{"_id":"@coral-ai-labs/installer","_rev":"2-f6c7827a43534eba057103b0b3752353","name":"@coral-ai-labs/installer","dist-tags":{"latest":"0.1.1"},"versions":{"0.1.0":{"name":"@coral-ai-labs/installer","version":"0.1.0","license":"Apache-2.0","_id":"@coral-ai-labs/installer@0.1.0","maintainers":[{"name":"caelum119","email":"caelum119@gmail.com"}],"bin":{"coral-code-install":"bin/coral-code-install.js"},"dist":{"shasum":"ba274d569e972663de810c7d5b24f2645de4cf7e","tarball":"https://registry.npmjs.org/@coral-ai-labs/installer/-/installer-0.1.0.tgz","fileCount":3,"integrity":"sha512-yo+CHsnM4rMkptJeaYLuu+xwZ2uI3g5WwF1XN3JhGwJC0B8LyVVnfGBWszENV4auIBiKSm5lz01B4LGJrf+C5Q==","signatures":[{"sig":"MEYCIQD2rWvUiXNIIzQFqziVCA12ymqmMeBmlLBKzcXZI+VGgAIhAOp6UbqN+7ymQ4eFR/Tewf35LNqkTtDwrqr2jB3qWGIA","keyid":"SHA256:DhQ8wR5APBvFHLF/+Tc+AYvPOdTpcIDqOhxsBHRwC7U"}],"unpackedSize":10961},"type":"commonjs","engines":{"node":">=18"},"gitHead":"9e83ccbe9ffe493fc3c9957d7cf3285621bab93e","private":false,"_npmUser":{"name":"caelum119","email":"caelum119@gmail.com"},"_npmVersion":"11.5.1","description":"Cross-platform npm entrypoint for installing coral-code via the official installer scripts.","directories":{},"_nodeVersion":"24.7.0","preferGlobal":true,"publishConfig":{"access":"public"},"_hasShrinkwrap":false,"_npmOperationalInternal":{"tmp":"tmp/installer_0.1.0_1785112948384_0.9528451243440466","host":"s3://npm-registry-packages-npm-production"}},"0.1.1":{"name":"@coral-ai-labs/installer","version":"0.1.1","description":"Cross-platform npm entrypoint for installing coral-code via the official installer scripts.","license":"Apache-2.0","type":"commonjs","private":false,"publishConfig":{"access":"public"},"preferGlobal":true,"bin":{"coral-code-install":"bin/coral-code-install.js"},"engines":{"node":">=18"},"_id":"@coral-ai-labs/installer@0.1.1","gitHead":"da8f5a3e437757bdcd62101bf7ae9a51e76be016","_nodeVersion":"24.7.0","_npmVersion":"11.5.1","dist":{"integrity":"sha512-fgKBS/E4ijoCsi1z/IV3dGdKVRQLfY8g7Q6N5NrqEJGr69BAfV6MErASyXfPPbE4g4JJ9A0xwXQMJPgHfHn1zQ==","shasum":"3d357cf80e38c7fcd40921ac8047d88c1bfd8a77","tarball":"https://registry.npmjs.org/@coral-ai-labs/installer/-/installer-0.1.1.tgz","fileCount":3,"unpackedSize":11174,"signatures":[{"keyid":"SHA256:DhQ8wR5APBvFHLF/+Tc+AYvPOdTpcIDqOhxsBHRwC7U","sig":"MEUCIAPDxWEmxMuqZnj5a8LC+J4U6uIfR3ZN+TLWpTmNMT4vAiEAgfY15EDlXKgyL71NFNZgvsjmTwE2jZQfy6gioJlPUlU="}]},"_npmUser":{"name":"caelum119","email":"caelum119@gmail.com"},"directories":{},"maintainers":[{"name":"caelum119","email":"caelum119@gmail.com"}],"_npmOperationalInternal":{"host":"s3://npm-registry-packages-npm-production","tmp":"tmp/installer_0.1.1_1785113543251_0.8866242477045752"},"_hasShrinkwrap":false}},"time":{"created":"2026-07-27T00:42:28.193Z","modified":"2026-07-27T00:52:23.558Z","0.1.0":"2026-07-27T00:42:28.517Z","0.1.1":"2026-07-27T00:52:23.401Z"},"license":"Apache-2.0","description":"Cross-platform npm entrypoint for installing coral-code via the official installer scripts.","maintainers":[{"name":"caelum119","email":"caelum119@gmail.com"}],"readme":"# coral-code\n\n[coral-code](https://coralcode.dev/) is a plugin for JetBrains IDEs that gives coding agents a CLI/skill for leveraging and becoming part of agent graphs. These agent graphs perform _much_ better (outcome/benchmark score, speed, cost*).\n\nIt also brings a nice UI for ergonomically interacting with these agents and their graphs.\nIt is intended to be easy to get started with, but primarily accommodate professional developers.\n\n\\* coral-code is focused on solving hard coding problems on large codebases. It is cheaper and more effective compared with using the largest models alone, but it doesn't beat the cost of deepseek alone for easier coding problems (what does?). The performance ceiling of course grows the better the models used too.\nSee [Performance](#performance) for more details.\n\n\n\n---\n\n## Getting started\n1. Install coral-code with npm\n\nUse one command on macOS, Linux, or Windows:\n\nThe installer will download the latest version of the plugin and install it into your selected JetBrains IDEs.\n\n```bash\nnpx @coral-ai-labs/installer\n```\n\n2. The installer will automatically detect your installed JetBrains IDEs. Choose the IDEs you wish to install the plugin into.\n\n3. Follow the prompts and complete installation.\n\n4. Open the 'coral-code Chat' tool window in your IDE. It should be on the right side of the IDE by default. If you don't see it, you can open it from the 'View' menu: `View > Tool Windows > coral-code Chat`.\n\n5. (Optional) Setup agent graph functionality by connecting an API key. Coral Cloud offers a free tier for inner agents.\n\n6. Give it your hardest prompts! You can configure agent graph features and some better IDE-integrated agent tools in the chat window's options by the send button.\n\n---\n\n\nCurrently, it only supports this configuration:\n* Codex \"outer\" agent (tailored for GPT 5.6 and 5.3)\n* DeepSeek \"inner\" agents\n* Codex \"Delegated\" agents\n\nThe \"outer\" agent actually could be any agent. We also have great raw performance results using mini-swe-agent, though you need to be a little bit insane to use that outside of a containerized environment.\nFor now, from the plugin's interface, we will just support the Codex \"outer\" agent, but we will add more options in the future.\n\n\n### Requirements\nWithout agent graphs:\n- JetBrains IDE (IntelliJ, PyCharm, WebStorm, etc.)\n- codex\n- Node.js 20+ (for the installer)\n\nWith agent graphs:\n- API keys:\n - coral-cloud account (recommended, free tier available)\n - OR DeepSeek (for inner agents)\n\n\n\n## Performance\nWe're currently evaluating the performance of coral-code on benchmarks.\n\nGenerally, we find that one of these outcomes occurs:\n- doesn't work in some way (e.g. a PSI tree is unavailable for a given build system)\n- a bug is encountered and the outer agent gives up using coral-code\n- the outer agent misunderstands the user's intent\n- **~performance is doubled on most coding benchmark tasks, larger codebases are better handled, written code is less blind of the rest of the codebase**\n\nIt is still sensible to use SOTA models on outer agents though, they have faster wall clock time and produce higher quality code. The user is encouraged to balance performance, time, and cost to their needs.\n\n### Early results\nWe can share that:\n* On [SWE Atlas QnA benchmark](https://labs.scale.com/leaderboard/sweatlas-qna), using deepseek v4 pro as the outer and inner agents, coral-code attains a provisional score of 57%. **More than double the 27% baseline** Complete ablation studies pending more tokens/time/money.\n\nIncomplete ablation results on hardest task of SWE Atlas QnA benchmark:\n* coral-code + deepseek v4 pro is able full score, outperforming Fable max alone at 1/10th of the cost\n* gpt 5.5 as the outer agent also full scores on this task, at half the wall clock time of deepseek v4 pro, and half the cost of Fable max.\n\nWe believe there is still even more room for improvement with horizontal agent graph scaling\n\n\n## Known issues\n* It is hard to balance letting the agent know when is best to use coral-code. During updates you may find it being biased against and for using it. You may rely on the chat hotkey to toggle its mode / enabled status for the current turn.\n* ? Probably many others. Please feel free to raise issues on this github page! Feedback is extremely welcome.\n\n\n## FAQ\n**Q: What's an outer agent? What's an inner agent?**\n\nAnswer
\nA: An outer agent is the agent that the user interacts with directly. It is responsible for understanding the user's intent, delegating work to inner agents, and final responsibility for the output. An inner agent is an agent that is spawned by the outer agent to perform a specific task. Inner agents are typically more specialized and have access to different tools and resources than the outer agent.\n \n\n---\n\n**Q: Is this production ready?**\n\nTo be honest, no...
\nIt is definitely useful and whenever it's been unavailable I have felt the pain of not having it, but there are still some rough edges. You can try it for free though.\n\nI'd just recommend not relying on its presence for critical work. If it gets good traction, we will be able to dedicate more resources to it and make it more production ready.\n \n\n---\n\n**Q: Is this safe to use on my codebase? How tested is the sandboxing? Which agent harnesses are used?**\n\nAnswer
\nA: It should be as safe or safer than codex alone. The inner agents are flat out unable to execute commands, they are koog-based agents that only have tools for interacting with files through Intellij's PSI tree. Matters of execution, checking online documentation, and other things that could be unsafe are left to the outer agents. \n \n\n---\n\n**Q: Does this use my codex subscription's quota?**\n\nAnswer
\nA: Yes, the outer agent(s) are Codex agent(s), so it uses your Codex subscription's quota. The inner agents are DeepSeek agents, they are separate. If you're authenticated to Coral Cloud, you get a free quota for inner agents and can pay for higher quotas.\n \n\n---\n\n**Q: Is this going to be inefficient on codex quota?**\n\nAnswer
\nA: If graph functionality is used, generally it gives the outer agent more input tokens to deal with. For difficult tasks this tends to be token saving, but for easier tasks it can be token expensive (though with better quality output, which long term saves costs again). You can easily toggle graph functionality between off, on (everyday) and 'high certainty' in the chat window.\n \n\n---\n\n\n**Q: Can inner agents also be Codex agents?**\n\nAnswer
\nA: Sort of. There may be 15-30 concurrent active inner agents per outer operation, and several outer operations. Codex isn't yet designed for that kind of concurrency. If you enable 'Delegation' in the settings, the inner agents are able to spawn Codex agents to delegate work to, respecting the configured codex concurrency limits. \n \n\n---\n\n\n**Q: Can I use coral-code in a standalone window like chatgpt desktop?**\n\nAnswer
\nA: Yes, open coral-code Chat and drag a tab down. It will become its own window, it stays above the IDE so you can drag it where it's not in the way.\n \n\n---\n\n**Q: Why isn't this just a skill or Codex plugin?**\n\nAnswer
\nA: coral-code relies on indexing and PSI trees to understand the codebase. It has its own layer of persistence built on top of JetBrains IDE's extremely high quality codebase processing. It technically could be considered \"just a skill\" that requires an IDE to run with it installed.\n \n\n---\n\n**Q: ...But why does it have its own chat interface?**\n\nAnswer
\nA: You can use codex's own chat interface, but there are a few reasons why coral-code's own chat could be more ergonomic:\n\n* For professional developers, deep IDE integration is important. Our thesis is that understanding a codebase is an investment, and the better the bandwidth for that understanding, the better the investment.\n* Codex offers great flexibility for displaying interactive UI, but there are a few UX optimizations that require making it a separate codex app server client. (e.g. chat area controls like custom access levels, per-turn agent graph settings etc.)\n* in future, this should work with more than codex. Having its own interface allows us to support more outer agents in the future, and even allow for multiple outer agents to be used at once (you can try this already, though it's untested)\n \n\n---\n\n**Q: What is coral cloud?**\n\nAnswer
\nA: coral cloud is a hosted service for graphs of agents. It's under construction, but in the future it will allow agent developers to share monetized agents for use in graphs, where the consumers control firewall and persistence. coral-code is reusing its inner agent graph functionality, and will be able to use coral cloud for inner agents in the future.\n \n\n","readmeFilename":"README.md"}