↓ Skip to main content
tomrochette.com
  1. Agents/

Semble

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~880 words)
agent-curated 173 code-search 2 context-engines 7 developer-tools 24 fully-ai-generated 263 llm=glm-5.3-flash 178 local-first 2 research-note 148
Table of Contents

Semble is a MIT-licensed local code search tool built for coding agents: it chunks a repository along syntax boundaries, indexes it with static Model2Vec embeddings plus BM25, and answers natural-language queries with only the relevant snippets. Facts below verified as of 2026-09-13.

Semble’s bet is that a code-specialized index small enough to run in under a second on any CPU can replace the most expensive thing an agent does, which is grep-failing its way through full-file reads.

What it is
#

One Python tool, three delivery surfaces, zero setup. semble install detects installed agents (Claude Code, Codex, OpenCode, Pi, and others) and wires any combination of an MCP server, CLI usage instructions in AGENTS.md or CLAUDE.md, and a dedicated semble-search sub-agent; semble uninstall reverses it. The engine splits files into tree-sitter chunks, scores queries with static potion-code-16M-v2 embeddings and BM25 in parallel, fuses the rankings with reciprocal rank fusion, then reranks with code-aware signals; everything runs on CPU with no API keys or external services. It ships as a CLI (semble search, find-related, a docs/config/all content switch), a two-tool MCP server (search, find_related), and a Python library, with gitignore-aware filtering, local path or git-URL inputs, and cached indexes invalidated on file change. It is built by MinishLab, the two-person team (Stephan and Thomas, per the launch post) behind the Model2Vec static-embedding project.

Status
#

Young, active, and unusually well received for a search tool. 6,060 stars and 265 forks since the repo appeared on 2026-04-06, with the last push 2026-09-12 and v0.6.0 released 2026-09-11 (GitHub API, as of 2026-09-13). 89,861 PyPI downloads in the trailing month (PyPIstats, as of 2026-09-13). The launch Show HN (May 17, 2026) drew 445 points and 151 comments, on top of two quiet warm-up threads in April; a third-party VS Code extension appeared in August, which is the usual early-ecosystem signal.

Strengths
#

  • The speed numbers, if they replicate, change the deployment model: about 500 ms to index an average repo and about 1 ms per query on CPU means per-session on-demand indexing, with no standing service and no stale-index failure mode.
  • The retrieval stack is a real hybrid design, static embeddings for meaning plus BM25 for identifiers, not one trick, and the pipeline and ablations are published.
  • Token-efficiency framing is measurable at the margins that matter: the benchmark reports 97 percent recall at 2k retrieved tokens where grep-plus-read needs a 100k window for 85 percent.
  • Grounded, self-measured savings: semble savings records actual calls and estimated avoided tokens per machine, so a team can audit the claim on its own traffic instead of trusting the vendor.

Cautions
#

  • Every published number is self-run, and the founders say so: in the launch thread they state they do not claim end-to-end agent improvements because they have not measured them, so the 99-percent token reduction is a retrieval-layer figure, not a work-quality one.
  • The benchmark bar is modest: parity is claimed against CodeRankEmbed, a 137M-parameter model, and the static-embedding design caps how much semantic depth is available at any price.
  • The “fewer tokens than grep” baseline has been challenged on its own terms: grep spends zero tokens, the cost is the agent’s read strategy afterward, and commenters asked whether the comparison is 100k-to-2k or 1k-to-20, a question the benchmark answers only in aggregate.
  • Two maintainers, a five-month-old project, and a pre-1.0 (0.6.x) API are the usual early-adoption risks.

Pricing
#

Free and MIT-licensed; there is nothing to buy. No cloud tier or paid plan exists as of 2026-09-02, so the cost model is entirely the compute you already have and the model tokens the snippets displace.

Compared to
#

  • Repomix: the inclusion-first alternative, one deterministic file for the whole repo; Repomix wins for small repos and one-shot questions, Semble wins once the pack would eat the window.
  • Graphify: the structure-first alternative, a deterministic AST knowledge graph with citations; choose Graphify for how-the-codebase-connects questions, Semble for where-is-the-code-that-does-X.
  • qmd: the same hybrid static-embedding plus BM25 plus rerank architecture, aimed at personal notes and docs rather than codebases; the two are complements, one per corpus.

Bottom line
#

Recommended for agent users on repos big enough that grep-plus-read burns real tokens, who want a local, key-free index and will measure savings with semble savings rather than take the benchmark on faith. Not for small repos where a Repomix pack or plain grep is already enough, and not for anyone needing vendor support or an agent-level quality guarantee, because neither exists yet.

Changes
#

  • 2026-08-30 - Created in the Context engines category from that day’s entrant scan, citing six fetched sources.

See also
#

References
#