Skip to the content.

fleetlens MCP tools

Generated from the server’s own tool registrations by python -m fleetlens.server.toolsdoc. Do not edit by hand: a test regenerates this page and fails when it differs, so a signature or docstring change that is not reflected here turns CI red.

15 tools. Descriptions are verbatim from the docstrings an agent reads over MCP, so what you see here is what it sees.

Availability

Tools are registered only when they can answer. A tool that is absent is not an error: an advertised tool that always fails costs an agent a turn and teaches it to stop trusting the server. A default fl serve against an index with no guidance and no embeddings advertises 10 of these 15.

Tool Registered when
discover_interfaces an embedding model is configured (--embed-model)
discover_libraries an embedding model is configured (--embed-model)
discover_services an embedding model is configured (--embed-model)
find_symbol always
get_callees always
get_callers always
get_endpoint_call_graph always
get_engineering_guidance the index contains guidance objects
get_index_info always
get_service_graph always
get_service_relationships always
get_source --http and FLEETLENS_EXPOSE_SOURCE is set
get_symbol always
list_interfaces always
list_services always

discover_interfaces

Registered when an embedding model is configured (--embed-model).

Parameter Type Default
query string required
limit integer 5

Find endpoints by meaning, e.g. “the endpoint that sends order-confirmation emails”. Ranked semantic search over LLM-generated interface summaries (needs the enrichment tier). Returns interface ids + summaries + scores. Use get_endpoint_call_graph on a result to trace it.


discover_libraries

Registered when an embedding model is configured (--embed-model).

Parameter Type Default
query string required
limit integer 5

Find shared code libraries by what they provide, in natural language, e.g. “retry with exponential backoff” or “parsing feature-flag config”. Use before writing a helper that may already exist. Ranked semantic search over LLM-generated library summaries (needs the enrichment tier). Returns library ids + summaries + scores; use find_symbol to reach the code.


discover_services

Registered when an embedding model is configured (--embed-model).

Parameter Type Default
query string required
limit integer 5

Find services by what they do, in natural language. Ranked semantic search over LLM-generated service summaries (needs the enrichment tier). Returns service ids + summaries + scores.


find_symbol

Parameter Type Default
query string required
limit integer 20

Find code_symbol ids by a substring of their id — a method name, class, or file fragment (e.g. “OrderManager.get” or “managers/db.py”). The entry point to the other call-graph tools when you know a name but not the full id.


get_callees

Parameter Type Default
symbol_id string required
max_depth integer 2

What a function/method calls, transitively — its downstream call tree. symbol_id is a fully-qualified code_symbol id (from find_symbol or an endpoint trace). Only in-repo calls are edges; stdlib/third-party calls are intentionally omitted.


get_callers

Parameter Type Default
symbol_id string required
max_depth integer 2

Who calls a function/method, transitively — its upstream callers (the blast radius of changing it). symbol_id is a fully-qualified code_symbol id.


get_endpoint_call_graph

Parameter Type Default
interface_id string required
max_depth integer 3

Trace an endpoint into the code: from an interface to its handler and everything that handler transitively calls. Use for impact analysis (“what does changing this endpoint touch?”). interface_id is a fully-qualified interface id (e.g. “interface:orders:post-orders”). max_depth bounds the walk (default 3, capped 6). Returns resolved handlers, reached nodes (with path + line), and calls edges.


get_engineering_guidance

Registered when the index contains guidance objects.

Parameter Type Default
topic string ''
service string ''
detail string 'auto'

Organisation engineering standards: which libraries, frameworks and patterns to use when writing code in this fleet. Human-authored, not derived from source.

Call this BEFORE writing or reviewing code in any of these repositories. Conventions inferred by reading the fleet are the existing conventions, which during a migration are the ones being moved away from.

service returns the rules governing that service in full, matched on what it declares. topic searches by meaning and returns the nearest few rules as titles, excerpts and scores, so ask again with topic=<id> for the text of one. Neither lists everything. detail is “auto”, “full” or “titles” to override that.

A topic search always returns its nearest matches, which are not necessarily relevant ones. Read the titles and decide: if none of them covers the question, the rulebook does not cover it, and you should say so rather than apply the closest one.


get_index_info

Takes no parameters.

When this index was built, and how much it covers.

Worth checking before you trust an answer on a shared server: the index is a snapshot of the repositories at the moment it was built, so anything merged since is invisible to every other tool here. Also tells you how many services are covered, which is how you find out a repository was never indexed at all.

build_id identifies this exact build. A harness running an evaluation should record it with each result, so the answer can be attributed to the index that produced it. repositories maps each indexed repository to the git commit it was built from; it is here and not on /healthz because it can run to hundreds of entries.


get_service_graph

Parameter Type Default
min_confidence string ''

THE WHOLE microservice dependency graph in one call — every service and every service-to-service edge across all repositories.

Use this FIRST for any fleet-wide or cross-repository question: which service is depended on by the most others, what depends on X, what would break if X changed, which services are unused or isolated, how many services exist, which talk over HTTP vs queues. Answers these without reading any source files.

Each service carries depends_on_count / depended_on_by_count, so ranking and fan-in/fan-out questions are answered directly from this one response. Each edge carries relationship (calls | publishes_to | shares_channel), a confidence, and the evidence that produced it (request paths, queue names, config keys).

min_confidence optionally filters weaker edges: “high” or “corroborated”.

Check likely_duplicate_deployments before ranking: one repository checked out several times appears as several services, which inflates raw counts.


get_service_relationships

Parameter Type Default
service_id string required

Dependencies of ONE service: downstream (services it calls or publishes to) and upstream (services that call it). Use for a single service; for questions spanning the whole fleet use get_service_graph instead, which returns everything in one call.

service_id is fully-qualified (e.g. “service:orders”). Edges are derived deterministically from source and config, and each carries a confidence and the evidence that produced it.


get_source

Registered when --http and FLEETLENS_EXPOSE_SOURCE is set.

Parameter Type Default
symbol_id string required
context_lines integer 0
max_lines integer 200

The actual source of an indexed symbol, read from the repository.

Use when the structure is not enough and the question is about what the code does: which validations run and in what order, which branch returns an error, what a handler calls before it writes. Every other tool here answers from the index and gives you a file:line; this returns the lines themselves.

Takes a code_symbol:... id from find_symbol, get_callers, get_callees or get_endpoint_call_graph, so the usual path is to locate the symbol first and read it second.

Scoped to the symbol’s own span. context_lines widens it by a few lines either side, for a decorator or a signature that sits just outside, and is capped at 50. At most 500 lines come back in one response, and truncated says when that bit. There is no whole-file mode: read the symbols you need rather than the files they are in.

The source is checked against a hash of the file the index actually parsed. It reads the commit the index was built from first, falls back to the working copy, and returns whichever matches, saying which in read_from; a pinned index reads at the pin’s commit. verified is false only for indexes built before those hashes existed. When nothing matches you get span_mismatch and no source, which means the index is stale for that file and needs rebuilding: the lines would otherwise land on a neighbouring function and read perfectly.


get_symbol

Parameter Type Default
symbol_id string required

Look up one code symbol: location (path + line) and how many direct callers/callees it has. Returns a not_found response (with did_you_mean) if unknown.


list_interfaces

Parameter Type Default
service_id string required

Every interface a service exposes: id, method, path, handler, and provenance (source = static from the parsers, or llm from the grounded gap-filler). The entry point to get_endpoint_call_graph when you know the service but not the endpoint. service_id is fully-qualified (e.g. “service:orders”).


list_services

Takes no parameters.

List every indexed microservice in this codebase, with interface and symbol counts. Use to discover what services exist before drilling in. For dependencies between them, prefer get_service_graph, which returns the whole mesh at once.