — includes additional tools that use internal JetBrains APIs or experimental designs. This build is not published to the Marketplace because internal APIs are unsupported, may break across IDE versions, and are flagged during Marketplace review.

JetBrains Marketplace

The Marketplace is the recommended installation channel — your IDE checks for updates automatically.

Marketplace submissions require a review by JetBrains (typically a few business days), so not every incremental release is pushed there. Instead, we publish to the Marketplace periodically and tag that commit as marketplace-latest. Each subsequent GitHub release accumulates a changelog of all changes since the last Marketplace publish, so users can see exactly what's new.

Getting updates before the Marketplace

If you want to test a specific fix or feature that hasn't reached the Marketplace yet:

1. Go to GitHub Releases and download the agentbridge-<version>.zip for the version you want.
2. In your JetBrains IDE: Settings → Plugins → ⚙ → Install Plugin from Disk → select the downloaded ZIP.
3. Restart the IDE.

When the next Marketplace update arrives, it will overwrite the manually installed version and you'll be back on the automatic update track.

Tags

Tag	Purpose
v<major>.<minor>.<patch>	Every release (created automatically by CI)
marketplace-latest	Points to the commit currently published on the JetBrains Marketplace

Project Structure

agentbridge/
├── plugin-core/           # Main plugin — UI, agent clients, MCP tools (Java 21)
├── mcp-server/            # Bundled MCP stdio server (JAR)
├── plugin-experimental/   # Experimental tools (run_inspections, execute_query)
└── integration-tests/     # Integration test suite

Documentation

Document	Description
QUICK-START.md	Fast setup instructions
FEATURES.md	Complete MCP tool reference
DEVELOPMENT.md	Build, deploy, architecture, extending for new agents
INSTALLATION.md	Detailed installation for all platforms
TESTING.md	Test running and coverage
Releases	Per-version changelog and downloads
docs/SESSION-RESUME.md	Cross-client session migration
docs/PERMISSIONS.md	Per-agent tool permission architecture

Contributing

Contributions are welcome. See CONTRIBUTING.md for guidelines.

Security

To report security vulnerabilities, see SECURITY.md.

License

Copyright 2026 Henrik Westergård

Licensed under the Apache License, Version 2.0.

---

AgentBridge is an independent open-source project. It is not affiliated with or endorsed by GitHub, Anthropic, Amazon, JetBrains, or OpenAI.