Beamscope

Compiler-accurate code intelligence for BEAM codebases (Erlang and Elixir).

Chunking and call-graph extraction are built on :epp and Code.string_to_quoted/2 — the same frontends the compiler itself uses — instead of a generic tree-sitter grammar. On Erlang specifically, this means macros are seen exactly as the compiler sees them: a call hidden behind a -defined macro resolves to its real target, not an opaque token.

Built to reduce the token cost of AI coding agents (Claude Code, and similar tools) working against large Erlang/Elixir codebases — in benchmarks across three real codebases, using beamscope instead of raw grep/read cuts token usage by 90–100% on call-graph queries. See ENGINEERING.md for the architecture decisions, benchmark methodology, and full results.

How it works

flowchart LR
    Client["AI agent / MCP client\n(Claude Code, etc.)"] -->|HTTP JSON-RPC| MCP["Beamscope.MCP\n(Plug + Bandit)"]
    MCP --> Repo["Beamscope.Repo"]

    Repo --> CGStore["Callgraph.Store"]
    Repo --> SearchStore["Search.Store"]

    CGStore --> CGPipeline["Callgraph.Pipeline\n(:epp / Code.string_to_quoted)"]
    SearchStore --> ChunkPipeline["Chunking.Pipeline"]
    SearchStore --> Embeddings["Embeddings\n(Bumblebee/Nx/Torchx, optional)"]

    CGPipeline --> Source[("your repo's\nsource files")]
    ChunkPipeline --> Source

    CGStore -. persists .-> CGFile[(".beamscope/callgraph.json")]
    SearchStore -. persists .-> SearchFile[(".beamscope/search.dets")]

Repo is the single entry point both the MCP server and direct callers go through. The call graph and search index are each built once per repo_path, cached in memory, and persisted to disk so a restart doesn't mean re-parsing or re-embedding the whole repo (see Limitations for what's not yet incremental about that). search_code's embedding step is the only part of this that touches the optional bumblebee/nx/torchx deps — everything else works with zero ML dependencies. See ENGINEERING.md for why each piece is built the way it is.

Status

Chunking and call-graph parity are validated against real production Erlang codebases. The MCP server (call graph + semantic search) works end-to-end and is verified as a real mix dependency in an external Elixir app. No incremental indexing yet — every index build reprocesses the whole repo from scratch (see Limitations). Published on Hex — see Setup.

Setup

Add to the consuming project's mix.exs:

def deps do
  [
    {:beamscope, "~> 0.1.0"}
  ]
end

mix deps.get

If the consuming project uses Igniter, mix igniter.install beamscope does the same mix.exs edit and prints a notice listing what's wired up right now.

Prefer to track the git repo directly (a specific commit, or main) instead of a Hex release:

{:beamscope, git: "https://github.com/mangalakader/beamscope.git"}
# or, against a local checkout:
# {:beamscope, path: "../beamlens_spike"}

mix igniter.install beamscope@git:https://github.com/mangalakader/beamscope.git

Ignore the build artifacts. Once you index a repo (see below), beamscope writes <repo_path>/.beamscope/ there — a rebuildable cache, like _build/, not something to commit or ship. Add to the consuming project's .gitignore (and .dockerignore, if you build container images):

.beamscope/

Optional — only needed for search_code (semantic search) — add the ML deps too:

{:bumblebee, "~> 0.7"},
{:nx, "~> 0.12"},
{:torchx, "~> 0.12"}

Building Torchx's NIF needs a C/C++ toolchain and cmake on the machine running mix deps.get/mix compile the first time — install that first if you don't already have it (e.g. brew install cmake on macOS). Everything after that is automatic — no Ollama, Qdrant, or Docker to run.

First run

Start the MCP server:

mix beamscope.mcp                # http://localhost:9877/mcp
mix beamscope.mcp --port 8080

Connect an MCP client to that URL as a remote HTTP server (not a spawned stdio subprocess). There's no separate "index this repo" step: every tool call takes an explicit repo_path, and the first call for a given path builds (and caches, and persists to <repo_path>/.beamscope/) whatever it needs on demand.

get_callers/get_callees/find_call_path build the call graph on first use — seconds for a small repo, longer for a real production-sized one (this is the one-time cost of walking and parsing every file).
search_code additionally chunks the repo and embeds every chunk. The first search_code call for a repo is the slowest call you'll make — it's also the point where the embedding model gets downloaded, if this is the first time it's run on this machine.

Every call after the first for a given repo_path is served from an in-memory cache, and the on-disk .beamscope/ files survive a server restart too — nothing needs to rebuild just because the server restarted.

sequenceDiagram
    participant C as Caller
    participant R as Repo
    participant Disk as .beamscope/ on disk

    C->>R: first call (repo_path)
    Note over R: no in-memory cache, no disk file yet
    R->>R: build (parse / chunk / embed)
    R->>Disk: persist (atomic write)
    R-->>C: result

    C->>R: next call (same repo_path)
    Note over R: in-memory cache hit
    R-->>C: result (fast)

    Note over R: server restarts
    C->>R: call after restart
    R->>Disk: load persisted file
    Note over R: cached in memory again
    R-->>C: result (no rebuild)

Without the MCP server, the same operations are available directly:

alias Beamscope.Repo

{:ok, %{callers: callers}} = Repo.callers("/path/to/repo", "my_module", "my_function")

{:ok, %{exact_matches: exact, semantic_matches: semantic}} =
  Repo.search("/path/to/repo", "where session tokens get validated", limit: 5)

Usage

Call graph — who calls what, and how to get from A to B:

alias Beamscope.Repo

{:ok, %{callers: callers}} = Repo.callers("/path/to/repo", "my_module", "my_function")
{:ok, %{callees: callees}} = Repo.callees("/path/to/repo", "my_module", "my_function")
{:ok, %{path: path}} = Repo.call_path("/path/to/repo", "mod_a", "foo", "mod_b", "bar")

Each caller/callee comes back enriched with its definition's file_path/start_line/end_line, so acting on a result doesn't require re-reading the whole file it lives in.

Semantic search — chunk-level embeddings, searchable by natural-language query, entirely in-process (no external service, no Ollama/Qdrant/Docker):

Repo.search("/path/to/repo", "where is the session token validated", limit: 5)
# {:ok, %{
#   exact_matches: [%{file_path:, line:, text:}, ...],
#   semantic_matches: [%{file_path:, symbol:, start_line:, end_line:, kind:, score:}, ...],
#   semantic_error: nil
# }}

exact_matches and semantic_matches are two separate lists, not one blended ranking. exact_matches is a literal, in-process grep for identifier-like terms in the query and needs no ML deps; use it (or get_callers/get_callees once you have a name) for exact-name lookups. Lean on semantic_matches when you don't know what to grep for. If the optional ML deps aren't installed, semantic_matches comes back empty with semantic_error: :embeddings_not_available rather than the whole call failing.

Chunking — the lower-level building block, if you need function/attribute-level chunks directly:

alias Beamscope.Chunking.Pipeline

result = Pipeline.chunk_repo("/path/to/repo", max_concurrency: 8)
result.chunks   # [%{symbol:, start_line:, end_line:, text:, kind:, file_path:, warning:}, ...]
result.errors   # [{path, reason}, ...] — timeouts/crashes, doesn't fail the whole run

Supports .erl/.hrl (via :epp), .ex/.exs (via Code.string_to_quoted/2), and falls back to line-window chunking for everything else (docs, config files) or files that fail to parse.

Benchmarking your own repo

mix beamscope.benchmark --repo /path/to/repo [--repo /path/to/repo2] [--output docs/benchmarks/]

Auto-discovers representative tasks in the repo, measures real token counts and latency for beamscope vs. a grep/read baseline, and writes a timestamped Markdown report. See ENGINEERING.md for the methodology and results this same tool produced against MongooseIM, amoc-arsenal-xmpp, and the Elixir language's own source.

The token-count/reduction table works with no extra setup. The latency comparison table needs Benchee; add {:benchee, "~> 1.3", only: :dev} to your own mix.exs deps to get it — without it, the benchmark still runs and reports token counts, just without that section.

Roadmap

Planned, not yet built:

Incremental indexing — skip re-processing files that haven't changed since the last build, instead of always rebuilding from scratch (see Limitations).
Concurrent/pipelined embedding — parallelize search_code's chunk-embedding step to speed up cold indexing of large repos.
Export/import of a built index — move a repo's .beamscope/ index between machines without rebuilding.
A web-based call-graph visualizer — browse get_callers/ get_callees/find_call_path results interactively instead of only through MCP tool calls or direct API calls.
Embedding/search-index quality metrics — surface how well search_code's semantic matches are actually performing, beyond the benchmark tool's task-by-task quality notes.

None of these block current usage — each is additive to what's already working.

Limitations

No incremental indexing. Every index build — call graph or search — reprocesses every file in the repo from scratch; nothing is tracked about what changed since the last build. Repo.reindex/2 means "discard the cache and rebuild everything," not "update only what changed." For a small repo that's fast enough not to matter; for a large production codebase it's the same one-time cost as the very first build, paid again on every reindex. The mental model: the index is a rebuildable cache, not a live-updating one.

Development

mix deps.get
mix test

Tests tagged :external (Beamscope.Embeddings/Beamscope.Search.Store real embedding tests) are excluded by default since they hit a real model download + CPU inference — run mix test --include external to include them.

Real-world parity fixtures (full MongooseIM and amoc-arsenal-xmpp checkouts) live under priv/fixtures/ but are gitignored — they're research artifacts for validating parity against a reference pipeline, not package fixtures. The small synthetic .erl/.ex files alongside them are real, tracked test fixtures.

License

MIT — see LICENSE.

Next Page → Engineering notes