Sigil is a coding assistant plugin for IntelliJ that integrates with the IDE instead of just wrapping the CLI. It's powered by OpenCode.
Most coding assistant plugins I tried fell into one of two camps: a thin CLI wrapper with no real IDE integration, or a bloated plugin that pushes its own workflow and ideas on you. This one tries to stay out of the way.
The bigger problem I have with coding assistants is that they tend to take the wheel. They write code, you accept it, and you move on. That makes it easy to ship stuff you don't fully understand. This plugin focuses on review as a first-class step - you can review code the agent wrote inline in the editor before anything reaches GitHub or your CI. The goal is to catch mistakes early and make sure the code you commit actually meets your standards.
This plugin was built almost entirely with open-weight LLMs - GLM 5.2, Kimi 5.6, Mimo 2.5, and DeepSeek V4. No Claude, no GPT-5, no closed-model pair programmer behind the scenes. The code, was written, reviewed, and iterated through that pipeline.
I did this as a deliberate experiment: can you build a real, shipping IntelliJ plugin using only open weights? Not a toy demo, but something with SSE streaming, MCP integration, editor gutter comments, async filesystem watchers, and a Compose UI. The answer so far is "mostly, and it's getting closer every month." Some parts took more retries than they would have with a closed model, and some prompts needed tighter scaffolding, but the result is a working plugin that I actually use day-to-day.
If you're curious how open weights compare for practical software work, this repo is one data point. Your mileage may vary, but it's a lot more viable than it used to be.
The main chat interface with sessions on the left and the conversation in the middle. Tool pills and markdown responses show up inline.
- AI chat inside the IDE - Ask questions, paste code, attach files, and get answers in a side panel. Responses render markdown, code blocks, and inline tool usage.
- Sessions - Keep conversations organized. Each session tracks its own token usage and estimated cost.
- Context panel - See how much context you're using, what model is active, and a rough cost breakdown.
- Code review - The plugin can request reviews from the OpenCode server and show the results both in a dedicated Review tab and as inline comments in the editor gutter. You can resolve, reply to, or delete review comments directly from the editor.
- Slash commands - Type
/in the input to trigger commands like/clearand/cancel, plus any commands exposed by the connected OpenCode server. - Tool permissions and MCP support - Connect to MCP servers and pick which tools the agent is allowed to call. There's also built-in support for registering IntelliJ's own bundled MCP server, so the agent can call IDE actions and tools.
Session list with titles, age, cost, and token usage per conversation.
Context tab showing token breakdown, active model, and total cost.
The review feature is one of the more useful parts. There are a few commands that drive it:
/review— asks OpenCode to look over the current change and write review comments to.review/./review-perform— runs a stricter, adversarial review that goes deeper into potential issues./review-perform-gaming— same as/review-perform, but tuned for game-engine-specific code./review-resolve— sends all open comments to the agent and asks it to fix the code and mark them resolved./review-recheck— re-runs the review with awareness of existing comments and your replies, so the agent can verify whether unresolved issues were actually fixed.
When a review finishes, you get:
- A Review tab that lists files, comment counts, and line change summaries.
- Inline comments in the editor gutter that you can resolve, reply to, delete, or add your own comments for the AI to address.
- A chat message summarizing what the review found.
You can also add your own comments by right-clicking a line and choosing Add Review Comment:
Right-click a line in the editor and choose "Add Review Comment" to drop in your own comment.
Review tab listing files with comment counts and change totals.
A single review comment in the editor gutter with resolve, delete, and reply options.
Multiple AI review comments overlaid directly on the source code.
The review workflow stores comments as JSON files under .review/, one file per reviewed source path. Comments track a line range, message, severity, and status. The plugin reads these files to render the Review tab and inline gutter comments, and writes them back when comments are resolved or deleted.
Because the comments live in the repo as regular files, they can be committed alongside the code. That means a branch can carry its own review state, and another developer (or an LLM agent) can check out the branch, act on open comments, and resolve them. The OpenCode server has no native review concept, so the plugin uses the filesystem instead of adding server-side storage or new endpoints. This keeps the review protocol simple: anything that can read and write JSON files in .review/ can participate.
This plugin uses Jewel, JetBrains' Compose-based UI toolkit for IntelliJ plugins. Jewel is still evolving, so some UI details might shift or break with future IntelliJ or Jewel updates. If something looks off after an update, that's probably why.
The chat UI could have been built with raw Swing, but that meant writing custom paint code, fighting BasicComboBoxUI, and manually keeping up with every IntelliJ theme change. Jewel is bundled with the IntelliJ Platform and bridges directly into the IDE's theme system, so the plugin looks at home in Darcula, New UI, or whatever theme the user has picked. The UI code is declarative Compose instead of hand-rolled Swing, which makes it easier to iterate on. The tradeoff is that Jewel's APIs are still moving, and it adds some runtime overhead, but for a chat tool window the integration win outweighs the churn.
- IntelliJ IDEA 2025.2 or newer.
- The
opencodeCLI installed and available on your system PATH, or configured in the plugin settings.
Download the latest ZIP from the GitHub releases page, then install it through Settings → Plugins → Install Plugin from Disk.