ProtoMap

ProtoMap from HoundDog.ai is a gRPC / Protobuf context engine and model harness: it scans source into an org-wide graph, then serves engineers and coding agents over the ProtoMap MCP for call sites, blast radius, field lineage and more. On their synthetic ~5M LOC benchmark they report about 13× fewer tokens, 7.7× faster answers, and 6.4× lower cost versus a bare agent (see their table; numbers move with models and repos).

Flow: protomap scan → graph (repos, services, methods, fields, messages) → optional BYO-model false-positive prune → MCP query. Languages: Go, Java, Kotlin, C#, Python, TypeScript, JavaScript, Rust, Swift; mono- and multi-repo. Self-hosted via Docker + PostgreSQL (trial installer). They state SOC 2, regular pentests, and membership in Anthropic Defense Access and OpenAI Daybreak Blue. Roadmap mentions REST/OpenAPI, GraphQL and Thrift.

Highlights:

• Product https://protomap.ai/ , quickstart https://docs.protomap.ai/quickstart/ ;
• Installer repo https://github.com/hounddogai/protomap ;
• Works with Claude Code, Codex, Antigravity, VS Code, OpenCode and other MCP clients;
• Analysis aimed at a modern laptop — no GPU required per their copy;
• Best for large gRPC microservice codebases, not a universal code-search replacement.

Why it matters

Cross-repo gRPC blast-radius work burns tokens and still misses indirect callers. ProtoMap builds a deterministic graph from code so agents stop rediscovering the same edges, and “who calls this RPC?” becomes one tool call. For platform teams already paying agent bills, that is a direct cost and accuracy lever; the Product Hunt launch makes trial easy.

Limits

Today is gRPC/Protobuf-first; other protocols are forthcoming. Benchmarks are synthetic and model-specific. Self-host needs Docker Compose plus RAM/disk; repos without gRPC see little gain. Some review steps expect a BYO model.

1. Read https://docs.protomap.ai/quickstart/ ; confirm git, recent Docker Compose, RAM and disk;
2. Clone https://github.com/hounddogai/protomap and run ./self-hosted/scripts/install.sh (Trial can use local PostgreSQL);
3. protomap scan . (or a public microservices-demo), then wire protomap mcp serve into Claude Code / Codex / etc.;
4. Ask MCP tools for call sites, blast radius, dead methods; watch the roadmap for more protocols.