knitli.com

What It Is

Knitli’s current flagship project, CodeWeaver, is an open-source code search server positioned as the “context layer” for AI Agents. It connects to tools such as Claude Code, Cursor, and Copilot via MCP, allowing agents to find relevant functions, classes, and snippets in a codebase using natural language, instead of inefficiently stuffing entire files into the context window.

Core Capabilities

CodeWeaver’s core feature is “always-hybrid” retrieval: keyword matching is used to find exact names, while semantic search finds conceptually related code, and the two are automatically combined every time. Results are then reranked by a second-stage model. It uses tree-sitter for structured parsing, with full syntax parsing for 27 languages and language-aware chunking for 166+ languages. It supports 17 embedding providers, including local options such as FastEmbed and SentenceTransformers, as well as cloud services like OpenAI, AWS Bedrock, and Voyage. Its MCP tool description is around 500 tokens, which keeps context overhead lower than LSP-style approaches with many tools.

Pricing and Deployment

The page states that CodeWeaver is released under the MIT OR Apache-2.0 open-source license and is available on PyPI and GitHub. No commercial pricing or paid plan is listed. For deployment, it supports fully offline operation, while cloud-based embeddings can also be used for better results. This makes it appealing for internal networks, air-gapped environments, or privacy-sensitive codebases.

Pros and Cons

Its strengths are that it is open source, uses a broadly compatible integration method, and has fairly opinionated defaults, so users do not need to manually combine keyword and semantic retrieval. Its language coverage is also broad. The drawbacks are equally clear: it is currently only at v0.1.2, and the official notes still mention bugs and rough edges; the documentation is still being improved, and installation/configuration requires a certain level of technical comfort. Performance will depend on codebase size and configuration. Direct symbol lookup is not yet exposed, and semantic matching cannot replace the precise navigation offered by an LSP.

Who It’s For and Access from China

It is best suited to developers who make heavy use of AI coding tools and want to control what context an Agent can see, especially in multi-IDE, multi-AI-client, or offline deployment scenarios. It is less suitable for users who just want an out-of-the-box all-in-one IDE experience. The page does not specify access conditions from China. Dependencies such as GitHub, PyPI, and OpenAI may involve network or payment uncertainty in mainland China, so users may want to prioritize local embedding configurations or compare alternatives such as Serena, Continue.dev, Aider, and Cursor.