reflens

Give your AI coding agent the full, lossless context of any reference repository — and let it reason over a codebase far larger than its context window.

Point it at a local folder, a GitHub URL, or a Repomix dump. reflens indexes it once and serves it to OpenCode / Claude Code (or any MCP host) as tools your agent calls on its own.

The problem

You want your coding agent to learn from a flagship repo and apply its patterns to your project. But the repo is 100k–1M+ tokens — it does not fit in the context window. Today you either:

• paste fragments and hope they're the right ones, or

• clone it into your workspace and let the agent blind-grep it every session (slow, token-hungry, no orientation), or

• dump it with Repomix and watch it overflow the window.

All three silently lose context. Silent truncation is the bug — the agent confidently reasons about code it never actually saw.

Related MCP server: mcplens

The approach: two tiers (the honest part)

You cannot fit a 25M-token repo into a 200K window losslessly — that's physics, not engineering. Any tool that claims otherwise is truncating behind your back. reflens refuses to, and gives you two tiers plus a way to prove nothing was lost:

The agent reasons from Tier 1 and expands into Tier 2 (exact source) only when a task needs the gnarly detail. Retrieval is the safety net, not the primary mechanism. Where the digest hits a budget, it prints the exact tool call to reach the rest — so nothing becomes unreachable.

→ Full design in ARCHITECTURE.md.

Quickstart (60 seconds)

# 1. Install (isolated; pipx recommended)
pipx install "git+https://github.com/cybertronayush/reflens"

# 2. Wire it into your agent (edits OpenCode + Claude Code configs, adds usage guidance)
reflens install both

# 3. Stock the library — any local dir, GitHub URL, or repomix .md
reflens add https://github.com/tiangolo/fastapi --name fastapi
reflens add /path/to/your/reference-repo --name myref
reflens add ./repomix-output.md --name dump

# 4. Prove nothing was lost (directory ingests are byte-exact)
reflens verify myref

# 5. Restart OpenCode / Claude Code, then just ask:
#    "Use reflens to learn fastapi's dependency-injection pattern and apply it to my app."

That's it. The agent calls the tools automatically — no slash-commands, no per-repo setup.

How your agent uses it

Once installed, every model in every project gets the reflens_* tools (global MCP server) plus a usage note in your global AGENTS.md / CLAUDE.md. A typical agent flow:

reflens_modules(repo)              → the module map (table of contents)
  → reflens_map(repo)              → architecture brief (hubs, conventions, decisions)
  → reflens_map(repo, path_glob)   → zoom into a module at signature detail
  → reflens_search(repo, query)    → find the relevant code (hybrid lexical+semantic)
  → reflens_read(repo, target)     → byte-exact source of a file or symbol
  → reflens_neighbors / _history   → dependencies / git history

You never write commands for the agent. For 100% reliability on a given task, just name the repo ("learn from the fastapi repo…").

MCP tools

CLI reference

reflens add <source> --name <n> [--semantic] [--max-file-bytes N] [--include-binary]
reflens list
reflens modules <name>                          # table of contents
reflens map <name> [--level 0|1|2] [--glob 'src/**'] [--budget N]
reflens search <name> "<query>" [-k N] [--mode auto|lexical|semantic|hybrid]
reflens read <name> <path|symbol> [--start N --end N]
reflens neighbors <name> <path|symbol>
reflens history <name> [path]
reflens verify <name>                            # SHA-256 round-trip + completeness + coverage
reflens enrich <name> [--model ...]              # optional LLM per-module summaries
reflens remove <name> -y
reflens install [opencode|claude|both]           # wire the MCP server + agent guidance
reflens serve                                    # the MCP stdio server (hosts launch this)

What gets extracted

• Python → exact, via the stdlib ast (classes, methods, functions, constants, module docstrings, imports).

• TypeScript / JavaScript / Go / Rust / Java / Kotlin / C / C++ / C# / Ruby / PHP / Swift / Scala / Shell → signature outlines via a tuned regex extractor (optionally tree-sitter with pip install 'reflens[code]').

• Markdown → heading outline (docs/specs/ADRs become navigable).

• Everything else → stored losslessly + full-text searchable.

Losslessness — and proof

$ reflens verify myref
{ "ok": true,
  "files": 1750, "verified": 1750, "failed": [],
  "completeness": { "declared_files": 1750, "indexed_files": 1750, "drift_detected": false },
  "extraction": { "code_files": 1287, "with_symbols": 1235, "coverage_pct": 96.0 } }

Every stored file is content-addressed and re-hashed on read; verify reconstructs all of them and compares SHA-256. Directory/git ingests are byte-identical to source. Repomix --compress dumps are lossless with respect to the dump (their bodies were already stripped) — reflens detects and warns about this; ingest the directory for true source fidelity.

Semantic search (opt-in)

Lexical FTS5 is the instant default and is excellent for code (symbol names, error strings). For concept queries ("how do they handle retries?"), build embeddings:

pipx install "reflens[semantic] @ git+https://github.com/cybertronayush/reflens"
reflens add /path/to/repo --name myref --semantic

Embeddings use fastembed (ONNX, no torch). It's opt-in because CPU embedding is slow (~25 chunks/s); the vector matrix is cached in-process so repeat queries are ~milliseconds.

Compared to

reflens is for "I want my agent to learn from N flagship repos I don't want cluttering my workspace." For a single repo you're actively editing, your editor's built-in tools are fine.

Honest limitations

• It's navigable, not omniscient. The agent must query well; the architecture-first design + AGENTS.md guidance steer it, but a lazy agent still gets shallow context. Name the repo for reliability.

• Semantic ingest is slow (CPU embeddings). Lexical-only is instant and the default.

• reflens_history needs a live git source dir — unavailable for URL-cloned or repomix repos (the digest still shows recent commit subjects).

• Not yet benchmarked against the "just clone it" baseline. The design is sound; measured task-success deltas are future work.

Install & setup

reflens has zero required runtime dependencies — it runs on the standard library alone (sqlite3 with FTS5, ast, and a hand-rolled MCP stdio server). The core works on Python 3.10+. The optional extras (semantic, code) pull fastembed/tree-sitter, whose wheels can lag the newest Python, so for those use Python 3.12 (recommended).

Pick one install path, then do the same three post-install steps.

Path A — pipx (isolated, simplest)

pipx install "reflens[semantic] @ git+https://github.com/cybertronayush/reflens"
# core only: pipx install "git+https://github.com/cybertronayush/reflens"

Path B — from source (this is exactly how the reference setup runs)

git clone https://github.com/cybertronayush/reflens && cd reflens
python3.12 -m venv .venv
.venv/bin/pip install -e ".[semantic,code,tokens]"

# make the `reflens` command available globally (PATH must include ~/.local/bin)
mkdir -p ~/.local/bin
ln -sf "$PWD/.venv/bin/reflens" ~/.local/bin/reflens

Then (any path) — wire it in, restart, stock the library

reflens install both        # registers the MCP server in OpenCode + Claude Code
                            # and writes a usage block into their global AGENTS.md / CLAUDE.md
# → restart OpenCode / Claude Code so they launch the server
reflens add <dir|git-url|repomix.md> --name myref   # populate the library (repeat per repo)
reflens list                # confirm

reflens install wires each host to launch the server via the interpreter that has reflens installed, e.g.:

// ~/.config/opencode/opencode.json  →  mcp.reflens
{ "type": "local",
  "command": ["/abs/path/.venv/bin/python", "-m", "reflens", "serve"],
  "enabled": true }
// ~/.claude.json  →  mcpServers.reflens  (command/args form, same interpreter)

Local state lives in ~/.reflens (override with REFLENS_HOME). Nothing leaves your machine. To update reflens itself, git pull (source) or re-run the pipx install, then restart your agent.

Contributing

git clone https://github.com/cybertronayush/reflens && cd reflens
python3 -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/python -m pytest -q

See CONTRIBUTING.md and ARCHITECTURE.md.

License

Apache-2.0.