Go beyond the initial patch
File owners trace callers, configuration, generated paths and runtime branches that sit outside the coding agent's active context.
Finish the whole change with your coding agent.
You set the task in Codex, Claude or JetBrains Junie. From that same first turn, coral-code coordinates focused agents across the codebase so affected files, runtime paths, tests and cleanup surface before you ship. You remain in charge.
- Your coding agent
- Leads implementation
- Agent Graph
- Up to hundreds, each with a bounded job
- Your role
- Decides what ships
Outer agent support
JetBrains early preview
One lead agent. Many focused owners.
Keep Codex, JetBrains Junie or Claude as the agent you direct. coral-code supplies JetBrains project structure, recorded runtime paths and file-level responsibilities while the harness keeps its own model and permission controls.
You set
Task + certainty level
Coral assigns
Files + runtime paths + reviews
Local correctness is not codebase completeness.
Carry the change through tests and cleanup
Focused reviews look for partial migrations, weak failure tests, stale adapters and obsolete code after the main implementation lands.
Keep unproven areas visible
Unvisited files, incomplete reports and missing runtime coverage stay explicit instead of becoming implied confidence.
Complete engineering has to cover more than the issue.
SWE Atlas is a benchmark for codebase-scale engineering rather than issue resolution alone. It tests whether an agent can understand an existing codebase, write tests that expose failures and carry a refactor across every affected path.
Research and Coral benchmark evidence
SWE Atlas evaluates codebase understanding, test completeness and refactoring quality. Coral's full-run and controlled studies test whether broader coordination helps coding agents recover more of that work.
Historical comparison snapshot
The same DeepSeek model and Mini-SWE-Agent harness, with and without Coral, across all 124 tasks on 29 July 2026.
Configuration
Harness
Resolve rate
Comparison
coral-code + DeepSeek V4 ProFull 124-task run
Mini-SWE-Agent + Coral
62%
#1
Opus 4.8Public result
Claude Code · xHigh
57.26%
Surpassed
GLM 5.2Public result
Mini-SWE-Agent
48.12%
Surpassed
GPT-5.6-SolPublic result
Codex · xHigh
46.00%
Surpassed
GPT 5.5Public result
Codex · xHigh
45.43%
Surpassed
Muse Spark 1.1Public result
Mini-SWE-Agent · xHigh
42.20%
Surpassed
GPT 5.4Public result
Codex · xHigh
40.80%
Surpassed
Opus 4.7Public result
Claude Code
40.32%
Surpassed
Fable-5Public result
Claude Code · xHigh
39.00%
Surpassed
GPT 5.4Public result
Mini-SWE-Agent · xHigh
36.30%
Surpassed
Opus 4.6Public result
Claude Code
33.30%
Surpassed
GPT 5.3Public result
Codex · xHigh
32.60%
Surpassed
Sonnet 4.6Public result
Claude Code
31.20%
Surpassed
Opus 4.6Public result
Mini-SWE-Agent
30.00%
Surpassed
DeepSeek V4 ProSame-model public baseline
Mini-SWE-Agent
27.15%
#11 baseline
On 29 July 2026, DeepSeek V4 Pro with Mini-SWE-Agent scored 27.15% and ranked #11. The same model and harness with coral-code reported 62% across all 124 tasks. The historical table lists the public configurations between those results on that date. Scale has since added entries, including a later Opus 5 result that Coral does not claim to surpass. Separate controlled studies measured 32.3% to 62.1% with Opus and 29.0% to 50.8% with DeepSeek V4 Pro using the full Coral coordination stack. Scale AI methodology and live leaderboard .
Watch responsibility move through the graph.
Graph and Timeline show who owns each question, which dependencies move between agents, and where work is blocked. Reported model calls are visible; unavailable usage data stays marked as unknown.
Follow the task as it develops.
Operation graph
Source editor
Four decisions stay with you.
Set the standard, anchor the search, let focused owners work, then resolve what remains before the task ends.
Set the bar
Choose your coding agent, its model and permissions. Use Everyday for routine work or High Certainty when the cost of missing part of the change is higher.
Anchor the investigation
IntelliJ PSI exposes symbols, usages and project boundaries. Recorded coverage and DataFlows connect the task to the paths the application runs.
Let file owners converge
Bounded file and question agents work in parallel, exchange discoveries and bring in relevant owners as new dependencies surface.
Resolve the gaps
Open unfinished reports, inspect changed files, request another pass or end the task when the remaining risk is understood.
Install coral-code for your next codebase change.
Choose a compatible JetBrains IDE during installation. Current builds need a local JetBrains backend and support Codex, Claude and JetBrains Junie in Agent Chat.
- Run the command for your operating system.
- Read the early-preview disclaimer, then type
okto continue. - Restart your IDE if JetBrains reports that the plugin cannot reload.
macOS / Linux
$ curl -fsSL https://downloads.coralcode.dev/install.sh | sh
Windows / PowerShell
PS> irm https://downloads.coralcode.dev/install.ps1 | iex
Early preview
This preview includes bugs and incomplete features. The default configuration uses DeepSeek models.
Start free. Add agent capacity when the work needs it.
Sign in to Coral Cloud for a free inner-agent quota. Your outer agent uses its existing subscription.
Add Coral Cloud quota as you run more inner-agent work.
Team subscriptions and annual agreements include managed usage, support and private deployment options.
coral-code@coralos.aiRun the task without leaving JetBrains.
Agent Chat, the project graph, task controls and usage records share one workspace. Move between implementation, delegation and review without reconstructing context in another tool.
Install the early preview
See who owns each part
Graph + Timeline
Trace dependency ownership, inter-agent messages and reported model calls over time.
Direct your coding agent
Tasks + controls
Resume or pause a task, change its model, and keep permissions and Agent Graph mode with the active harness.
Review completion evidence
Changes + usage
Compare file activity with owner reports, token snapshots and known usage gaps.
Review the unfinished work, not just the diff.
In this recorded task, six of eight owner reports are complete while two continue. Open either report, inspect an affected file yourself, or wait before making the final call.
FAQ.
Models, quota, safety and installation.
Updated .
FAQ.
Models, quota, safety and installation.
Updated .
What is coral-code?
coral-code?coral-code is a JetBrains plugin that puts Agent Chat, Agent Graph, task controls and usage records in one workspace. It works with Codex, JetBrains Junie and Claude.
How does coral-code help me complete a change?
coral-code help me complete a change?Agent Graph divides the task among bounded file, question and DataFlow owners from the start. They inspect dependencies in parallel and share what they find with the coding agent.
Which coding agents and IDEs does coral-code support?
coral-code support?Current builds require a compatible JetBrains IDE and local backend. Agent Chat supports Codex, JetBrains Junie and Claude. Codex uses its account, Junie routes models through Coral Cloud, and Claude uses its own credential.
What is Agent Graph, and when should I use it?
Agent Graph uses IntelliJ PSI, file relationships, DataFlows and recorded operations to coordinate focused agents.
Use Off for a small edit, Everyday for routine codebase work, and High Certainty when you need broader investigation, explicit gap checks and stronger completion evidence.
What evidence can I inspect before accepting a result?
Open reports, Graph and Timeline activity, changed files, attempts and observed model usage in the JetBrains workspace.
Incomplete reports and unavailable usage data remain labelled.
How do accounts, models and quota work?
Codex uses its existing account and quota. Junie routes model requests through Coral Cloud. Claude uses its own account and credential. Agent Graph agents use the model selected in coral-code settings.
The usage view labels each observed source and marks unavailable account data.
How does coral-code handle safety and production use?
coral-code handle safety and production use?Agent Chat uses the access and approval controls of the selected coding agent. Agent Graph agents use project-scoped PSI tools. The selected coding agent keeps command execution and web access behind its approval controls.
coral-code is an early preview, has not been audited as a sandbox and does not replace code review, testing or human approval.
How do I install coral-code?
coral-code?Run the installer for your operating system and accept the early-preview disclaimer. Restart the IDE if JetBrains cannot reload the plugin.
Open the install commands
Infrastructure used by coral-code
Use the same coordination layer in your own agent systems.
CoralOS provides the Server, Cloud, model proxy and Console behind coral-code. Connect your own agent frameworks, route models, apply budgets and keep operational history under one authority.