Claude Code on steroids. Codex on steroids. Any AI coding agent — on steroids.

yoyo is an MCP server that gives your agent 30 AST-grounded tools to read, understand, and edit code. No hallucinated file paths. No guessing. Facts from the source.

99% eval accuracy across 4 languages, 8 real codebases — vs 26% baseline (Claude Code alone).

Why

Your AI agent reads code like a human with no IDE: grep, cat, hope. It hallucinates function names. It misses callers. It patches the wrong file.

yoyo gives it what it was missing: a structured interface to the codebase. The agent calls `symbol` instead of `cat`. It calls `blast_radius` before deleting. It calls `flow` to trace a request end to end. It patches by function name, not line number.

The eval gap is the proof: 99% vs 26%. Same model. Same tasks. Different tools.

Language focus

> Rust · Go · Zig · TypeScript — four languages, done deep.

| Language | bake | symbol | trace_down | endpoints | write tools |

|---|---|---|---|---|---|

| Rust | ✅ | ✅ | ✅ | ✅ actix/rocket | ✅ |

| Go | ✅ | ✅ | ✅ | ✅ gin/echo/net-http | ✅ |

| Zig | ✅ | ✅ | — | — | ✅ |

| TypeScript | ✅ | ✅ | — | ✅ express | ✅ |

Not every language. The four where systems-level code intelligence matters most.

---

## The combinations are the point

One trick is fine. Fifty moves chained is transcendent.

| Combination | What it does |

|---|---|

| `supersearch` → `symbol` → `patch` | find it, read it, change it ⚡ |

| `blast_radius` → `health` → `graph_delete` | who calls this? is it dead? remove it safely |

| `flow` → `multi_patch` | trace the full request path, fix it end-to-end in one shot |

| `bake` → `semantic_search` → `suggest_placement` | where does this new function belong? |

| `architecture_map` → `all_endpoints` → `graph_create` | understand the shape, find the gap, fill it |

No single tool is the point. The orchestration is.