Check the answer. Then make the change.
Send product feedback through the issue tracker.
CoralCode
Install
CoralCode
Install
A missing dependency can send you back into the investigation. Coral Code brings findings from relevant files and recorded execution paths to Codex, Claude Code or Junie, so you can check the answer before changing the code.
You still have to trace callers and check whether an answer covers the behaviour you asked about. When your coding agent misses a connection, you reopen the investigation and supply the missing context yourself.
Coral Code uses project structure and recorded runtime paths to give agents focused areas to investigate. They exchange findings and return source references and unchecked questions to your coding agent.
Or import a coverage profile. The recorded execution path identifies files involved in the behaviour you want to investigate.
File agents investigate their assigned code and ask agents responsible for connected files for relevant findings.
Follow the source references, see what agents checked and review the questions they could not resolve.
Your coding agent uses the findings to answer or edit. You run the checks and decide what ships. JetBrains Local History records attributed edits.
We asked questions about how code behaves across 11 repositories. An answer counted as fully resolved only when it met every graded requirement.
The test checks whether an answer supplies the details you would otherwise have to trace yourself. A team pilot measures whether that reduces investigation, review and correction work.
99/124
30 more fully resolved answers. A 43.5% relative improvement over the strongest comparison in our evaluation.
Company-run evaluation. Code understanding, rather than shipped changes or measured developer-time savings.
See the evaluation and methodsCoral Code used the Codex harness with DeepSeek V4.1 Flash for the outer agent and graph, and Ling 3.0 Flash for initial file agents. The comparison changes both models and coordination.
Claude Opus 4.5 graded the answers at 99/124 for Coral Code and 69/124 for Claude Opus 5.5 (xHigh). GPT-5.5 rescored the same answers at 96/124 and 63/124. This second check grades the same work; it is not another run.
In this separate test, Coral Code High met all 67 grading checks. Codex with Graphify met 56, Codex with Augment MCP met 46, and the agent without added context tools met 45. Each check covered a required part of an answer. These are not counts of completed software changes.
| Task (rubrics) | Codex + Graphify | Codex + Augment MCP | Agent without added context tools | Coral Code High |
|---|---|---|---|---|
| baa0b · trufflehog (26) | 22/26 | 15/26 | 20/26 | 26/26 |
| ba9a9 · simple-login (10) | 9/10 | 8/10 | 5/10 | 10/10 |
| ba9ac · simple-login (10) | 9/10 | 5/10 | 6/10 | 10/10 |
| ba9c5 · scapy (12) | 11/12 | 11/12 | 8/12 | 12/12 |
| ba999 · wp-calypso (9) | 5/9 | 7/9 | 6/9 | 9/9 |
| Total (67) | 56/67 | 46/67 | 45/67 | 67/67 |
A context graph represents relationships between code elements. Coral Code also assigns focused investigations to agents responsible for that code. They exchange findings and return source references and unresolved questions to your coding agent.
Connect the Codex or Claude account you use in the terminal. Choose models for the coding task and focused investigation, and review their usage as you work.
A cheaper answer can cost you more if you have to investigate it again. Coral Code uses smaller models for focused file work. In a pilot, compare total AI spend with the time you spend investigating, reviewing and correcting answers.
Your frontier model plans, answers or edits with prepared project evidence.
Each investigates its assigned file and returns findings. Review supporting-agent usage alongside the coding agent’s usage.
Agents would propose and test changes against a goal and checks you define.
Set a coding agent’s responsibilities, tools and limits within the coordinated investigation.
Coordinate several developers and their agents on a shared working tree with explicit ownership.
Run a four-to-six-week pilot with one JetBrains team. Compare your current workflow with Coral Code on agreed work. Track investigation and review time, corrections, accepted changes and total AI spend. You own the data.
Send product feedback through the issue tracker.