LLM 토큰을 소모하지 않고 코드 영향도, 호출 그래프, 심볼 관계를 파악할 수 있는 AI 에이전트용 구조 분석 도구를 제공합니다.
다른 이름 GhostlyGawd/codeweb, @ghostlygawd/codeweb
- 스타
- 0
- Tools
- —
- 다운로드, 30일
- 282
- 최신 버전
- 0.15.0
- 마지막 커밋
- 2026년 11월 9일
- 라이선스
- MIT
codeweb MCP란?
LLM 토큰을 소모하지 않고 코드 영향도, 호출 그래프, 심볼 관계를 파악할 수 있는 AI 에이전트용 구조 분석 도구를 제공합니다.
프로젝트 README에서 생성된 요약
codeweb MCP server 정보
Codeweb은 AI 에이전트가 호출 그래프와 심볼 관계를 통해 코드 변경의 영향도를 이해할 수 있도록 설계된 구조 분석 도구를 제공합니다. 계정, 서버 또는 텔레메트리 없이 로컬 머신에서 완전히 작동합니다. 이 시스템은 정적 분석을 수행하여 함수, 클래스 및 메서드를 매핑하므로, 에이전트가 grep과 같이 정확도가 낮은 방식에 의존하는 대신 코드베이스 구조를 직접 쿼리할 수 있습니다.
개발자는 이 MCP server를 사용하여 codeweb_impact, codeweb_callers, codeweb_find_similar와 같은 28가지 특화된 도구를 에이전트에게 제공할 수 있습니다. 이러한 도구는 범위가 제한된 구조적 답변을 반환하므로, 에이전트의 컨텍스트 창을 작게 유지하고 구현에만 집중할 수 있도록 돕습니다. 분석 과정이 결정론적이며 매핑을 위해 LLM 토큰을 사용하지 않기 때문에 대규모 저장소에서도 효율적이고 비용 효율적입니다. 이 서버는 약 3초 만에 3,000개의 심볼을 매핑할 수 있으며, 중복된 로직, 데드 코드 및 도메인 결합에 대한 명확한 뷰를 제공합니다.
활용 사례
- 1코드 변경 영향도 평가AI 에이전트가 코드를 작성하기 전에 특정 함수 수정이 미칠 영향 범위를 결정할 수 있습니다.
- 2모든 함수 호출자 식별개발자는 의존성 파괴를 방지하기 위해 특정 함수를 호출하는 모든 심볼을 찾는 데 이 서버를 사용할 수 있습니다.
- 3로직 중복 감지이 도구는 코드베이스 전반에서 서로 다른 함수가 동일한 작업을 수행하는 사례를 식별하는 데 도움을 줍니다.
- 4구조적 회귀 테스트 자동화gate 기능을 사용하여 의존성 사이클이나 고립된 심볼을 유발하는 pull request를 차단할 수 있습니다.
프로젝트의 README를 바탕으로 작성되었습니다. 모든 활용 사례는 해당 내용의 인용문과 대조하여 확인합니다.
codeweb MCP 설치 방법
형식 확인: 2026년 11월 10일Cursor에 추가Goose에 추가Claude Code에 codeweb MCP 설치
$ claude mcp add codeweb -- npx -y @ghostlygawd/codewebCursor에 codeweb MCP 설치
~/.cursor/mcp.json
{
"mcpServers": {
"codeweb": {
"command": "npx",
"args": [
"-y",
"@ghostlygawd/codeweb"
]
}
}
}VS Code에 codeweb MCP 설치
.vscode/mcp.json
{
"servers": {
"codeweb": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@ghostlygawd/codeweb"
]
}
}
}VS Code CLI에 codeweb MCP 설치
$ code --add-mcp '{"name":"codeweb","type":"stdio","command":"npx","args":["-y","@ghostlygawd/codeweb"]}'Claude Desktop에 codeweb MCP 설치
claude_desktop_config.json
{
"mcpServers": {
"codeweb": {
"command": "npx",
"args": [
"-y",
"@ghostlygawd/codeweb"
]
}
}
}Codex에 codeweb MCP 설치
$ codex mcp add codeweb -- npx -y @ghostlygawd/codewebCodex config.toml에 codeweb MCP 설치
~/.codex/config.toml
[mcp_servers.codeweb] command = "npx" args = ["-y", "@ghostlygawd/codeweb"]
Gemini CLI에 codeweb MCP 설치
$ gemini mcp add -s user codeweb npx -- -y @ghostlygawd/codewebGemini settings.json에 codeweb MCP 설치
~/.gemini/settings.json
{
"mcpServers": {
"codeweb": {
"command": "npx",
"args": [
"-y",
"@ghostlygawd/codeweb"
]
}
}
}자동 검사
2026년 11월 14일에 실행됨전체 보고서 →자동 검사를 통과했습니다. 이는 보안 감사가 아닙니다. 코드를 직접 읽고 서버에 필요한 최소한의 권한만 부여하십시오. 검사 방식
- 소스 저장소 연결 및 접근 가능ghostlygawd/codeweb is public and reachablegithub.com/ghostlygawd/codeweb ↗통과
- 패키지가 이 저장소를 소스로 명시@ghostlygawd/codeweb names ghostlygawd/codeweb as sourcenpmjs.com/package/@ghostlygawd/codeweb ↗통과
- 빌드 출처(Sigstore 증명)Sigstore attestation for @ghostlygawd/codewebnpmjs.com/package/@ghostlygawd/codeweb/v/0.15.0#provenance ↗통과
- 오픈 소스 라이선스 (OSI)MITspdx.org/licenses/MIT.html ↗통과
- 최근 1년 내 커밋 또는 릴리스Last commit or release 2026-11-09github.com/ghostlygawd/codeweb ↗통과
- OpenSSF Scorecard검사되지 않음
- 패키지 이름이 인기 패키지와 유사하지 않음No popular package one edit away from @ghostlygawd/codeweb통과
- 패키지 및 해당 종속성에 중요한 보안 권고 없음2 resolved dependencies checked; no critical advisoryosv.dev ↗통과
- 모델에 대한 숨겨진 지침 없음0 tool descriptions and the README read; no override phrases, credential paths or hidden characters통과
- 파괴적인 도구에 라벨 표시됨검사되지 않음
- 원격 엔드포인트가 MCP 핸드셰이크에 응답검사되지 않음
README
GitHub에서 보기Free & MIT-licensed. Runs entirely on your machine — no account, no server, no telemetry. Reads your code; never executes it.
Website · Gate every PR · See it in action · Install · Use · For agents (MCP: Model Context Protocol) · How it works · Free forever · Changelog
See what your AI edits affect.
npx -y @ghostlygawd/codeweb .
Structural checks for AI code changes. codeweb's gate builds the call graph before and after a change, then fails the pull request on three regressions: a new dependency cycle, a new body-confirmed duplication, or an existing non-exported symbol that becomes an orphan with no mapped callers. Gate conditions and limits.
No LLM is in the checks. The gate uses zero model tokens for its static analysis. A passing check does not prove the program works. Analysis can miss dynamic relationships; inspect unresolved results and run the relevant tests. Pure removals pass the structural rules.
The verdict uses the mapped graph, not a model, and costs zero LLM tokens. Exported symbols are exempt
from the diff’s orphan check. Inspect analysis.checks: skipped
checks are not passes, and refresh drops the evidence needed to evaluate duplication.
A green result does not establish complete dependency coverage or behavioral correctness. Pure removals do not count as structural regressions. Put it on your PRs →
The gate needs a map to compare, and that map is worth reading on its own.
codeweb reads your code. It maps supported functions and the calls it can resolve between them. It maps 3,000 symbols in approximately 3 seconds. Static analysis produces the same map from the same code. No LLM is in the mapping loop.
Static extraction can omit unsupported layouts, ambiguous calls, and runtime dispatch; the map is not a completeness guarantee.
Your coding agents query the map instead of using grep. In measured tests, agents that used grep missed more than half of a function's real callers (see the measurements). An incomplete caller list can cause an agent to break code that it did not inspect.
- Give agents structural data: The Model Context Protocol (MCP) server provides 28 tools,
including
codeweb_impact,codeweb_callers, andcodeweb_find_similar. - Keep answers small: Each query returns a bounded structural answer. Your agents can use their remaining context for implementation work.
- Inspect the same data: The interactive report shows the extracted codebase map.
The map also shows relationships that are not visible in one file. These relationships include duplicated logic, dead code, hotspots, and tangled domains.
Try it on your repo
cd your-project
npx -y @ghostlygawd/codeweb .
For a repository with 3,000 symbols, the first map takes approximately 3 seconds. Open
.codeweb/report.html to inspect the map.
Gate every pull request
From a Codeweb source checkout, run the verdict locally; the GitHub Action can run it on every pull request:
node scripts/ci-gate.mjs --base origin/main --target src # exit 1 on a structural regression
The gate builds the graph from the pull request base and from its head, then diffs them. It posts a sticky structural review comment either way, so reviewers see the blast radius without installing anything.
Pin the Action to a release tag, not a moving branch: a floating ref can change your gate's
verdict semantics under you. Pin codeweb-ref to the same tag so the engine is fixed too. The
workflow YAML, the monorepo matrix form, and every input live in
docs/ci-gate.md.
Want the gate hosted — no workflow YAML, cached base graphs, and history across every repo in the org? That is the planned scope of codeweb Teams, the paid half of the boundary. Running it yourself stays free forever.
See it in action
Each screenshot below shows a generated report for axios (280 symbols, 7 domains). The screenshots are not mockups.
codeweb found 3 real duplications in axios and rejected 12 false positives. Read the case study, or inspect the live map.
Know what an edit breaks — before you write
Select a function in the live map. The map highlights the function's blast radius and shows the symbols that the change can affect.
Your agents can get the same answer from the codeweb_impact MCP tool before they edit the code.
Navigate the whole system
The force-directed map shows the extracted symbols. You can collapse symbols into domains. Search, drag, zoom, or select a node to trace its callers and dependencies.
Findings — stop guessing what to refactor
The Findings tab ranks duplication, highly connected hotspots, and likely dead code. Select a row to inspect the symbol's callers and dependencies.
See duplication density, and where domains tangle
codeweb works at symbol resolution. It maps functions, classes, methods, and the call and import edges between them. A file-level scanner can show that two modules are similar. codeweb can show that two functions do the same work, identify their callers, and calculate the effect of a merge.
Benchmarks
- Find callers before an edit: In the v0.9.0 efficiency pilot, agents found 74% of a function's real callers with codeweb and 44% with grep at the same context spend. A missed caller can cause an edit to break working code.
- Calculate the effect of a change: One codeweb call returned one small answer. A simulated graph-assisted grep loop used 126 times the context size in the recorded comparison. This measures context-size differences, not total agent-session token savings.
- Detect duplicate code: codeweb found every planted duplicate with zero false alarms, including renamed copies. Text search found 0% of the renamed copies.
- Check deterministic results: Tests compared codeweb with the TypeScript compiler and other independent implementations more than 490,000 times, with zero disagreements.
- Map and query quickly: The first map takes approximately 3 seconds for a repository with 3,000 symbols. Queries take approximately 0.1 seconds. A repository with twice as many symbols takes approximately 1.3 times as long to map.
- Understand the limits: A new map after a very large edit can take more time. Agents also completed simple tasks successfully without codeweb.
Methodology, raw data, and per-claim receipts:
the evidence ledger. Benchmark your own
repo from a Codeweb source checkout: npm run bench -- <path>/.codeweb/graph.json. CI re-runs the performance budgets on
every PR; breaking a published number fails the build.
codeweb also keeps a local activity tally. From a Codeweb source checkout, run npm run stats to see it:
codeweb this month: 41 pre-edit card(s) · 5 card-named caller(s) followed · 2 regression(s) flagged · 120 queries served
To evaluate a dependency, point codeweb at a repository that you do not own:
/codeweb https://github.com/owner/repo. codeweb makes a read-only clone, maps the clone, and
adds an adoption review. codeweb does not execute the target code.
Install
Free & MIT-licensed. Runs entirely on your machine — no account, no server, no telemetry. Reads your code; never executes it.
- codeweb requires Node.js ≥ 22.
- codeweb has zero required dependencies. CI verifies operation with an empty
node_modulesdirectory. - The optional
web-tree-sitterwasm grammar improves extraction. codeweb does not require it. - CI publishes releases with npm provenance. Run
npm audit signaturesto verify a release.
Using Claude Code? Install the plugin to add the /codeweb command, automatic pre-edit
impact cards, and all 28 tools:
/plugin marketplace add GhostlyGawd/codeweb
/plugin install codeweb
Restart Claude Code to register the /codeweb command, agents, and skill.
Using Codex? Install the bundled coding skill and local MCP server:
codex plugin marketplace add GhostlyGawd/codeweb
codex plugin add codeweb@codeweb
Reload Codex, then use $codeweb to inspect callers before an edit, compare existing
implementations, or review a structural change. The plugin uses your existing tool permissions.
The plugin/MCP workflow is verified with Codex CLI 0.159.2 and 0.160.0. If a client does not expose bundled tools, use the standalone MCP recipe below with the coding skill. Manage each configured entry when disabling or removing Codeweb. This workflow does not claim automatic hook delivery.
Choose your MCP client: The setup page provides one recipe for Claude Code, Cursor, Windsurf, Gemini CLI, or Codex. Print a recipe and check local setup from your project:
npx -y @ghostlygawd/codeweb setup --client cursor
npx -y @ghostlygawd/codeweb doctor
Setup prints configuration without replacing files. Doctor checks the local server and graph;
a successful local check does not prove an editor connection. Restart your client, ask your
agents to run codeweb_callers, and check a returned source location.
Map a repository without an AI agent: Run one command from your project directory:
cd your-project
npx -y @ghostlygawd/codeweb . # ~3 s for 3,000 symbols — then open .codeweb/report.html
For a temporary evaluation, use the npx command. It creates the map without a permanent
installation.
Run the engine from a clone:
git clone https://github.com/GhostlyGawd/codeweb.git
node codeweb/scripts/run.mjs /path/to/your/project
docs/cli.md lists each executable, flag, and exit code.
VS Code: editor/vscode-codeweb shows an
N callers · blast M lens above each mapped symbol. Select the lens to open the report.
What you can do
Each link lands on full docs, flags, and examples in the reference.
- Know before you edit: Find callers, calculate the effect of a change, and check for an existing implementation. → Query the graph · context & pre-flight
- Gate every edit: Get a structural regression result for an edit, pull request, or
architecture rule.
→ The
diffverdict · the PR gate · the capability suite - Clean up, ranked: Rank consolidation and dead-code work by evidence.
→
optimize·hotspots·campaign·trend
Use
/codeweb # map the current project
/codeweb src/payments --depth symbol # deep-dive one subsystem
/codeweb https://github.com/owner/repo # external review before adopting
/codeweb owner/repo --open # clone, map, and open the report
Available flags include --depth module|symbol|auto, --engine hybrid|read|tools,
--focus <glob>, --mode internal|external, and --open. See commands/codeweb.md for details.
codeweb writes all outputs to <target>/.codeweb/. Agents and other tools can read graph.json.
You can open report.html. codeweb also creates Markdown versions.
See the description of each output file.
Use it as an MCP tool
scripts/mcp-server.mjs is a zero-dependency Model Context Protocol (MCP) stdio server. It gives
MCP clients access to 28 tools. The tools help the client orient, read the structure,
check before writing, gate an edit, and plan cleanup.
The September 14 packaged assessment exercised the baseline loop with Codex CLI 0.154.0; it did not certify every tool in every client. Claude runtime compatibility and other clients were not tested in that assessment. See the recorded scope.
The plugin registers the server automatically. To register the standalone server, run:
claude mcp add codeweb -- npx -y -p @ghostlygawd/codeweb codeweb-mcp
Requires Codeweb 0.15.0 or later. Use the installed MCP server, or register a source checkout using its absolute path as shown in the reference.
The loop an agent runs:
- Orient with
codeweb_brief(orcodeweb_findwhen no symbol name is known). - Before editing, call
codeweb_refresh {baseline:true}once to capture fresh source. Askcodeweb_explain, thencodeweb_context,codeweb_impact, orcodeweb_dependentsabout the symbol before changing it. - After editing, call
codeweb_diff {before:"baseline",refresh:true}. - If red, repair the code and repeat that same diff against the same baseline.
Do not capture another baseline during repair:
baseline:truereplaces the original. Ordinary and automatic refreshes preserve it. Start a new baseline for the next task.
Known unsupported same-line JS/TS declarations produce status: "inconclusive" and
ok:false instead of a clean gate. Inspect the diagnostic locations and callers in source.
The detector covers known layouts; no diagnostic does not prove complete extraction.
Read analysis.checks: a skipped check is not a pass. Refresh drops overlap evidence, so
this loop does not evaluate duplication; structural green does not prove behavioral
correctness. Run the project's tests too. If the baseline is missing, edited source cannot
reconstruct it — recover the pre-edit source before beginning again.
For npm v0.14.0, use the snapshot quickstart. For the unreleased checkout, try the tiny cycle → repair → green walkthrough. Clients that hide the server's built-in instructions can paste the rules snippet instead.
The server includes these agent-specific features:
- Optional
graphargument: The server finds the nearest map when you omitgraph. If no map exists, the error directs the agent tocodeweb_map. - Budgeted responses: Responses include the highest-ranked items and the true totals. A context response that was approximately 300 KB is now approximately 10 KB.
- Staleness information: A stale result identifies its state and directs the agent to
codeweb_refresh— on the orient tools and on every spawned advisor answer.
All 28 tools, grouped and explained →
How it works
For JavaScript, TypeScript, Python, Rust, Go, Java, C#, Ruby, PHP, Kotlin, Swift, C, and C++, codeweb uses a deterministic Node pipeline by default. One command creates the map. No LLM is in the pipeline, and the same input produces the same bytes.
The map pipeline has the four stages in the following diagram. scripts/run.mjs also creates
optimize.md after overlap analysis and before report rendering.
- Extract (
extract-symbols.mjs) parses each source file into atomic nodes such as functions, classes, and methods. It also records call and import edges. If a bare call can refer to more than one definition, codeweb omits the edge instead of guessing. Per-file caching makes extraction incremental and byte-identical to a full rebuild. An imported.jsonfile enters the map as a file-level node without being parsed. An unreferenced.jsonfile stays out of the map, which prevents lock-file noise. - Cluster (
cluster3.mjs) removes genuine utility hubs and groups the remaining nodes into directory-anchored semantic domains. - Overlap (
overlap.mjs) detects duplicated logic and parallel implementations. It compares each candidate with the actual function bodies by using token-shingle similarity. This check prevents name coincidences from becoming findings. A structural pass over identifier-normalized skeletons also finds renamed Type-2 clones (find-similar --structural). - Render (
build-report.mjs) convertsgraph.jsoninto the self-containedreport.htmlandreport.mdfiles.
For a language that the extractor cannot parse, codeweb uses the agent path.
codeweb-dissector agents extract nodes and edges for each subsystem. codeweb-domain-mapper
then assigns domains and overlaps.
Both paths produce the same graph.json schema. In external mode, each path also adds an
adoption verdict.
Versus a language server (LSP): an LSP answers one hop on demand — definitions, direct references — inside an editor session. codeweb builds one deterministic whole-graph artifact: transitive impact, duplication with body evidence, dead code, and domain coupling.
Agents query that artifact over MCP, and CI diffs it to gate a PR. The two compose — codeweb replaces the grep loop, not your language server.
Choose a reading path in the documentation index, which also explains repository ownership. The component map details the engine.
Roadmap
- Support more first-class languages: codeweb currently supports thirteen native languages:
JavaScript, TypeScript, Python, Rust, Go, Java, C#, Ruby, PHP, Kotlin,
Swift, C, and C++. Other languages use the agent fallback. Dynamic-dispatch AST tiers
cover JS/TS, Java, C#, Python, Go, Rust, Ruby, PHP, C, and C++. Kotlin and Swift
dispatch requires a trusted wasm grammar at the pinned ABI. See
scripts/grammars/PROVENANCE.md.
Recent releases added the agent-intelligence suite (hotspots, campaign, reading-order, Type-2 clone detection, and suppression memory), a live interactive demo, Go and Rust on the fast path, duplication trend data, and the one-command CI regression gate with a GitHub Action. codeweb currently provides 28 tools.
codeweb MCP 질문
codeweb MCP를 Claude Code에 추가하려면 어떻게 하나요?
터미널에서 claude mcp add codeweb -- npx -y @ghostlygawd/codeweb를 실행합니다. 모든 프로젝트에서 사용할 수 있게 하려면 --scope user를 추가합니다.
codeweb MCP를 Cursor에 추가하려면 어떻게 하나요?
~/.cursor/mcp.json를 열고 mcpServers 아래에 서버를 추가합니다. 위 설치 섹션에 표시된 Cursor 구성을 사용합니다.
codeweb은 분석을 위해 LLM을 사용하나요?
아니요, 이 도구는 정적 분석을 사용하며 매핑이나 구조 확인을 위해 LLM 토큰을 전혀 사용하지 않습니다.
호출자를 찾는 데 있어 grep을 사용하는 것과 비교하면 어떤가요?
측정된 테스트 결과에 따르면, 동일한 컨텍스트 비용을 사용했을 때 grep을 사용한 에이전트는 실제 호출자의 44%만 찾아낸 반면, codeweb을 사용한 에이전트는 74%를 찾아냈습니다.
제 코드가 서버로 전송되거나 텔레메트리에 사용되나요?
아니요, 이 도구는 계정, 서버, 텔레메트리 없이 사용자의 로컬 머신에서 완전히 실행됩니다.
gate 기능은 어떤 종류의 구조적 회귀를 감지하나요?
gate는 새로운 의존성 사이클, 새로운 본문 확인 중복(body-confirmed duplications), 또는 고립된 상태가 된 기존 비내보내기(non-exported) 심볼을 식별합니다.