Carrick

Check a consumer service's calls against a producer's operations: missing or unused operations, plus stored type verdicts (coverage in type_verdicts). Read status first. It reads compatible, incompatible, partially_checked, or unresolved where nothing was compared, and pairs_compared and pairs_uncompared say how much of the pair the check reached. Call it before removing or changing any route, shape, field, or event. Consumers that reach the producer through a published SDK package have no call site of their own and are listed separately in via_sdk, never merged into the direct pairs. For whole-project topology, use get_service_graph first. To read both sides' types for the operations a verdict names, use get_contract_pair. Reference: https://docs.carrick.tools/mcp-tools

Find functions that already do what you are about to write, and places this project has built the same thing twice. Pass functions with one entry per new or edited function (name and file when it is already indexed, one plain sentence when it is not) and each entry comes back with the closest functions in other files. Pass audit: true instead, with no functions, for a whole-project audit that groups the functions describing the same behaviour. A call with neither, or with both, is refused. Matching reads how the index describes a function rather than its body, and every row carries the file and line span of the function it names. For one concept rather than a set of functions, use search_by_intent.

List the operations one service exposes: HTTP endpoints, GraphQL fields, socket events. Call it when unsure which operations a service has or whether one exists. For the types of a known endpoint, use get_endpoint_types. When you already know the operation and need its producer or its consumers, use get_operation.

List recorded direct callers of a named function, with their service, file and line span. Zero recorded callers is not safe-deletion evidence; callback and value references are unmeasured. Call it before changing or removing a function to see what a signature change would break; pass depth for transitive callers. Rows name the enclosing caller, not the call-site line. For what a function calls rather than who calls it, use list_function_intents.

Put one consumer and one producer side by side, operation by operation: the producer's request and response types, every consumer call site with the type that call site expects, and the type verdict the scan stored, with its scanner version. Call it when you are about to change a shape and need to know what the other side already believes it is, or when a compatibility verdict names a problem you cannot place. It reports what the index holds and adds no verdict of its own. It shows HTTP operations side by side, and only the operations a consumer call wins onto; for GraphQL, socket and pub/sub pairs, check_compatibility carries the stored verdicts. For the producer's types alone, use get_endpoint_types; for which calls match at all and which do not, use check_compatibility. Reference: https://docs.carrick.tools/mcp-tools

Read the extracted TypeScript request/response types for one endpoint instead of guessing from sample JSON. Call it before building a request body or parsing a response. For nested types, follow with get_type_definition. To read these types against the type each consumer call site expects, use get_contract_pair. Reference: https://docs.carrick.tools/mcp-tools

Read the findings list on main: type mismatches between endpoints and their callers, and groups of functions that look like copies of each other, built after each scan of main. Each row carries every place as file and line, what to check, and the call that drills into it. get_project_map shows the first rows under FOUND ON MAIN; this tool lists them all, paged.

Start here when you know an operation but not the service: given a method and a path, it returns every service that serves it, every call site that reaches it, and the recorded calls that resolved to no producer, with a file and line on each. Use it before changing or removing a route to see the blast radius, and to answer "who serves this" without guessing a service first. Every response also names what the index holds one method or one path segment away, so a call recorded under a neighbouring method or path version cannot read as a confident zero. Bounded to one operation; for the whole topology use get_service_graph, for request/response types get_endpoint_types. Reference: https://docs.carrick.tools/mcp-tools

Call this first in a session: a compact plain-text map of the project, naming the sibling services this project indexes, the contracts this repo depends on, who consumes this repo, and the unmatched calls with how many of them match no route, with the drill-down tool at the end of every section. Cheap; pass service to split THIS REPO from its siblings, detail for more rows per section.

Read npm dependencies: one service's full package list, or org-wide conflicts when called without a service. Conflicts are graded on the semver ranges written in package.json, not on version strings: a package is listed only when no single install satisfies every service's range; error means the ranges span different majors, warning means they miss within one major. Each conflict carries a one-sentence summary naming the groups. Call it before adding or bumping a package.

Return the whole cross-service call graph in one call: consumer -> producer edges across every protocol, unmatched calls, orphaned endpoints. Use it first for topology: who depends on a service, what a change ripples into. An edge carrying via_sdk reaches its producer through a published package rather than a call site of its own. An edge carrying same_service has one service at both ends, a service calling its own operation, and is listed after the cross-service ones. For one pair in depth, use check_compatibility. For the types on both sides of one pair's operations, use get_contract_pair. On a large project, call get_project_map first: it summarises this graph section by section and names the drill-down tool for each.

Get the fully resolved TypeScript definition of a named type, with transitive dependencies. Call it after get_endpoint_types when a type references nested types you need expanded. Accepts the Endpoint_ alias or a real symbol name.

Enumerate this project's recorded outbound-call candidates: SDK calls into declared dependencies, calls to declared external domains, and env-var URL bases, each with its service, file and line. Use it to inventory or audit what a service talks to outside itself; for an audit, start with view "targets" to see every distinct target with counts, then fetch call sites per target. Rows are candidates, not verified egress.

List indexed functions with their one-line intents, paginated; filter by service or substring. The index covers every function with an intent, not just exported ones. Use it to browse what a service contains or compare what sibling repos have built. Pass file to answer "what can I call in this file" without reading it, with compact for signature-only rows. For concept queries ('dedupe users by email'), use search_by_intent instead.

List this workspace's Carrick projects and their connected repos. Call it when unsure which project or repo value to pass. Metadata only, no scan data.

List every service in the project with endpoint and call counts. Call it when unsure what exists or which repo owns a name. For the call graph, use get_service_graph.

Search every function in this project by what it does, including sibling repos you do not have checked out. Pass a plain-English concept ('verify a webhook signature'); matches need no keyword overlap. Use it before writing any helper, parser, validator or domain function, even when you are sure it is new — grep can prove a name is absent but never that the behaviour is. A row whose only tie to the query is a word most functions share is counted in the response rather than ranked. To browse instead, use list_function_intents. For a set of functions you are about to write or have just written, and for a whole-project reuse audit, use find_similar.