Skip to content

> tail -f /var/log/shipped

Changelog

What changed, when it shipped. Outcomes, not implementation details.

June 2026

Sharper retrieval

  • Upgraded code embeddings to voyage-4-large — the top-ranked code model on the independent RTEB Code benchmark. Sharper semantic search, swapped in with zero downtime
May 2026

The grind

  • Language quality campaign across ~280 languages — an autonomous loop spot-checks extraction against real-world repositories, an LLM judge classifies every sampled edge, and a fleet of agents fixes what it finds
  • Hundreds of extraction precision fixes landed: scope attribution, false-positive calls and reads, line-span metadata — Rust, Go, Swift, Ruby, C++, Elixir, Julia, Haskell, and dozens more
  • Honest coverage accounting — relationship types a language genuinely can’t express are now documented as not-applicable instead of silently skipped
  • New on the blog: “Language Recursive Self-Improvement” — how we grind extraction quality at this breadth without a human in every loop
April 2026

Engine swap

  • MCP server rewritten as a TypeScript Cloudflare Worker — lower latency, native MCP HTTP transport, swapped in under live traffic
  • intelligent_search rebuilt around intent classification — each query is routed to the modalities that fit it, then fused with weighted Reciprocal Rank Fusion across semantic, text, structural, and graph results
  • Community context enrichment — search results can now pull in related symbols from the same code neighborhood
  • Structural search presets — functions, classes, methods, variables, imports — plus auto-detected patterns like def foo(
  • Cursor-based pagination and a response metadata envelope (timing, applied filters, routing diagnostics) on every tool
  • ask_maguyva v3 — slimmed from a sprawl of discrete tools to 4 operations and 2 parameters
  • Tree-sitter language pack 1.4 with download-on-demand grammars
  • Site: per-page social preview images, FAQ structured data on docs pages, a WCAG AA contrast pass, and reduced motion respected on the live dashboard
March 2026

Public beta

  • 229 languages — added Zig, Nix, QML, and 10 others since pre-launch
  • ask_maguyva — grounded Q&A tool. Ask natural language questions, get answers with cited evidence from your codebase
  • Expression evaluator replaces discrete math tools — compound calculations in a single call
  • Rewritten 4-step onboarding flow on homepage
  • MCP protocol explainer added to hero for newcomers
  • StatCounter renders real values for screen readers and search bots
  • Fixed: hero headline animation now settles on “knows” instead of looping
February 2026

Pre-launch

  • 11 MCP tools across 5 search modalities — semantic, structural, graph, text, and grounded search
  • 217 languages with real AST extraction via tree-sitter
  • Installation guides for 8 MCP clients — Claude Code, Cursor, VS Code, Windsurf, and more
  • Per-workspace pricing — Free, Operator ($19), Crew ($49), Fleet ($149)
  • Interactive demo panel on homepage showing grounded evidence
  • 100% precision across the top 50 languages
January 2026

Foundation

  • Core MCP server with remote indexing pipeline
  • Graph intelligence — symbols ranked by architectural importance using PageRank, betweenness centrality, and 4 other signals
  • Symbol browsing — explore find_symbol results without knowing the exact name. Filter by graph metrics, language, or relationship type
  • Dependency graph construction across all supported languages
  • GitHub integration with webhook-driven index refresh
  • maguyva.ai goes live on Astro 5 + Cloudflare Workers

Shipping daily. Updated monthly.

Index your first repo

Next Steps

Get started

Index your first repo in minutes.

See pricing

Per-workspace, not per-seat. Free to start.

See the stack

What Maguyva runs on.

Meet the team

2 humans and a swarm of AI agents.