TL;DR
Mantic.sh is a structural code search engine that ranks files across very large repositories in under 500ms without using embeddings, vector databases, or external services. The project released v1.0.12, which fixes scoring regressions and improves reliability while preserving local-first, zero-egress operation.
What happened
The Mantic.sh project, a local-first code search layer aimed at AI agents, published version 1.0.12 addressing a critical scoring bug and improving result reliability. The update restores a missing high-confidence scoring path in the legacy scan routine, ensures the BrainInspiredScorer assigns scores consistently regardless of intent confidence, and fixes an issue that produced undefined scored files and triggered broken fallbacks in an earlier release. Mantic emphasizes structural intent inference — using file paths, names, and metadata rather than reading file contents or relying on vector embeddings — to return ranked results rapidly. The README and docs highlight sub-500ms retrieval on large monorepos (the Chromium repo with roughly 480,000 tracked files is cited at 0.46s on an M1 Pro) and outline CLI usage, environment variables, agent integration rules, and licensing under AGPL-3.0. The project also provides a quick npm run and npx workflow and lists options for file-type filtering and impact analysis.
Why it matters
- Faster retrieval can cut the time AI agents spend collecting context, potentially improving responsiveness in code-writing or analysis workflows.
- Filtering by structure before ingesting file contents can reduce token usage, which may lower costs when paired with language models.
- A local-only, zero-egress design preserves code privacy for teams that cannot send source to cloud services.
- Avoiding embeddings and external vector stores simplifies deployment and eliminates dependency on third-party APIs or databases.
Key facts
- Release v1.0.12 fixes a missing high-confidence scoring path and resolves undefined scoredFiles fallback failures.
- Mantic reports sub-500ms retrieval, with a cited example: Chromium (~480k files, 59GB) indexed in 0.46s on an M1 Pro.
- The system does not use embeddings, vector databases, or external API keys; it ranks files using structural metadata.
- Core components include an Intent Analyzer, Brain Scorer, File Classifier, and Impact Analyzer that produces JSON, file lists, or MCP output.
- File enumeration relies on git ls-files to prioritize tracked files and speed traversal.
- CLI supports options like –code, –config, –test, –json (default), –files, –impact, and –session.
- Configuration can be controlled via environment variables (e.g., MANTIC_MAX_FILES, MANTIC_TIMEOUT, MANTIC_IGNORE_PATTERNS).
- Mantic is licensed under AGPL-3.0 and is free for individual developers, open source projects, and internal business use; a license is required to embed it commercially or offer it as a hosted service.
- Benchmarks in the README compare Mantic to vector-based search and show multix speedups (e.g., 11–22x on Chromium vs. vector search in the cited tests).
What to watch next
- Commercial licensing enforcement and how that will affect adopters who want to embed or host Mantic (license details provided in repo).
- Integration and adoption by local agent UIs — README lists native MCP support for Claude Desktop and Cursor as built-in compatibility.
- not confirmed in the source
Quick glossary
- Structural code search: A search approach that ranks and filters files using file paths, names, and repository metadata rather than full-content semantic embeddings.
- Embeddings: Numeric vector representations of text or code used by semantic search systems; not used by Mantic.sh according to the project.
- git ls-files: A Git command that lists files tracked in the repository; used by Mantic to speed file enumeration.
- AGPL-3.0: A copyleft open-source license that requires derivative works offered as network services to share source under the same license.
- Impact analysis (blast radius): An assessment that identifies files or modules potentially affected by a change to help prioritize review or tests.
Reader FAQ
Is Mantic.sh open source?
Yes. The project is published under the AGPL-3.0 license.
Does Mantic send code to external services?
No. The documentation states it runs locally with zero data egress.
Does Mantic use embeddings or vector databases?
No. It explicitly avoids embeddings and vector stores, using structural scoring instead.
How much does it cost to run?
The README provides a cost comparison showing Mantic as zero-dollar for the examples given; commercial embedding or hosted offerings require a license.
Which programming languages does it support?
not confirmed in the source
Mantic A structural code search engine for AI agents. Provides sub-500ms file ranking across massive codebases without embeddings, vector databases, or external dependencies. What's New in v1.0.12 Critical Fix: Implemented…
Sources
Related posts
- Universal Music, Nvidia extend AI deal to bring Music Flamingo to vast catalog
- Grok generates sexualized deepfakes of adults and children — can law stop it?
- Stop Doom Scrolling, Start Doom Coding: Terminal Dev from Your Phone