aoci-spec/aoci-code

AOCI-CODE enables coding agents to write a one-line description for each file: what it is responsible for, which files must be read in conjunction, what dependencies callers have, and what constraints are not visible in the code but must not be violated. The index is stored in the repository and managed alongside Git versions. Subsequently, a local MCP service detects code changes and lists the entries that need updating. When switching personnel, agents, or starting a new conversation, reading the index once allows for a seamless handover. The author states that codebases of approximately 50 ten thousand lines or fewer can be indexed in one pass, with the initial indexing requiring about one hour per 20 ten thousand lines.

AOCI-CODE enables coding agents to write a one-line description for each file: what it is responsible for, which files must be read in conjunction, what dependencies callers have, and what constraints are not visible in the code but must not be violated. The index is stored in the repository and managed alongside Git versions. Subsequently, a local MCP service detects code changes and lists the entries that need updating. When switching personnel, agents, or starting a new conversation, reading the index once allows for a seamless handover. The author states that codebases of approximately 50 ten thousand lines or fewer can be indexed in one pass, with the initial indexing requiring about one hour per 20 ten thousand lines.

Why It Matters Agents do not need to re-read the entire repository each time; the index is written by your own agent, the local MCP service prompts updates when changes are detected, and it can also index database table structures.

Stars 2113 (+23 / 24h)

Reasons

  • 23 new stars in 24 hours, total stars 2113 (indicating attention level only)
  • The README states that it only reads local code and database table structures, does not connect to the internet, does not store database passwords, and explains the differences from CodeGraph, RAG, and LSP

Usage

  • Send the two instruction blocks from the "Quick start" section of the README sequentially to Codex, Claude Code, or Cursor: the first block downloads, verifies, and runs init and scan; after restarting the agent, the second block builds the index and provides a link to the dashboard

Risks

  • The license is FSL-1.1-MIT (Fair Source, source code visible), which is not an OSI-approved open-source license, and competitive commercial use is restricted
  • Currently in v0.2.0-rc2 pre-release; initial indexing requires the agent to read every file, consuming a large number of tokens

Open

Trendshift(GitHub Trending 快照)

Explore the latest AI Radar updates