scan_project
Find your starting point.
Scan the project for risk signals. See which files have the most going on, before spending a model call.
A SECOND OPINION FOR YOUR CODEOPEN SOURCE
Catch the bug.
Before it
ships.
Find likely runtime failures in JavaScript and TypeScript. Give your coding agent the context to see what it missed.
Seven MCP tools. The model CLI you already use.
reserveStock.tsSOURCE / 01
18
const onHand = await getStock(sku);19
20
if (onHand < qty) {21
return {ok: false, remaining: onHand};22
}23
24
await recordReservation(orderId, sku, qty);25
26
await setStock(sku, onHand - qty);27
28
return {ok: true, remaining: onHand - qty}; One await. Two callers. A race.
Likely failure foundL26
Stock can oversell.
Concurrent calls read the same quantity.
The last write wins.
race_conditionscore 0.82 A REAL FINDING FROM THE DEMO
LESS GUESSWORK. MORE CONTEXT.Scroll to explore
Fits the way
you already build.
Claude Code
Codex
GitHub Copilot
YOUR CLI.YOUR MODEL ACCESS.
THE WORKFLOW
A little context.
A much better review.
From the first signal to the second opinion.
Follow one real bug through the tool.
map_dependencies
Put the code in context.
Follow the imports. See who calls a file and which tests connect to it, so a local change gets a wider review.
predict_failures
Ask for the failure.
An independent model checks the code. Here, two concurrent reservations can read the same stock and overwrite each other.
predict_failures
Fix it. Check it again.
Move the stock check and decrement before the await. Re-check the file, then test the behavior across the feature.
01 / 04 SCROLL TO CONTINUE
Skip to the results predictive-debuggerRECORDED SESSION
› scan_project demo/src
FILERISK DENSITY
auditLog.ts3 async boundaries · 1 branch0.58reserveStock.ts3 async boundaries · 1 branch0.51stockStore.ts1 async boundary · 1 branch0.35formatMoney.tscyclomatic complexity 10.05priceCart.tscyclomatic complexity 10.025 source files scanned · no model call
Risk density helps prioritize reading. It is not a probability that a file contains a bug.
› map_dependencies stockStore.ts
reserveStock.tsimports at line 1imports↓
stockStore.tsselected file5 files indexed · 0 unresolved imports
› predict_failures reserveStock.ts
18
const onHand = await getStock(sku);19
20
if (onHand < qty) {21
return {ok: false, remaining: onHand};22
}23
24
await recordReservation(orderId, sku, qty);25
26
await setStock(sku, onHand - qty);race_condition
0.82Concurrent reservations can overwrite the same stock value.
Read the recorded verdict
onHand is read at line 18 before two awaits; concurrent reserveStock calls for the same sku both read the same onHand and setStock(sku, onHand-qty) overwrites each other, letting stock oversell despite the negative-stock guard.
Confidence score from this recorded prediction, not a measured accuracy rate.
› predict_failures fixed/reserveStock.ts
− const onHand = await getStock(sku);
− await setStock(sku, onHand - qty);
+ const remaining = tryReserve(sku, qty);
// Check and decrement before the await.
Key lines from the fix. The full demo also handles insufficient stock and adds tryReserve to stockStore.ts.
No defect demonstrated.
Recorded re-check · pattern: none
tryReserve and stock update are synchronous and atomic before the await, so no race window exists between check and reservation; audit log failure would just propagate as a rejected promise.
A clean file verdict does not prove the whole feature is correct.
Actual v0.8.1 output · Claude provider · September 2026
Read the tool reference
MEASURED / V0.9.0
More signal.
Fewer tokens.
Same 49 files. Three coding agents.
Compare an agent reading the code with an agent using Predictive Debugger.
24 bugs · 25 clean files · 2 sessions per setup
Claude Code
38%
fewer tokens than direct reading
CLI 2.1.283 · Sonnet
Agent reads files279k
With v0.8.2771k
With v0.9.0173k
| Setup | Bugs / 24 | False alarms |
|---|---|---|
| Agent reads files | 21, 21 | 0 |
| With v0.8.2 | 21, 21 | 0 |
| With v0.9.0 | 21, 22 | 0 |
One more bug found across two sessions than direct reading. 78% fewer tokens than v0.8.2.
GitHub Copilot
32%
fewer tokens than direct reading
CLI 1.0.82 · Sonnet
Agent reads files883k
With v0.8.21.50M
With v0.9.0598k
| Setup | Bugs / 24 | False alarms |
|---|---|---|
| Agent reads files | 20, 20 | 1 |
| With v0.8.2 | 22, 24 | 0 |
| With v0.9.0 | 22, 22 | 0 |
Two more bugs per session than direct reading. v0.8.2 found two more across both sessions, at 2.5× the tokens.
Codex
69%
fewer tokens than direct reading
CLI 0.157.1 · GPT-5.6 sol
Agent reads files1.01M
With v0.8.2Unavailable
With v0.9.0308k
| Setup | Bugs / 24 | False alarms |
|---|---|---|
| Agent reads files | 22, 21 | 0 |
| With v0.8.2 | — | — |
| With v0.9.0 | 23, 23 | 2, 0 |
23 of 24 bugs found in each session. Two false alarms in the first, none in the second. v0.8.2 could not run with approvals off.
Tokens include the calling agent and every model call made by the tool, averaged over two sessions. All charts use the same scale. Bug counts show each session; false alarms are totals unless shown separately. The answer key informed development except for 12 held-out files. These results describe this benchmark, not accuracy on every project. CommonJS context and verdict caching shipped after these measurements.
A SMALL, FOCUSED TOOLKIT
Seven tools.
Better questions.
Local analysis narrows the search.
An independent model gives you another view.
01Local
scan_projectKnow where to look.
Rank source files by risk density. Start with the code that deserves a closer read.
02Local
analyze_fileRead the signals.
Inspect complexity, async boundaries, and the signals behind a file’s risk score.
03Local
map_dependenciesFollow the connections.
Trace imports, reverse imports, and connected tests back to their source lines.
04Model call
predict_failuresGet a second opinion.
Ask an independent model for likely runtime failures, with a line number and a reason.
05Local
analyze_logsMake sense of the logs.
Surface unusual log entries, ranked by severity and unusual wording. Requires Python 3.
06Local
list_providersCheck your connection.
See which supported CLIs are installed and check their sign-in status.
07Local
check_typesCheck the types.
Get compiler diagnostics for selected JavaScript and TypeScript files. No model call needed.
YOUR NEXT REVIEW
Give your agent
a second set
of eyes.
Add the MCP server. Restart your agent.
Ask it where your code might fail.
01 / CONNECT THE TOOL
claude mcp add --scope project predictive-debugger -- npx -y predictive-debugger@latestAdds the server to this project. Use --scope user for every project.
codex mcp add predictive-debugger -- npx -y predictive-debugger@latestAdds the server for your user. See the setup guide for project-only configuration.
copilot mcp add predictive-debugger -- npx -y predictive-debugger@latestAdds the server for your user. See the setup guide for project-only configuration.
02 / START A CONVERSATION
“Use Predictive Debugger to find the riskiest files in src/.”
Node.js 22+ · A supported CLI, installed and signed in.
Prediction calls use your CLI’s model access and allowance.
A FEW DETAILS
Before you start.
Does my code leave my machine?
Static analysis, dependency maps, and log analysis run locally. Predictions send source and bounded dependency context to your CLI’s model provider. Credentials stay with your CLI. See the security model.
Do I need a separate API key?
No. Predictions use the Claude Code, Codex, or GitHub Copilot CLI you are already signed in to, including its usage allowance.
What code can it review?
JavaScript and TypeScript, including JSX, TSX, and decorators. Vue and Svelte single-file components are not supported. Files over 4 MB are rejected, and large predictions may cover selected declarations. See the tool reference.
Does it replace tests or code review?
No. Risk scores and predictions can be wrong. A clean verdict on one file does not prove a feature works. Use the findings to guide your review, then verify behavior with tests and the surrounding code.
Is there a VS Code extension?
There is an experimental extension that can show findings in the Problems panel. It currently requires a local build and is not on the Marketplace. Try the VS Code preview.