semble — Fast Token-Efficient Code Search for Agents

~98% fewer tokens than grep+read. Index in ~250ms. Query in ~1.5ms. No GPU, no API key.

Semble returns only the relevant code snippets agents need, without grepping full files or reading directories. A natural-language or symbol query like "authentication flow" or "save_pretrained" returns exact chunks with file paths and line ranges — nothing more.

Installation

MCP (Claude Code — recommended)

# Requires uv: https://docs.astral.sh/uv/getting-started/installation/
claude mcp add semble -s user -- uvx --from "semble[mcp]" semble

MCP (Codex)

Add to ~/.codex/config.toml:

[mcp_servers.semble]
command = "uvx"
args = ["--from", "semble[mcp]", "semble"]

MCP (Cursor)

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "semble": {
      "command": "uvx",
      "args": ["--from", "semble[mcp]", "semble"]
    }
  }
}

MCP (OpenCode)

Add to ~/.opencode/config.json:

{
  "mcp": {
    "semble": {
      "type": "local",
      "command": ["uvx", "--from", "semble[mcp]", "semble"]
    }
  }
}

CLI / pip

pip install semble        # pip
uv tool install semble    # uv (recommended for CLI use)

Skill (any platform)

npx skills add https://github.com/akillness/jeo-skills --skill semble

When to use

• Search a codebase by describing behavior in natural language ("how is rate limiting handled")
• Look up a symbol or identifier without knowing the exact file ("save_pretrained")
• Discover code semantically similar to a known location (find-related)
• Give an agent token-efficient access to any repo via MCP instead of letting it grep/read full files
• Index a remote git repo without cloning first

Do not use when

• You need to read a full file or directory listing → use native Read, Glob tools
• You need regex or exact-string search → Grep is more appropriate
• The repo is too small to justify indexing (a few files) — just read them directly
• You need to run tests, build, or execute code — this is a search-only tool

CLI usage

# Natural-language search in a local repo
semble search "authentication flow" ./my-project

# Symbol search
semble search "save_pretrained" ./my-project

# Search with a limit on returned chunks
semble search "save model to disk" ./my-project --top-k 10

# Search a remote git repo (no clone needed)
semble search "save model to disk" https://github.com/MinishLab/model2vec

# Find semantically similar code given a known file+line
semble find-related src/auth.py 42 ./my-project

# Show token savings vs grep+read for the last query
semble savings
semble savings --verbose

Python library

from semble import SembleIndex

# Index local directory
index = SembleIndex.from_path("./my-project")

# Index remote repository (no clone required)
index = SembleIndex.from_git("https://github.com/MinishLab/model2vec")

# Natural-language or symbol query
results = index.search("save model to disk", top_k=3)

# Find semantically similar code to a known chunk
related = index.find_related(results[0], top_k=3)

# Inspect results
result = results[0]
print(result.chunk.file_path)   # "model2vec/model.py"
print(result.chunk.start_line)  # 127
print(result.chunk.end_line)    # 150
print(result.chunk.content)     # the function/class body

AGENTS.md / CLAUDE.md integration

Add this section to your project's AGENTS.md or CLAUDE.md to enable semble for all agents:

## Code Search

Use `semble search` to find code by describing what it does or naming a symbol, instead of grep:

​```bash
semble search "authentication flow" ./my-project
semble search "save_pretrained" ./my-project
semble search "save model to disk" ./my-project --top-k 10
​```

Use `semble find-related` to discover code similar to a known location (pass `file_path` and `line` from a prior search result):

​```bash
semble find-related src/auth.py 42 ./my-project
​```

`path` defaults to the current directory when omitted; git URLs are accepted.

If `semble` is not on `$PATH`, use `uvx --from "semble[mcp]" semble` in its place.

For Claude Code sub-agents, initialize once in the project root:

semble init

Performance benchmarks

Operating rules

1. Prefer MCP installation for interactive agent use; prefer CLI/pip for scripting and CI.
2. Use --top-k to limit results and keep context small — default is often too generous for agent prompts.
3. Use find-related after search when you need to expand from one known chunk into similar code.
4. Use semble init in project roots to pre-warm the index for Claude Code sub-agents.
5. If semble is not on $PATH, replace with uvx --from "semble[mcp]" semble in scripts.
6. Treat semble as the first pass — read full files only when the returned chunk is insufficient context.

Examples

# Search for how a feature is implemented
semble search "rate limiting middleware" ./api-service

# Find all code related to database migrations
semble search "database migration" ./backend --top-k 5

# Explore similar code patterns near a known function
semble find-related src/middleware/auth.py 88 ./api-service

# Index and query a remote library without cloning
semble search "tokenizer padding" https://github.com/huggingface/transformers

Source: MinishLab/semble — MIT License