Open source · Local-first · MCP-native

Give coding agents the right context, not more context.

FireLens is the open-source local code search engine for AI coding agents — lexical, semantic, and graph retrieval fused with RRF, native MCP tools, and Mojo-accelerated hot paths. Your code never leaves your machine.

Explore FireLens How it works

firelens / search_code

localhost

⌕where are indexes persisted?↵

Hybrid RRF · lexical + semantic Auto-selected

⌑

default_database_path

app/storage/database.py:46–82

bounded

BM25 lexical hits and semantic hits fused by reciprocal-rank fusion.

rrf k=60 · rank 1 of 5

Local-firstYour code never leaves your machine.

Hybrid retrievalLexical + semantic, fused with RRF.

Graph searchQuery the repo dependency graph.

Mojo accelerationMeasured speed where it wins.

From repository to useful context.

FireLens turns a checked-out codebase into a local search engine your agent can query — lexical, fuzzy, semantic, hybrid, and graph — without moving source code through another service.

01 · INPUT

Repository

⌄ app/ ⌄ indexing/ embedder.py storage/database.py

Python-first parser · root .gitignore

02 · PERSIST

SQLite index

symbolsfts5 indexchunks + vectorsgraph edges

firelens.dbAtomic replacement · shared read lock

03 · RETRIEVE

Hybrid route

≡ Lexical SQLite FTS5 · BM25≈ Fuzzy Mojo / Python✦ Semantic 768-d vectors⊕ Hybrid RRF fusion⬡ Graph dependency edges

Auto picks per query · Mojo only after the local threshold is met

04 · DELIVER

Agent context

mcp / search_codetop_k: 5snippet: 4,000 chars max

CLI · Streamlit · local STDIO MCP

Acceleration where it earns its place.

These numbers show how much faster is the Mojo acceleration than Python for the same query, on the same machine. Timings measure complete calls, including Python-to-native buffer preparation.

Read the benchmark method

Fuzzy search · Mojo vs Python · 512 candidates4.04× faster

The same fuzzy query drops from an 11.257 ms Python median to 2.788 ms with Mojo.

Python p95 11.388 ms · Mojo p95 2.850 ms · 30 iterations

Semantic search · Mojo vs Python · 50k × 7681.19× faster

Ranking 50,000 vectors takes 6.490 ms with Mojo against 7.696 ms in Python and NumPy.

Python p95 8.039 ms · Mojo p95 7.456 ms · 30 iterations

Auto routes Mojo fuzzy at ≥4 candidates and semantic ranking at ≥30,000 rows.

Apple M4 Max · 16 logical CPUs · 30 measured iterations

The local engine in a hosted world.

Sourcebot and Sourcegraph made code search a product you host and browse. FireLens gives your agents deeper retrieval, zero infrastructure, and code that never leaves the machine it lives on.

Capability	FireLens	Sourcebot	Sourcegraph
Runs where your code lives	Local CLI and MCP — no server, no Docker	Self-hosted web server	Cloud or enterprise install
Lexical search	SQLite FTS5 · BM25 ranking	Regex and symbols	Regex and symbols
Semantic search	Local 768-d vector embeddings	—	—
Hybrid ranking	RRF fusion of lexical + semantic	—	—
Graph search	Dependency-graph traversal	—	—
Accelerated hot paths	Experimental Mojo · up to 4.04×	—	—
Built for coding agents	MCP-native from day one	Browser-first experience	Browser-first experience
Where your code lives	Never leaves your disk	Indexed on a server	Indexed on a server

Based on publicly documented features. Sourcebot and Sourcegraph are trademarks of their respective owners.

Less setup. More signal.

Install FireLens, index a local repository, and ask for the code your agent needs next.

Read the docs

terminal

# install
$ git clone https://github.com/aslanbayli/firelens.git
$ cd firelens && uv sync

# retrieve
$ uv run firelens index /path/to/repo
$ uv run firelens search /path/to/repo "where are indexes persisted?" --mode auto

Designed for complex codebases.

Does FireLens send my source code anywhere?+

No. FireLens indexes local repository paths and persists indexes in SQLite on your disk. Its retrieval surfaces are a local CLI, Streamlit, and STDIO MCP tools — there is no cloud component.

What makes the search hybrid?+

Lexical hits come from SQLite FTS5 with BM25 ranking; semantic hits come from local vector embeddings. Reciprocal-rank fusion (RRF) merges the two lists, so exact keywords and natural-language intent reinforce each other instead of competing.

What is graph search?+

FireLens represents your repository as a dependency graph of imports, symbols, and call relationships, then searches over it. That answers structural questions keyword search can't — like who calls a symbol or what a module depends on.

Which repositories are supported today?+

The current parser is Python-first. The architecture keeps the parser boundary explicit so additional languages can be added without changing retrieval contracts.

Is Mojo required?+

No. Python and NumPy provide the reference path. Mojo is optional and used only for fuzzy and semantic workloads where the benchmark supports it.

How do coding agents connect?+

FireLens exposes local STDIO MCP tools for index status, indexing, and bounded code search.

Your context window is valuable.

Give your agent better context.

Explore FireLens