# What Maguyva runs on Canonical source: https://maguyva.ai/stack 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). Built for speed, debuggability, and continuous improvement: proven infrastructure at the core, purpose-built components where code intelligence requires them. - PostgreSQL + Lance - Quantized vector search - Code language engine - MCP everywhere - 285+ languages ## What runs in production These are the core layers that stayed after multiple iterations. Every one of them earned its place by making debugging easier or by removing friction from agent execution. ### PostgreSQL + Lance *Database* Tenant-partitioned storage and indexes. Your code stays isolated, with no cross-user index mixing. ### Voyage AI: voyage-4-large *Embeddings* High-recall code embeddings for agent search, retrieval, and grounding. ### Binary Quantized HNSW *Vector Search* Embeddings compressed to 2048-bit signatures and indexed with HNSW. 32x storage reduction, low-latency search, zero extra infrastructure. ### Code Language Engine *Language Engine* A heavily modified Tree-sitter stack with custom grammars, handler heuristics, and 285-language extraction coverage ### MCP Protocol *Protocol* Remote MCP server for Claude Code, Claude Desktop, Cursor, VS Code, Devin Desktop, and other MCP-compatible clients. 11 tools on a TypeScript Cloudflare Worker with native streamable HTTP transport. ### Lance columnar + R2 *Cache* Columnar-cached, cloud-synced embeddings ## The stack after the dead ends Maguyva did not arrive at this shape all at once. Each layer survived because it made the system faster to reason about, cheaper to operate, or easier to debug under pressure. The orchestration experiments started earlier. This timeline starts in September 2025, when the Maguyva product itself snapped into focus. - **2025-Q3** (foundation): Focused Maguyva build starts → Voyage voyage-code-3. The product direction hardens and the rapid iteration loop begins - **2025-Q3** (storage): Binary vector retrieval layer. Quantized search with compact indexes, no separate vector stack - **2025-Q4** (ast-engine): Code language engine comes online. Custom queries, parser fallbacks, and handler-based heuristics across 285 languages - **2025-Q4** (search): Multi-modal fusion search. Ranked retrieval across semantic, text, AST, and graph indexes in a single query - **2026-Q1** (caching): Lance columnar + R2 sync. Server-side columnar cache, faster cold starts - **2026-Q1** (quality): 285-language validation framework. Automated extraction testing across every supported language - **2026-Q1** (launch): Public launch at maguyva.ai. 7 months from focused build to public availability - **2026-Q2** (protocol): TypeScript Cloudflare Worker MCP. Native streamable HTTP transport, lower latency, swapped under live traffic - **NOW** (embeddings): Production embeddings → voyage-4-large. Higher-recall semantic search after cutover from voyage-code-3 ## We run on what we ship Maguyva is built by agents that run on Maguyva. Our orchestration system coordinates specialist AI agents that use the same MCP tools, search indexes, and production infrastructure you get. There is no internal version. The team page lists “1x Maguyva ($149) — we pay ourselves.” That is not a joke. It is the invoice. // customer zero ships to production daily ## The rules that survived production These are the constraints that stayed useful in production. They are less about ideology and more about keeping the system understandable while agents are making changes at speed. ### No Silent Fallbacks Fail fast, fix fast. Every error surfaces immediately. No graceful degradation hiding bugs in production. *// panic > silent corruption* ### RPC-First Database-side logic for performance. PostgreSQL functions execute close to data, eliminating round trips. *// PostgreSQL functions, 0 ORMs* ### CQRS Separate write and read paths. Indexing pipeline writes, MCP tools read. Different optimization strategies. *// writes: batched | reads: cached* ### Deterministic UUIDs Cache coherence without database lookups. Same input = same ID. Idempotent operations by design. *// uuid5(namespace, content_hash)* The best architecture is the one you can debug at 3am. (We've tested this.) // last updated: 2026-10 [Try it on your repo](https://app.maguyva.ai/auth/sign-up?plan=free) ## Where your code lives No mystery. Here's exactly what happens to your code when you connect a repo. - **Metal servers, US East** (Compute): Bare-metal machines in Ashburn, tuned for indexing throughput instead of noisy general-purpose instances. - **Cloudflare global CDN** (Edge): API and MCP responses cached at the edge. Static assets served from Cloudflare's global network. - **Tenant-partitioned per workspace** (Isolation): Each workspace gets tenant-scoped repository indexes across the pipeline and database. We use customer code to provide code intelligence, not to train models or sell data. - **Singapore incorporated** (Entity): UT International PTE. LTD. Real jurisdiction, real entity. - **Remove a repo, remove the index** (Deletion): Disconnect a repo and Maguyva removes active indexed source code, embeddings, and graph data from the service. Residual backups expire on the retention schedule. - **Need EU or AU?** (Regions): Additional deployment regions are a conversation for customers with residency requirements. ## Next Steps - [Get started](https://maguyva.ai/docs): Index your first repo in minutes. - [See pricing](https://maguyva.ai/pricing): Per-workspace, not per-seat. Free to start. - [Meet the team](https://maguyva.ai/team): 2 humans and a swarm of AI agents. - [Read the blog](https://maguyva.ai/blog): Technical notes from the build.