ProtoMap makes Claude Code 6.4× faster and 5.3× cheaper across 9 gRPC repositories with 5 million lines of code

Opus 5.5 · high effort: time to answer

Median wall-clock time of 2 runs per prompt (their mean), log scale. The figure at the right of each row is how many times faster the answer came with ProtoMap.

Opus 5.5 · high effort: cost per answer

Median cost of 2 runs per prompt (their mean) at API prices, as Claude Code reports it, log scale.

Sonnet 5.5 · xhigh effort: time to answer

Median wall-clock time of 2 runs per prompt (their mean), log scale. The figure at the right of each row is how many times faster the answer came with ProtoMap.

Sonnet 5.5 · xhigh effort: cost per answer

Median cost of 2 runs per prompt (their mean) at API prices, as Claude Code reports it, log scale.

1. Quickstart Q01 List all the gRPC services implemented in the Swift repo (protomap-test-swift). Give the fully qualified name of each and how many RPCs it has.
2. Quickstart Q02 Show the 10 most connected gRPC services across all our repos: the ones with the most distinct other services calling them or being called by them. Give the inbound and outbound service counts for each.
3. Quickstart Q03 Which services depend on acme.loyalty.v1.LoyaltyMemberQueryService? For each dependent service, tell me which repo it lives in and which RPCs it calls.
4. Quickstart Q04 Which services call acme.customers.v1.CustomerService/ListCustomers? List every call site with file:line and the enclosing function, across all repos.
5. Quickstart Q05 What is the blast radius of changing acme.ledger.v1.LedgerAccount.currency_code? Which services actually use this field's data (read it and pass it along to another RPC or into a response), and through which RPCs?
6. Cross-service Q06 We're replacing the `customer` field of acme.customers.v1.UpdateCustomerRequest with a new message. Which code, in which repos, sets that field when calling acme.customers.v1.CustomerService/UpdateCustomer, and to which other RPCs does the UpdateCustomer handler pass the field on?
7. Cross-service Q07 Where does the `cart` in the response of acme.cart.v1.CartCalculationService/SimulateCart come from? Trace it back to the upstream RPC responses or request fields it is built from.
8. Cross-service Q08 Trace the `payout` field of acme.payments.v1.SchedulePayoutExpiryRequest two hops downstream: which RPCs does the SchedulePayoutExpiry handler send it to, and where do the handlers of those RPCs send it next? Name the services and repos at each hop.
9. Cross-service Q09 We want to remove the RPC acme.documents.v1.DocumentQueryService/ResolveDocuments. Find every place in every repo that calls it, including indirect calls through helpers, method names built from strings, registries, or reflection, and give file:line and the calling service for each so we can migrate them.
10. Cross-service Q10 What does acme.payments.v1.PaymentCalculationService depend on? List every downstream service it calls, which RPCs, and which repo owns each downstream service.
11. Cross-service Q11 If acme.pricing.v1.PriceListQueryService goes down, which services are affected? Include direct callers and services that depend on it transitively through the RPCs that call it, up to 3 hops away, and say how many hops away each one is.
12. Cross-service Q12 Can a request to acme.ticketing.v1.TicketTransferWorkflowService end up calling acme.ledger.v1.LedgerAccountValidationService? Show the shortest call chain hop by hop: the RPC called at each hop, and the repo and file:line where each call is made.
13. Cross-service Q13 Which code outside of gRPC handlers (event workers, batch jobs, CLIs, server startup) calls the ledger services (acme.ledger.v1.*)? List each caller, its repo, and the ledger RPCs it calls.
14. Cross-service Q14 Write API reference docs for acme.giftcards.v1.GiftCardService: for each RPC, its request and response messages, the file that implements it, and which services call it.
15. Cross-service Q15 Which RPCs of acme.checkout.v1.CheckoutSessionService are never called by any code in any of our repos? We want to delete dead endpoints, so don't list an RPC that is actually called somewhere. End your answer with one line of the form `UNUSED: <comma-separated RPC names>`.
16. Cross-service Q16 For a privacy review: which services' code actually reads a patient's national ID (acme.patients.v1.Patient.national_id) and passes it on, and through which RPCs does it travel?
17. Cross-service Q17 If acme.customers.v1.CustomerCalculationService/ExplainCustomerCalculation stopped returning the `result` field of its response, which callers would be affected? For each caller, which RPC requests or responses does that data flow into?
18. Cross-service Q18 What does the handler of acme.shipping.v1.ShipmentCalculationService/CalculateShipment call downstream, and which fields of its request end up in each downstream call?
19. Cross-service Q19 Which other domains does the checkout domain (acme.checkout.v1) call into, and which domains call checkout? Summarize the dependencies in both directions.
20. Cross-service Q20 We're making a breaking change to the acme.subscriptions.v1 protos. Which repos call or implement subscriptions services and would need to update and redeploy? For each repo, which subscriptions services does it use?

4.99M tracked lines: 1.64M of service code, 1.23M of protobuf (each repository holds copies of the files of the services it calls), and 2.12M of generated Go bindings, which the Go repository checks in. Calls and flows are the oracles' counts of call sites and field-level dataflows the code makes.

• Each run is a fresh claude -p session (Claude Code 2.1.306) in ~/protomap-test, which holds only the nine repositories, with nothing the agents could take answers from: the fixtures' answer keys, build outputs, and other projects were moved out for the runs.
• Both arms have the same read-only tools: Read, Grep, Glob, subagents, and read-only shell commands (rg, grep, find, git grep, awk, jq, and the like); anything else is denied. The ProtoMap arm adds the ProtoMap MCP server (Streamable HTTP, the shared graph of the nine repositories); the baseline arm has no MCP servers.
• User settings, CLAUDE.md files, and hooks were left out (--setting-sources project,local) so nothing but the prompt differs between runs.
• Opus 5.5 ran at high effort and Sonnet 5.5 at extra-high effort. Every prompt ran twice per model and arm (160 runs; a third round was cut short by usage limits and discarded); the charts show the median, which for two runs is their mean, and accuracy is the mean.
• Time is wall-clock time of the whole session; cost and tokens are what Claude Code reports at API prices, subagents and cache reads included. A run is stopped after 45 minutes and counts its 45 minutes and the cost it had reached.
• Grading is by code, with no model as judge: each answer must name the services, methods, repositories, or fields of its prompt's answer key (by full or short name, or the file its implementation is in). The dead-RPC prompt scores the F1 of the list the answer ends with, so a wrong RPC costs as much as a missed one.