Tool reference

The MCP server exposes seven tools. Type checking, static analysis, dependency mapping and log analysis run locally. Only predict_failures calls a model provider.

Source analysis supports .js, .jsx, .mjs, .cjs, .ts, .tsx, .mts and .cts, including decorators. Vue and Svelte single-file components are excluded. Files above 4 MB are rejected. Use absolute paths unless a parameter says otherwise.

scan_project

Rank a directory’s source files by riskDensity, the concentration of structural risk signals per line. This helps choose where to read first; it does not report confirmed defects.

Parameter	Default	Purpose
directory	Required	Directory to scan
limit	50	Maximum results, up to 500
includeTests	false	Include test files in the ranking
verbose	false	Include raw metric counts

Results include scanned, returned, orderedBy and ranked files. Paths are relative to the scanned directory. Unreadable files are reported separately so one failure does not discard the rest of the scan.

Tests are excluded by default because mocked awaits and other test structures can rank highly without indicating production risk. Set includeTests: true when reviewing the test suite itself.

analyze_file

Pass file to get AST complexity metrics, riskScore, riskDensity and the signals behind the scores. A parse failure returns parseError and a zero score; that zero is not evidence that the file is safe.

check_types

Pass files with up to 20 absolute paths from one project. Optional project selects a tsconfig.json or jsconfig.json; otherwise the nearest config is discovered. The bundled TypeScript compiler loads the complete config’s root files, preserving ambient declarations, aliases and inherited options. Only selected files’ diagnostics are returned. Nothing is emitted or executed, and no provider is called. The reply still occupies the calling agent’s context.

Without a config, mode: "inferred" uses ES2022, bundler module resolution, JS checking, null checking and implicit-this checking on the selected roots. Configured JavaScript checking and file-level opt-outs are respected. This is compiler evidence under those settings, not proof of a runtime defect or safety.

Results include compiler, project, checked, skipped, diagnostics, contextIssues and truncated. Each diagnostic has a code, message and source position. Missing dependencies and invalid compiler options appear separately as context issues. status: "unavailable" reports a config failure or limit; it never means the files are clean.

Limits are 1,000 root files, 1,000 compiler reads, 4 MB per file, 32 MB total reads and 100 returned diagnostics. Project references are currently unsupported; select a leaf config. The compiler runs synchronously, so these size limits are not a hard execution timeout. The package includes the compiler and its standard declarations; a project-local TypeScript installation is not required.

map_dependencies

Find a file’s imports, reverse imports and tests connected through imports. Each relationship has a via chain with paths, import kinds and source lines. These are static file relationships; they do not establish runtime callers or test execution coverage.

Parameter	Default	Purpose
directory	Required	Project directory; outside source is excluded
file	Required	Source file, absolute or relative to directory
depth	1	Import hops in each direction, from 1 to 3
limit	50	Total neighbors across both lists, up to 200
maxFiles	1000	Maximum discovered source files, up to 2,000

The map includes tests and follows ESM imports and re-exports, type import expressions and literal dynamic imports. test: true follows test-path naming conventions. CommonJS requires, import assignments and nonliteral dynamic imports remain unresolved.

Build, vendor and hidden directories are excluded. The index refreshes on every request. Work is bounded by 20,000 directory entries, 64 directory levels, 4 MB per source file, 32 MB of source reads and 10,000 import references. Serialized replies are capped at 32,000 characters.

Field	Meaning
unresolved	Unresolved outgoing imports from the requested file
coverage.unresolved	Unresolved references across the scanned project
issues	Read and parse failures, plus scan-limit details
coverage.scanLimited	The scan reached a discovery or processing limit
truncated	Some reply entries were omitted

A missing relationship in a partial scan is not proof of absence. The tool uses no model calls, but its replies occupy the calling agent’s context. See the local measurements.

analyze_logs

Pass logFile to score log lines by severity and unusual wording, returning the anomalies first. Optional threshold ranges from 0 to 1 and defaults to 0.5. Requires Python 3; see setup.

predict_failures

Combine static analysis with an independent verdict from a supported CLI’s model. Each file requires a model call and uses the provider’s usage allowance.

Parameter	Default	Purpose
file	Required unless files is supplied	One source file
files	None	Non-empty batch of source paths; takes precedence over file
concurrency	4	Concurrent model calls, from 1 to 8
provider	First installed CLI	claude, codex or copilot
model	CLI default	Model override passed to the provider
calleeContext	true	Include bounded imported definitions and referenced types
multi	false	Request all demonstrable findings; experimental
logFile	None	Log file to include in the combined score
verbose	false	Include static metrics and the full log breakdown

Use files for a change set. Up to eight small files share a model call, with the grouped prompt bounded at 120,000 characters. Large files run alone with their existing source allowance. Groups run concurrently; lower concurrency if the provider starts rate-limiting. Duplicate resolved paths are removed.

The reply contains results in input order and lists failures separately. Each verdict is checked against its own file’s source ranges. Missing, duplicate or incomplete verdicts are unavailable, never clean. A file whose group verdict names a defect with a score under 0.80 is reviewed again on its own, and that verdict replaces the group’s: grouped scores near the 0.70 gate were unreliable in both directions. So is a file the group calls clean when it contains an async function that awaits a read and later awaits a write, the shape of a lost update that group replies missed. A failed CLI call affects its group and is not retried; retry only files whose assessment failed or is unavailable.

Within one server session, a file whose review input is unchanged (its source, imported definitions, provider and model) returns its earlier verdict with cached: true instead of calling the model again, for up to an hour. Failed and unavailable verdicts are never reused.

Only JavaScript/TypeScript source files are sent to the model; other paths fail before any provider call. Each review call starts the CLI in a new, empty temporary folder that is deleted afterwards, so the CLI does not load your project’s CLAUDE.md or AGENTS.md into the review. SECURITY.md explains why.

Reading a verdict

The reply includes pattern, score, line, reason, status and actionable. Use actionable when deciding whether to present a prediction as a defect.

status	Meaning
actionable	A reported defect has a model score of at least 0.7
uncertain	A named defect scored below the reporting threshold. The reply adds check: read the cited lines and confirm or dismiss it
none	No defect reported
unavailable	No usable model verdict; this is not a clean result

checked lists the bug categories the model says it considered. It is a self-report, and an empty list means no coverage was reported. Additional findings appear in findings when returned. The reporting threshold was measured on single-finding replies, so multi: true is less well characterized.

Large files send at most 120,000 source characters, selected as whole declarations with original line numbers. The reply discloses truncation; omitted code remains unreviewed. There is no automatic second model pass.

Dependency context

By default, the model also receives bounded supporting definitions from local dependencies. Set calleeContext: false to send only the source file. Third-party packages are never read for this context.

Supported cases include direct ESM imports, declared imported objects, referenced types, imported constructors, named/default binding re-exports and unambiguous export * barrels through at most four files. Local tsconfig.json path mappings support relative config inheritance.

CommonJS exports, namespace re-exports, package-based config inheritance and injected instance methods remain unresolved. Conflicting or unreadable wildcard branches remain unknown. Each collection parses at most 24 dependency files, each no larger than 4 MB. Each requested export has a 128-step traversal limit, including cached paths.

The text budget follows source length, with a 1,000-character minimum and 16,000-character maximum. It includes at most 12 definitions of 3,000 characters each. Oversized imported objects and static class members prioritize the called member, referenced fields and helpers. Retained members stay in source order, and omissions are marked. Dynamic definitions, duplicate overrides, inheritance and decorators retain prefix truncation. This does not establish complete state flow through the program.

Reviewing code an agent just wrote

The server advertises the review routing in MCP instructions and repeats a short reminder in each prediction’s review field. Some clients do not forward server instructions, so the result also carries it.

For changes spanning files, a sub-agent should receive the changed files and intended behavior, review only that scope, and run in the background where the host supports it. The sub-agent rule’s benefit has not been measured. A clean per-file prediction cannot establish that the feature meets its requirements.

list_providers

Takes no parameters. Reports installed supported CLIs and their sign-in checks. Some credential-store checks are provisional; a live CLI call confirms access.

How the score works

riskScore measures accumulated structural signals such as nested loops, long functions and async boundaries. riskDensity adjusts for file length and reduces the influence of signals that accumulate with it. scan_project sorts by density. Both scores use smooth saturation to stay within 0 to 1 without flattening every complex file to the same maximum.

combinedScore is a separate blend used by the prediction pipeline:

Input	Weight
Model verdict	0.90 without logs
Static risk	0.10 without logs
Log anomalies, when available	0.15; the other weights are scaled to total 0.85

The model verdict dominates. The actionable gate uses the model’s score, not combinedScore. The implementation is in score.ts and confidence.ts. See benchmark results for measurements and their limits.