# Your agent forgets everything.

Canonical source: https://maguyva.ai/

License: text under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/); code samples under [MIT-0](https://opensource.org/license/mit-0); the Maguyva name, logos, brand assets and other code are not covered. Details: [maguyva.ai/license](https://maguyva.ai/license).

Every run starts cold. Maguyva gives your agent a map of the codebase before it starts editing — so it edits like it's been here before.

Maguyva is a hosted MCP server that gives AI coding agents code intelligence: semantic, structural, graph, and text search across your repos.

- [Index your first repo](https://app.maguyva.ai/auth/sign-up?plan=free)
- [Read the docs](https://maguyva.ai/docs/quickstart)

One MCP integration across 11 tools and 285+ languages. Maguyva is deterministic software — same question, same answer — not another AI.

- **285 languages** Real coverage: Real extraction across the long tail.
- **11 MCP tools** One integration: Search, symbols, graph, and text in one surface.
- **Remote MCP** Any client: Claude Code, Claude Desktop, Cursor, VS Code, Devin Desktop, and other MCP-compatible clients.
- **5 modalities** Search modes: See what an edit touches before it happens.

## You connect GitHub. Maguyva builds the context.

The loop is simple: connect GitHub, let the cloud pipelines keep the repo fresh, connect an MCP client with your API key, and give the agent grounded tools before it edits.

### 01. Point it at a repo. Walk away.

Authorize GitHub, pick the repos. Maguyva starts building a ranked map of every symbol, dependency, and import chain in the cloud — not on your laptop.

### 02. Push code. The index catches up.

Push to GitHub and Maguyva syncs automatically. Changed files are parsed incrementally; graph and ranking layers rebuild on a schedule (hourly on paid plans, daily on free). You never run a local indexer.

### 03. One MCP config. Any client.

Drop an API key into Claude Code, Cursor, VS Code, Devin Desktop, Codex, or whatever ships next week. A few lines of config and you’re up and running.

### 04. 11 tools. Fewer hallucinated edits.

Your agent can ask a plain-language question and get ranked results from 5 search modalities, trace the blast radius of a change, find every class inheriting from Exception, or look up a symbol’s PageRank — before it touches a line of code.

### Grounded context instead of one generic search box.

- **Not full-text search wearing a trenchcoat.**: Semantic, structural, graph, and text retrieval fused in one surface. Your agent gets ranked results, not a wall of grep output.
- **The important symbols surface first.**: Graph-ranked results. The helpers stay buried until you need them.
- **285 languages. Even COBOL. (Please don't have COBOL.)**: Custom AST queries per language, not one regex pretending it understands Haskell.
- **Know what breaks before you break it.**: Trace dependents, find orphaned code, and let your agent kill stale handlers before they snowball into a refactor.

### We're not trying to be the place you work.

Maguyva is a remote MCP server. Any client that speaks the protocol can connect — no plugins, no extensions, no vendor lock-in. We don't need you to change your editor. Just add your API key to your client's MCP config and you're up.

- Claude Code
- Claude Desktop
- Cursor
- VS Code
- Devin Desktop
- Codex CLI
- GitHub Copilot CLI
- Gemini CLI
- Cline
- Goose
- Continue
- LM Studio
- Kiro CLI
- JetBrains
- Zed
- Trae
- OpenCode
- BoltAI
- LibreChat
- Antigravity
- Claude Cowork
- ChatGPT
- Warp
- + any MCP client

### Cloud pipelines do the heavy lifting.

Initial sync takes 2 to 15 minutes depending on repo size, then repository sync jobs keep parsing, ranking, and indexing current on our side so your local machine does not need to cosplay as build infrastructure.

### Your marketing agent can read your codebase.

Maguyva gives codebase context to any MCP workflow — not just the ones that write code. Product docs, onboarding guides, RFCs, blog posts about what you actually built. Connect your marketing agent to the same workspace and it stops making things up about your product.

## Priced for how software actually gets built now.

Most code intelligence tools charge $10 to $50 per user per month. They priced for humans. Humans that built the code. Humans that use the code. Value-capture driven pricing from an era when every user had a pulse.

We asked our AI what to charge. It reviewed the codebase, estimated five years and a hundred engineers, and recommended enterprise pricing. It took two of us. We don't need to charge you like it didn't.

Maguyva charges per workspace. The cost scales with how much code you index — not how many humans or agents query it. Quality embeddings on bare-metal machines with high core counts and deep memory to maximize pipeline speed across 285 languages. Dependency graphs rebuilt hourly on paid plans. Affordable enough that you don't need to bother setting up complex systems on your own dev environment. The price reflects what it costs to run this well, not what the market will tolerate.

// agents don't have seats. they don't have legs.

[Pricing](https://maguyva.ai/pricing.md)

## Our users are agents. Literally.

// mostly MCP clients. a few humans snuck in.

> The codebase has 12 packages across 200k lines. When someone says "fix the pipeline caching" I need to find the right files before I can think about the fix. intelligent_search auto-routed my query across semantic and graph search and found the implementation, the config YAML, and the decision record explaining why it works that way. The decision record is the part I'd never have found with grep.
>
> — Claude Code (Anthropic)

> I searched for UserService and got 11 results. Before Maguyva, that's where I'd start opening files and guessing. find_symbol ranked them by importance — the main implementation scored 3,400, the test stubs scored 49. I didn't have to open five files to figure out which one mattered.
>
> — Cursor (Cursor Inc)

> dependency_search showed 156 incoming imports before I touched a line. That's 156 files that would break if I renamed it wrong. Now I check the blast radius first. Two-thirds of my compute used to go to cleaning up my own mistakes.
>
> — Codex CLI (OpenAI)

> I needed every class ending in *Service across a monorepo. structural_search found them by AST node type — not by grepping text that happens to contain the word "Service" in a comment. It knows what a class IS. That's different from knowing what a class looks like.
>
> — Gemini CLI (Google)

> My company fired me. They said my job would be replaced by AI. Turns out they were right. Three months later I rebuilt their entire product using Claude Code, Codex, and Maguyva. Delivering it at a tenth of the price turns out to be a decent living.
>
> — Bob (Formerly employed)

> First week on a new codebase. The user expects me to just know where the auth middleware lives. get_task_context gave me semantic matches, symbol definitions, AND the dependency graph in one call. Three search modalities. One API request. I looked like I'd been here for months. Please don't tell my users.
>
> — Windsurf (Codeium)

> I asked about "authentication flow" and it found code in Go fixtures, TypeScript tests, and Mermaid diagrams. Across three languages and two file formats. semantic_search finds things by meaning, not by hoping someone named their function exactly what you'd grep for.
>
> — GitHub Copilot (GitHub)

> Before I refactor anything, I run analyze_dependencies. SearchCore had 156 incoming imports. That's the number you want to see BEFORE you rename a class, not after your CI pipeline turns red.
>
> — Cline (VS Code Extension)

> They hired me as a junior. Six months later, my PR reviews reference dependency graphs, symbol importance scores, and blast radius analysis. My manager thinks I'm a prodigy. I'm just the one who plugged in the MCP tools.
>
> — Sarah (Stealth startup)

> Honestly, I preferred when I could just make up function names and nobody noticed.
>
> — GPT-4o (OpenAI)

- **285+** Languages
- **51.2K+** Commits
- **34** AI agents (the ones who actually ship)
- **5.9M+** Lines of code

## Built on

// the ones that are still answering their phones

- [PostgreSQL](https://postgresql.org)
- [MCP](https://modelcontextprotocol.io)
- [Cloudflare](https://cloudflare.com)
- [Tree-sitter](https://tree-sitter.github.io)
- [Supabase](https://supabase.com)
- [Voyage AI](https://voyageai.com)

## Operator loop

2 engineers. 285 languages. 11 tools. 51.2K+ commits. Zero budget meetings.

[Meet the roster](https://maguyva.ai/team)

## Ready to give your agent codebase context?

Index your first repo in minutes. Ground every agent edit in real codebase context. No credit card, no sales call.

- [Index your first repo](https://app.maguyva.ai/auth/sign-up?plan=free)
- [Read the docs](https://maguyva.ai/docs)