by Anatoly Koptev

Krolik review, pricing and verdict

Self-hosted MCP server that gives coding agents a call graph and review memory

Best for
Senior backend engineer using Claude Code or Cursor on large monorepos
Not for
Junior developer who wants zero-config setup
  • Developer
  • Web search

Should you use Krolik?

Krolik is the product name for Vaelor, an Go engine with 1,422 commits that parses code with tree-sitter and stores it in Apache AGE. You run it yourself with Docker and point your agent at a local MCP URL. Expect to operate Postgres and an embedder if you want the full tool set.

debug_investigate: a 7-phase pipeline that bridges Prometheus alerts and Jaeger traces to the root-cause file:function in 5 seconds, a capability no other free MCP server provides.

Open source and easy to fork; the moat is the depth of the debug pipeline and taint analysis, which take real engineering effort to replicate. But competitors with larger communities can catch up.

Also for

  • Security engineer doing agent-assisted vulnerability triage
  • Backend engineers who want AI coding agents to navigate large monorepos without reading every file
Show 3 more
  • Security engineers using AI agents to automate taint tracking and injection vulnerability detection
  • Platform teams that need agents to debug production incidents from Prometheus alerts without manual log triage
  • Staff engineers doing large-scale refactors who need agents to understand the full blast radius before changing shared utilities

Also not for

  • Team evaluating MCP tools that needs community support and documentation

Main trade-off

2 GitHub stars as of June 2026: effectively zero community support, no external documentation, and no indication of production adoption.

Krolik pricing

Free and open source under Apache 2.0, self-hosted only. Vaelor's site says there is no hosted SaaS and no signup, so there is no plan to buy. Your running costs are your own server and, for the LLM-backed tools, an OpenAI-compatible endpoint of your choice. The agent client you attach it to is billed separately.

Install Krolik

claude mcp add -s user -t http vaelor http://127.0.0.1:8897/mcp

Before you start

  • Build and start the server: git clone https://github.com/anatolykoptev/vaelor, docker build -t vaelor ., then docker run -p 8897:8897 with LLM_API_BASE and LLM_API_KEY set
  • Register the server URL in your MCP client using the install command above
  • Optionally attach PostgreSQL with Apache AGE (DATABASE_URL) for the knowledge graph tools
  • Optionally add pgvector and an embedder (EMBED_URL) for semantic_search

How to invoke it

Exposed as MCP tools over Streamable HTTP on port 8897. Register the URL with your client (command under install). The agent calls tools by name (for example call_trace, understand, impact_analysis) with JSON arguments.

Works with

  • Claude Code
  • Cursor
  • Windsurf
  • Aider
  • Cline
  • Any MCP client

Required runtime

  • Docker (recommended) OR Go 1.24+ with a C compiler
  • OpenAI-compatible LLM endpoint for full capability (code_graph and repo_search require it; most other tools run without it)
  • PostgreSQL with Apache AGE (optional, for code_graph)
  • pgvector plus an embedder such as ox-embed-server (optional, for semantic_search and code_research reranking)

Where it runs

  • Desktop
  • CLI
  • API

Data it can reach

  • source code repositories
  • call graphs
  • symbol tables
  • Prometheus metrics
  • Jaeger traces
  • GitHub repositories
  • web pages

The permissions this connector asks for, as its own documentation states them.

What Krolik can do

26 actions an agent can call. Open one for its parameters and an example.

debug_investigateRuns a seven-phase incident pipeline: metric spikes, Jaeger failed traces, symbol resolution, callgraph walks, log excerpts, deployed-binary drift and LLM fusion, ranked to file:function.

Example call

{"arguments":{"depth":4,"query":"rate(http_requests_total{status=\"500\"}[5m]) > 0.1"},"tool":"debug_investigate"}

Parameters

ParameterTypeNeededDescription
querystringRequiredThe error message, Prometheus alert expression, or Jaeger trace to investigate.
depthnumberdefault 3OptionalHow many call levels to trace when following the root-cause hypothesis.

Returns

FieldTypeDescription
root_causestringFile path and function name identified as the root cause.
hypothesisstringExplanation of the failure chain.
semantic_searchFinds code by concept using hybrid BM25F lexical scoring, pgvector 768-dim embeddings, one-hop AGE graph expansion, and ColBERT cross-encoder reranking.

Example call

{"arguments":{"limit":10,"query":"user authentication token validation"},"tool":"semantic_search"}

Parameters

ParameterTypeNeededDescription
querystringRequiredNatural language description of the code to find.
limitnumberdefault 10OptionalMaximum number of results to return.

Returns

FieldTypeDescription
resultsarrayRanked list of file paths, symbol names, and relevance scores.
symbol_searchSearches functions, types, and constants by name pattern across all indexed files and languages.

Example call

{"tool": "symbol_search", "arguments": {"pattern": "HandleAuth*", "language": "go"}}

Parameters

ParameterTypeNeededDescription
patternstringRequiredName pattern to match, supports wildcards.
languagestringOptionalFilter results to one language (e.g. 'go', 'python', 'typescript').

Returns

FieldTypeDescription
symbolsarrayList of matching symbol names with file path and line number.
call_traceTraces call chains bidirectionally from a given symbol with configurable depth, showing all callers or all callees.

Example call

{"tool": "call_trace", "arguments": {"symbol": "CreateOrder", "direction": "up", "depth": 5}}

Parameters

ParameterTypeNeededDescription
symbolstringRequiredFully qualified or partial function or method name to trace.
directionstringdefault upOptionalTrace direction: 'up' to find callers or 'down' to find callees.
depthnumberdefault 5OptionalMaximum number of call levels to follow.

Returns

FieldTypeDescription
call_chainobjectTree structure of the call chain with file paths and line numbers.
impact_analysisCalculates the blast radius of changing a symbol: all direct and transitive callers up to depth 10 across the entire codebase.

Parameters

ParameterTypeNeededDescription
symbolstringRequiredSymbol name to analyze for change impact.
depthnumberdefault 5OptionalMaximum traversal depth for impact calculation, up to 10.

Returns

FieldTypeDescription
affected_symbolsarrayList of symbols affected by a change, with their file paths and distance from the target.
dataflowTracks values through the IL and CFG to find taint paths, dead stores and SQL or command injection sinks.

Parameters

ParameterTypeNeededDescription
symbolstringRequiredStarting function or symbol to trace dataflow from.
sourcestringOptionalTaint source to track (e.g. an HTTP input function).

Returns

FieldTypeDescription
taint_pathsarrayList of taint flow paths from source to potential sinks, with file and line numbers.
rewriteApplies structural AST search-replace patterns with wildcard bindings across 16 languages, with a dry-run preview before apply.

Parameters

ParameterTypeNeededDescription
patternstringRequiredTree-sitter AST pattern to find, supports $x wildcard bindings.
replacementstringRequiredReplacement pattern using the same $x bindings.
languagestringOptionalLanguage to run the rewrite in. Omit to apply across all indexed files.

Returns

FieldTypeDescription
changed_filesarrayList of files modified with change counts per file.
review_prComputes the differential blast radius of a pull request and stores one learning per changed symbol from the real review outcome.

Example call

{"tool": "review_pr", "arguments": {"pr_url": "https://github.com/org/repo/pull/42"}}

Parameters

ParameterTypeNeededDescription
pr_urlstringOptionalGitHub pull request URL to analyze.
diffstringOptionalRaw unified diff text if no PR URL is available.

Returns

FieldTypeDescription
reviewobjectBlast radius, risk score, and list of affected symbols with file paths.
Show all 26 actions
understandReturns type-aware symbol analysis with complexity metrics, including type signatures, references, and cyclomatic complexity.

Parameters

ParameterTypeNeededDescription
symbolstringRequiredSymbol to analyse, fully or partly qualified.

Returns

FieldTypeDescription
analysisobjectType signature, call count, complexity score, and all references in the codebase.
code_graphQueries the Apache AGE property graph using a natural language question translated to Cypher, returning graph relationships between symbols.

Parameters

ParameterTypeNeededDescription
querystringRequiredNatural language question about code relationships.
depthnumberdefault 3OptionalMaximum graph traversal depth for the generated Cypher query.

Returns

FieldTypeDescription
graph_resultsarrayMatching nodes and edges from the knowledge graph with file and line metadata.
dead_codeReports likely-unused symbols with a confidence score rather than a flat list; stricter verdicts need the ox-codes companion.

Example call

{"tool": "dead_code", "arguments": {"language": "typescript", "threshold": 0.8}}

Parameters

ParameterTypeNeededDescription
languagestringOptionalRestrict analysis to one language. Omit to scan all indexed languages.
thresholdnumberdefault 0.7OptionalConfidence threshold (0.0 to 1.0) above which a symbol is reported as dead.

Returns

FieldTypeDescription
dead_symbolsarrayList of likely-unused symbols with confidence scores and file paths.
code_healthGrades a repository from A to F using complexity, test coverage, dependency freshness and OSV vulnerabilities. Fuller grades use the ox-codes companion.

Example call

{"tool": "code_health", "arguments": {}}

Returns

FieldTypeDescription
gradestringLetter grade A through F for overall code health.
breakdownobjectPer-category scores: dead code, complexity, dependencies, test coverage.
repo_analyzePerforms a deep, quick, or issue/PR-focused analysis of a repository and returns a structured overview.

Example call

{"tool": "repo_analyze", "arguments": {"repo_path": "/workspace/my-service", "mode": "deep"}}

Parameters

ParameterTypeNeededDescription
repo_pathstringRequiredAbsolute path or GitHub URL of the repository to analyze.
modestringdefault quickOptionalAnalysis mode: 'deep' for full indexing, 'quick' for overview, 'issues' or 'prs' for issue/PR context.

Returns

FieldTypeDescription
analysisobjectRepository overview including languages, key modules, health grade, and open issues summary.
exploreReturns a fast repository overview without an LLM call, and without cloning for remote repositories.

Example call

{"tool": "explore", "arguments": {"repo_url": "https://github.com/org/repo"}}

Parameters

ParameterTypeNeededDescription
repo_urlstringOptionalGitHub repository URL. Omit to explore the currently indexed workspace.

Returns

FieldTypeDescription
overviewobjectRepository name, primary language, file count, key entry points, and README summary.
file_parseParses a single file with tree-sitter and returns either a flat list of symbols or the full AST in JSON.

Example call

{"tool": "file_parse", "arguments": {"file_path": "src/auth/token.go", "output": "symbols"}}

Parameters

ParameterTypeNeededDescription
file_pathstringRequiredRelative or absolute path to the file to parse.
outputstringdefault symbolsOptionalOutput format: 'symbols' for a flat list or 'ast' for the full AST.

Returns

FieldTypeDescription
resultobjectParsed symbols or AST depending on the output parameter.
dep_graphBuilds a dependency graph for the indexed codebase and outputs it in Mermaid, Graphviz DOT, or JSON format.

Example call

{"tool": "dep_graph", "arguments": {"format": "mermaid"}}

Parameters

ParameterTypeNeededDescription
formatstringdefault jsonOptionalOutput format: 'mermaid', 'graphviz', or 'json'.

Returns

FieldTypeDescription
graphstringDependency graph in the requested format.
code_compareCompares two repositories structurally across architecture, API surface, and code quality dimensions.

Example call

{"tool": "code_compare", "arguments": {"repo1": "https://github.com/org/service-v1", "repo2": "https://github.com/org/service-v2"}}

Parameters

ParameterTypeNeededDescription
repo1stringRequiredPath or URL of the first repository.
repo2stringRequiredThe repository to compare against, as a path or URL.

Returns

FieldTypeDescription
comparisonobjectSide-by-side comparison of architecture patterns, API symbols, and quality scores.
prepare_changeAssesses the risk of a planned code change by identifying affected symbols, tests, and dependent services before any code is written.

Parameters

ParameterTypeNeededDescription
symbolstringRequiredSymbol to be changed.
change_descriptionstringRequiredPlain English description of the intended change.

Returns

FieldTypeDescription
risk_reportobjectRisk score, list of affected tests, downstream services, and recommended validation steps.
code_researchPerforms BM25F plus embedding-based search optimized for large monorepos where semantic_search may be too broad.

Parameters

ParameterTypeNeededDescription
querystringRequiredPlain-English research question about the codebase.
limitnumberdefault 5OptionalMaximum number of results.

Returns

FieldTypeDescription
resultsarrayRanked list of relevant code snippets with file paths and match scores.
design_searchFinds DESIGN.md systems by UI description using multilingual-e5-large embeddings.

Example call

{"tool": "design_search", "arguments": {"description": "The checkout flow that shows a cart summary before payment"}}

Parameters

ParameterTypeNeededDescription
descriptionstringRequiredUI or workflow description to search for in the codebase.

Returns

FieldTypeDescription
matchesarrayMatching components with file paths and confidence scores.
resolve_frameUnminifies a JavaScript or TypeScript stack frame by resolving it through a source map back to the original file and line number.

Example call

{"tool": "resolve_frame", "arguments": {"frame": "at t.e.n.processEvent (bundle.min.js:1:43291)", "source_map_url": "https://cdn.example.com/bundle.min.js.map"}}

Parameters

ParameterTypeNeededDescription
framestringRequiredMinified stack frame text to unminify.
source_map_urlstringOptionalURL to the source map file. If omitted, Krolik attempts to locate it automatically.

Returns

FieldTypeDescription
original_framestringResolved original file path, line number, and function name.
site_analyzeCrawls a website to produce a tech stack fingerprint and SEO audit including meta tags, canonical URLs, and structured data.

Example call

{"tool": "site_analyze", "arguments": {"url": "https://example.com"}}

Parameters

ParameterTypeNeededDescription
urlstringRequiredRoot URL of the site to analyze.

Returns

FieldTypeDescription
tech_stackarrayDetected technologies (framework, CDN, analytics).
seo_issuesarrayList of SEO issues found with severity ratings.
site_crawlRuns a breadth-first crawl of a website and returns all discovered URLs with their status codes.

Parameters

ParameterTypeNeededDescription
urlstringRequiredStarting URL for the BFS crawl.
depthnumberdefault 2OptionalMaximum crawl depth from the starting URL.

Returns

FieldTypeDescription
pagesarrayCrawled URLs with HTTP status codes and link counts.
fleet_versionsDiffs pinned container image versions in configuration files against the versions actually running in the deployed environment.

Returns

FieldTypeDescription
driftarrayList of services where the pinned version differs from the deployed version.
repo_searchDiscovers GitHub repositories matching a query by combining SearXNG web search with GitHub API results, ranked by an LLM.

Parameters

ParameterTypeNeededDescription
querystringRequiredNatural language query to find matching repositories.
limitnumberdefault 5OptionalMaximum number of repositories to return.

Returns

FieldTypeDescription
repositoriesarrayRanked list of repository URLs with star counts and descriptions.
review_deltaComputes differential blast radius for a code diff and persists learnings, functionally identical to review_pr but accepts raw diffs directly.

Example call

{"tool": "review_delta", "arguments": {"diff": "--- a/auth.go\n+++ b/auth.go\n@@ -12,7 +12,8 @@"}}

Parameters

ParameterTypeNeededDescription
diffstringRequiredUnified diff text to analyze.

Returns

FieldTypeDescription
reviewobjectBlast radius, risk score, and affected symbols with file paths.

What Krolik is for

Krolik is a free, self-hosted MCP server that gives AI coding agents a persistent call graph across 16 languages. Built on the open-source Vaelor engine, it shows what breaks when a function changes, remembers past review verdicts per symbol, and traces production alerts to a file and function. It registers in Claude Code, Cursor, Windsurf, Aider and Cline.

  • Persistent call graph

    Symbols, calls, imports and test relationships are stored in an Apache AGE property graph on PostgreSQL, so code_graph can answer questions grep cannot.

  • Tree-sitter parsing with typed call edges

    Typed edges come from go/types for Go and bundled SCIP indexers for four further languages; everything else gets plain tree-sitter edges.

  • Blast radius before the edit

    impact_analysis defaults to depth 5, pages direct callers, and reorders hotspots by churn; prepare_change adds a dead-code check in the same call.

  • Review memory per symbol

    review_pr stores approve as good and request-changes as bad for each changed symbol, and understand returns the nearest learnings on later calls.

Show all 8 features
  • Seven-phase incident pipeline

    debug_investigate joins metric spikes, Jaeger failed traces, callgraph walks, log excerpts and deployed-binary drift into ranked file:function hypotheses.

  • Hybrid semantic search

    semantic_search blends BM25F, pgvector embeddings and one-hop graph expansion; it needs pgvector plus an embedder.

  • Structural rewrite and taint tracking

    rewrite applies wildcard AST patterns across every supported language with a dry run, and dataflow traces tainted input to SQL and command sinks.

  • One endpoint for every MCP client

    A single Streamable HTTP URL serves Claude Code, Cursor, Windsurf, Aider, Cline and any other MCP client.

What people use it for

  • Production incident debugging

    An agent receives a Prometheus alert, calls debug_investigate, and gets ranked file:function hypotheses from metrics, failed traces and callgraph walks.

  • Safe refactoring

    Before changing a function, agent calls impact_analysis with depth 10 to see every direct and transitive caller across the entire codebase.

  • Security taint tracking

    Agent calls dataflow_analyze to trace unescaped user inputs through the IL/CFG and identify potential injection points before they reach production.

  • Codebase-wide AST rewriting

    Agent calls rewrite with a tree-sitter pattern to replace a deprecated API across all 16 supported languages in one pass.

  • Automated PR review

    Before approving a pull request, the agent calls review_pr to see which symbols and tests it touches, and the verdict is stored for next time.

Strengths and limits

Strengths

  • Agents get blast radius, callers and prior review verdicts in one call, so a risky edit is flagged before CI instead of after.
  • Open source and self-hosted: the code and the graph stay on your own infrastructure, and there is no hosted plan to outgrow.
  • Covers a wide language list, including Kotlin, Swift, Vue and HTML, with type-aware call resolution on Go and, through SCIP, on four more.
  • Works with every MCP client from one URL, so a team mixing editors shares one index.

Limits

  • Only 5 GitHub stars on the Vaelor repository as of 21 October 2026, so community answers and third-party setup notes are scarce.
  • Full capability means running PostgreSQL with Apache AGE, pgvector and an embedder; code_graph and repo_search also need an LLM key.
  • Two names and two sites: krolik.tools lists 28 tools and 13 languages while vaelor.ai lists 35+ and sixteen, so check tools/list on your own instance.
  • No hosted option, so trying it means building a Docker image first.

Krolik alternatives

The nearest AI agent skills on Hokai for the same job.

SkillWorks withConnects byChecked
Krolik This pageClaude · Cursor · WindsurfMCP21 Oct
Godot AILets agents build scenes, edit scripts and run games in a live Godot editorClaude · Codex · Cursor · Gemini +1MCP15 Nov
AisyncAdvisory file claims and a shared journal for your AI coding agents.Claude · Codex · CursorSkill file11 Nov
CloakroomMasks secrets and personal data before your coding agent sees them.Claude · Codex · Cursor · GeminiSkill file11 Nov

Krolik facts

Protocol
MCP
Latency class
Fast
Open source
Yes

What changed

  • Release v1.21.0: Kotlin, Swift, HTML language support added source

Questions about Krolik

How much do you pay for Krolik?

Nothing for the software. The vendor sells no plan or subscription, so there is nothing to buy. You cover your own hosting and, for the LLM-backed tools, whichever OpenAI-compatible endpoint you connect. The agent client you attach it to bills separately.

Can you use Krolik without paying?

Yes, commercial use included. The engine and its two optional Rust companions, ox-embed-server and ox-codes, are open source and nothing is feature-gated. The limits are operational: code_graph needs PostgreSQL with Apache AGE, and semantic_search needs pgvector plus an embedder.

Which tools compete with Krolik in 2026?

Sourcegraph suits people searching code across an organisation, with no graph database to run. Greptile and CodeRabbit are hosted reviewers, the better choice if you only want pull request feedback. Vaelor's own FAQ says Cursor's built-in index serves only Cursor, so Krolik makes sense when several agents need one shared endpoint.

Is Krolik better than Sourcegraph?

They solve different jobs. Sourcegraph is a code search product for humans, while Krolik exposes call graphs, blast radius and review memory to agents over MCP, and you host it. Pick Sourcegraph if developers do the searching and you want zero operations, and pick Krolik if the reader is an AI agent and the code must stay on your infrastructure.

How do you set up Krolik?

Vaelor's site quotes five minutes: docker build, docker run pointed at your LLM endpoint, then register http://127.0.0.1:8897/mcp in your MCP client. Allow longer for code_graph, which needs PostgreSQL with Apache AGE, or semantic_search, which needs pgvector and an embedder. The first understand or call_trace call on a GitHub repo indexes it, which takes a few seconds.

About Krolik

Krolik is a self-hosted MCP server that indexes a codebase into a call graph, so an AI coding agent can see what a change will break before it edits anything. The open-source engine underneath is published as Vaelor: the GitHub repository was renamed from go-code to vaelor, is open source, and showed 1,422 commits and 5 stars when we read it on 21 October 2026. Krolik is the name on krolik.tools, and the README calls Vaelor "the open-source engine behind Krolik".

Read the full background

Tree-sitter parses sixteen languages according to vaelor.ai: Go, Python, TypeScript and JavaScript, Rust, Java, C, C++, Ruby, C#, PHP, Svelte, Astro, Vue, Kotlin, Swift and HTML. Call resolution is type-aware on Go through go/types and, through bundled SCIP indexers, on TypeScript, JavaScript, Python, Java and Rust; other languages fall back to tree-sitter call edges. With PostgreSQL and Apache AGE attached, symbols, calls, imports and test relationships persist as a property graph that code_graph queries in plain English. Adding pgvector and an embedder switches on semantic_search, which blends BM25F, vector search and one-hop graph expansion.

Vaelor's site advertises 35+ MCP tools over Streamable HTTP on one local port, while krolik.tools still lists 28, so run tools/list on your own instance to see the real set. The headline tools are impact_analysis and prepare_change for blast radius, call_trace for callers and callees, review_pr, which stores a verdict per changed symbol that understand returns on later calls, dataflow for taint tracking, rewrite for AST search and replace, and debug_investigate, a seven-phase pipeline from a Prometheus alert and Jaeger traces to a ranked file:function list. One URL registers it in Claude Code, Cursor, Windsurf, Aider and Cline.

Krolik is the wrong pick for some teams. Sourcegraph is built for people searching an organisation's code and needs no graph database, and hosted reviewers such as Greptile and CodeRabbit review pull requests without you operating Postgres or an embedder. Krolik asks you to run those pieces yourself, its documentation is a README plus a small docs site, and 5 GitHub stars means few people have shared setup notes. See the wider field in our AI coding assistants guide and the AI code review roundup, or let Smart Match shortlist for your stack.

Sources and guides

More Agent Skills on Hokai

  • FactIQGives an agent read-only SQL over 100M+ official economic and financial series.
  • spicedb-devPlan, build and test SpiceDB permissions without leaving Claude Code.
  • Skill CreatorDesign and package agent skills as self-contained SKILL.md files