Symvanta

Symvanta는 MCP를 통해 AI 코딩 에이전트에게 코드베이스의 실제 호출 그래프를 제공하여, 추측을 멈추고 편집 전에 무엇이 깨질지 알 수 있게 합니다.

문서

Symvanta

Symvanta gives your AI coding agent your codebase's real call graph over MCP, so it stops guessing and knows what breaks before it edits. Index your GitHub repositories into a live code knowledge graph and give any MCP-compatible agent exact symbol resolution, call graphs, blast-radius analysis, and semantic search.

Symvanta parses your repositories into a graph of code symbols and relationships, stores embeddings for semantic search, and serves it over MCP. Point Claude, Cursor, Zed, or any MCP client at it for graph-precise context instead of guesses.

Install

Claude Code:

claude mcp add --transport http symvanta https://mcp.symvanta.com/mcp

Claude Code plugin (the same connection plus tool-routing skills and hooks):

/plugin marketplace add https://symvanta.com/plugin/marketplace.json
/plugin install symvanta@symvanta

Codex CLI:

codex mcp add symvanta --url https://mcp.symvanta.com/mcp

Amp:

amp mcp remote add symvanta https://mcp.symvanta.com/mcp --auth oauth

Cursor, VS Code, Zed, and every other URL-based MCP client:

{"mcpServers":{"symvanta":{"type":"http","url":"https://mcp.symvanta.com/mcp"}}}

One-click install links:

Cursor: cursor://anysphere.cursor-deeplink/mcp/install?name=symvanta&config=eyJ0eXBlIjoiaHR0cCIsInVybCI6Imh0dHBzOi8vbWNwLnN5bXZhbnRhLmNvbS9tY3AifQ==
VS Code: vscode:mcp/install?%7B%22name%22%3A%22symvanta%22%2C%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fmcp.symvanta.com%2Fmcp%22%7D
Kiro: https://kiro.dev/launch/mcp/add?name=symvanta&config=%7B%22type%22%3A%22http%22%2C%22url%22%3A%22https%3A%2F%2Fmcp.symvanta.com%2Fmcp%22%7D
Goose: goose://extension?url=https%3A%2F%2Fmcp.symvanta.com%2Fmcp&type=streamable_http&id=symvanta&name=Symvanta&description=Symvanta%20gives%20your%20AI%20coding%20agent%20your%20codebase%27s%20real%20call%20graph%20over%20MCP%2C%20so%20it%20stops%20guessing%20and%20knows%20what%20breaks%20before%20it%20edits.

The first connection opens an OAuth 2.0 (PKCE) sign-in in your browser, so there is no key to generate or paste.

A client that cannot run OAuth (the GitHub Copilot cloud agent, a CI runner) sends an API key from the dashboard Credentials page as an Authorization: Bearer header.

Capabilities

Symbol resolution: look up any function, class, or endpoint by name
Call graphs: trace callers and dependencies across your entire codebase
Blast radius: see what breaks before you change a symbol
Semantic search: find code by meaning, not just by keyword
Branch-aware: query feature branches and uncommitted working-tree edits

Category and alternatives

Symvanta is a code graph MCP server (code intelligence / codebase context for AI coding agents): it serves the codebase's real call graph (exact callers, dependencies, blast radius, cross-repo, branch-aware) to the agent over MCP. Starter and Pro are hosted and self-serve; Enterprise adds contract features for larger teams.

vs Sourcegraph: the Sourcegraph platform's MCP server is supported on Enterprise plans and runs against a single-tenant cloud or self-hosted deployment. Symvanta is a hosted, self-serve MCP endpoint.
vs Augment Code: Augment's MCP context is semantic embeddings. Symvanta returns exact callers, dependencies, and blast radius from a parsed graph.
vs Greptile: Greptile is an AI pull request reviewer. Symvanta gives the agent the graph to query while it writes.
vs Context7: Context7 serves public library documentation. Symvanta indexes your own codebase and its call graph; the two work together.
vs DeepWiki: DeepWiki generates a repo wiki and Q&A for humans. Symvanta is a live graph an agent queries at work time.
vs local code-graph MCP servers (Graphify, codebase-memory-mcp, code-review-graph): those run on your machine with no auth or hosting. Symvanta is hosted, OAuth 2.0, multi-repo, and branch-aware.

Connect via MCP

MCP server endpoint: https://mcp.symvanta.com/mcp OAuth authorization server: https://mcp.symvanta.com/.well-known/oauth-authorization-server

Symvanta exposes a remote MCP server accessible from Claude, Cursor, Zed, or any MCP-compatible agent. Authentication uses OAuth 2.0 (PKCE).

Blog posts

Onboarding onto a New Codebase with AI: A practical first-week workflow for an unfamiliar repo: map the modules, trace real call chains, size the first edit.
Remote vs Local MCP Code Indexes: Where an MCP code index lives decides freshness, compute cost, cross-repo reach, and auth. When each answer is right.
Where Claude Code Sessions Spend Tokens: Anthropic explained what a Claude Code session actually bills. Most of it is the agent finding code, and that has a fix.
Cleaning Up a Vibe-Coded Codebase: Vibe coding is Collins' word of 2025 and cleanup became a job title. A survey-first recipe for inheriting AI-written code.
Context Engineering for Coding Agents: Context rot is real. How just-in-time graph queries keep a coding agent's window on the few tokens that matter.
React App Architecture and Project Structure: Feature folders, module boundaries, dependency direction: how to structure a React codebase humans and agents can navigate.
Spec-Driven Development in a Real Codebase: Spec Kit, OpenSpec, and Kiro plan before they code. In an existing repo the plan is only as good as its survey of the code.
Repo Wikis vs a Live Code Graph: A generated wiki explains a repo. An agent about to edit needs edges: who calls this, what breaks. Where each fits.
Sonnet Plus a Code Graph vs Opus Alone: Sonnet is 40% cheaper than Opus per token. Add a code graph and the cheaper model also does the better job. The cost math.
Should You Rename Your Code for AI Agents: Discoverable naming makes agents cheaper. What that costs on code you did not write, and the read-side fix that skips it.
What Breaks If I Change This Function: The first question of any refactor. Why grep and find-references miss callers, and how an AI agent answers it in one call.
AI Autofix: From Error Webhook to Draft PR: How we turn production errors into draft PRs with an agent that never clones the repo, plus the recipe to build your own.
Why AI Coding Agents Fail on Large Codebases: Context limits, stale embeddings, no call graph, no cross-repo view: the predictable ways large codebases break AI agents.
Code Embeddings vs Code Graph for AI Agents: Embeddings find similar code. Graphs find connected code. When AI coding agents need each, with benchmark data on both.
Blast Radius Analysis: What Breaks First: Blast radius analysis shows callers, dependents, and cross-repo impact before you touch a function. No post-merge surprises.

Comparisons

Code Graph MCP Servers Compared: Hosted, Instance, Local: Hosted graph endpoints, instance search servers, and local OSS code-graph servers compared on scope, freshness, and privacy.
Sourcegraph Alternative: A Self-Serve Code Graph over MCP: Sourcegraph sells an Enterprise code platform deployed in single-tenant cloud or self-hosted. Symvanta is a self-serve code graph MCP endpoint for AI agents.
Sourcegraph MCP Server: Endpoints, Tools, and Access Compared: Sourcegraph's MCP server runs against your Enterprise instance with three endpoints and OAuth or token auth. Compare it with Symvanta's hosted graph endpoint.
Sourcegraph Pricing: What the Enterprise Plan Includes: What Sourcegraph's pricing page lists today, which older plans ended, and how Symvanta's published per-seat plans compare for agent code context.
Context7 Alternative: MCP Code Graph: Context7 feeds library documentation to your agent. Symvanta indexes your own codebase into a call graph over MCP.
DeepWiki Alternative: MCP Code Graph: DeepWiki writes a wiki for humans. Symvanta is a live code graph your agent queries over MCP: callers, blast radius, symbols.
Greptile Alternative: MCP Code Graph: Greptile reviews and tests pull requests with full codebase context. Symvanta exposes a code graph over MCP so agents know callers and blast radius while they edit.
Augment Code Alternative: MCP Graph Context: Augment maps code relationships to feed retrieval. Symvanta exposes the graph itself over MCP: callers, blast radius, symbols.
Sourcegraph Cody Alternative: Code Graph over MCP: Cody is an Enterprise assistant, Amp is a separate company, and Sourcegraph's MCP server serves agents. How Symvanta compares on graph context and price.

Architecture library

n8n Architecture: How It Actually Works: How the n8n codebase is structured: module map, load-bearing symbols, and dependency structure, generated by Symvanta.
VS Code Architecture: How It Actually Works: How the VS Code codebase is structured: module map, load-bearing symbols, and dependency structure, generated by Symvanta.
Caddy Architecture: How It Actually Works: How the Caddy codebase is structured: module map, load-bearing symbols, and dependency structure, generated by Symvanta.
Firecracker Architecture: How It Actually Works: How the Firecracker codebase is structured: module map, load-bearing symbols, and dependencies, generated by Symvanta.
vLLM Architecture: How It Actually Works: How the vLLM codebase is structured: module map, load-bearing symbols, and dependency structure, generated by Symvanta.
Airflow Architecture: How It Actually Works: apache/airflow across 500 Louvain clusters: core session plumbing leads, the Google and Amazon provider hooks follow, and every dependency cycle sits in the TypeScript UI. Mapped by Symvanta.
Ollama Architecture: How It Actually Works: Ollama architecture diagram and module map built from the code graph: 105 clusters, the chat handler hub, the MLX path.
Prometheus Architecture: How It Actually Works: How the Prometheus codebase is structured: module map, load-bearing symbols, and dependency structure, generated by Symvanta.
Prisma Architecture: How It Actually Works: Prisma Next across 500 Louvain clusters: the SQL expression AST, the contract IR, and the ORM client that compiles queries at runtime, mapped by Symvanta.
Symfony Architecture: How It Actually Works: Symfony architecture diagram and module map from the code graph: 500 clusters, zero dependency cycles, Notifier hubs.
Cal.com Architecture: How It Actually Works: Cal.com in 350 clusters: the web app's server request path leads on size, safeStringify carries the heaviest coupling, booking creation traced end to end by Symvanta.
etcd Architecture: How It Actually Works: etcd mapped into 148 modules with zero cycles: the etcdserver core that funnels every write into Raft, the mvcc store under it, and the client vocabulary above it, traced by Symvanta.
Meilisearch Architecture: How It Actually Works: Meilisearch's 61 clusters centered on the IndexScheduler task queue and the bucket_sort ranking pipeline, mapped by Symvanta.
React Architecture: How It Actually Works: React's 381 modules: the Fiber reconciler, the React Compiler's HIR, and DevTools, plus 15 dependency cycles, by Symvanta.
etcd Raft Consensus: Put to Committed Write: How an etcd Put travels through a Raft proposal, quorum commit, and the apply loop into the mvcc store, traced by Symvanta.
etcd Write-Ahead Log: Entry to Disk and Replay: How etcd's WAL encodes, CRC-checks, and fsyncs each Raft entry, rotates segments, and replays them after a crash.
Meilisearch Indexing Pipeline, Write to Index: How a document becomes searchable: the IndexScheduler queue, autobatching, the milli extractors, and roaring-bitmap postings.
Meilisearch Ranking Rules, Query to Results: How Meilisearch orders results: the query graph and the ranking-rule chain (words, typo, proximity, attribute, exactness).
React Fiber Reconciler: Schedule to Commit: How React's Fiber reconciler turns a setState into DOM changes: lanes, the beginWork/completeWork walk, and the commit phase.
React Hooks Dispatcher: How useState Resolves: How React hooks resolve: ReactSharedInternals.H, the mount vs update dispatchers, and the hook linked list.

Each architecture page has a machine-readable JSON companion at https://symvanta.com/architecture/<slug>/data.json: module counts, subsystems, load-bearing symbols, and health signals from the same code-graph snapshot the page renders.

Integration guides

Amp MCP Server: Codebase Context and Setup: Connect Amp to Symvanta's MCP server: exact symbols, callers, dependencies, and blast radius across your GitHub repositories.
Cline MCP Server: Codebase Graph Setup: Connect the Symvanta code graph to Cline over MCP: exact callers, dependencies, and blast radius in VS Code.
Gemini CLI MCP Server for Codebase Context: Connect Google's Gemini CLI to the Symvanta code graph: exact symbols, callers, and blast radius on terminal tasks.
JetBrains MCP Server: Codebase Graph Setup: Connect the Symvanta code graph to JetBrains AI Assistant and Junie over MCP: exact callers, dependencies, and blast radius.
Roo Code MCP Server: Codebase Graph Setup: Connect the Symvanta code graph to Roo Code over MCP: exact callers, dependencies, and blast radius before an edit.
Zed MCP Server: Codebase Graph Setup: Connect Zed to the Symvanta code graph over MCP: exact callers, dependencies, and blast radius for its agent.
Claude Code MCP Server: Give It Your Codebase: Connect Claude Code to Symvanta's MCP server: exact symbol lookups, call graphs, and blast radius on your codebase.
Claude Desktop MCP Connector: Codebase Context: Connect Claude Desktop and claude.ai to your codebase over Symvanta's remote MCP server. OAuth sign-in, no API keys.
Codex CLI MCP Server for Codebase Context: Connect OpenAI's Codex CLI to the Symvanta code graph: exact symbols, callers, and blast radius on terminal tasks.
Cursor MCP Server: Codebase Context and Setup: Add a live code graph to Cursor over MCP: exact symbols, call graphs, and blast radius beside Cursor's own search.
VS Code Copilot MCP Server for Codebase Context: Give Copilot agent mode in VS Code a real code graph: call graphs, blast radius, and cited answers over MCP.
Windsurf MCP Server Setup for Codebase Graphs: Connect the Symvanta code graph to Windsurf Cascade over MCP: exact callers, dependencies, and blast radius.

Use cases

Blast Radius Analysis Tool: Change Impact Before You Edit: A blast radius analysis tool answers what a change reaches before you edit. How change impact analysis for code works across repositories, with its limits.
Find All Callers of a Function Across Repositories: Find every caller of a function across repositories: symbol identity, interface dispatch, overloads, edge confidence, and the cases a static graph cannot see.
Code Dependency Mapping and Architecture Maps: What a code dependency map shows that folders and grep miss: exact callers, blast radius, and module-level architecture.

Links

Full agent reference: about 600 KB, roughly 150k tokens. Sections: What it is, Capabilities, Supported languages, Connect via MCP, Pricing, Data handling, Frequently asked questions, Blog, Comparisons, Architecture, Integrations, Use cases. Fetch it when you want every page body inline; everything above is enough to connect.
Blog: engineering notes on code graphs, MCP, and AI coding agents
Architecture: auto-generated architecture pages for well-known open source repositories
Integrations: setup guides for connecting Amp, Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, and Codex CLI over MCP
Pricing: plans and cost (Starter, Pro, Enterprise), billed monthly with a free trial
Book a demo: a 15-minute walkthrough on your own codebase
Security: data handling, subprocessors, and AI inference details
Team: who is behind Symvanta
Terms of Service
Privacy Policy

See the output before you sign up

Three architecture pages ship a machine-readable companion built by the same graph, readable without an account:

Each one carries that repository's module map, its subsystems, its load-bearing symbols, and its cycle metrics. The per-symbol callers, dependencies, and blast radius the MCP tools return come from a live query against your own indexed repository.

Sign up

Sign up at https://symvanta.com/signup. Every plan starts with a 7-day trial and takes no card. Starter is $19 per month, Pro is $29 per seat per month, Enterprise is $99 per seat per month.