Sign in Subscribe

Cora Code v0.8: Brain Mode, MCP Server, and the Rename

Cora Code v0.8: Brain Mode, MCP Server, and the Rename

Cora Code v0.8 is the biggest release since the project started. It ships a new name, a hybrid code search engine called Brain Mode, and an MCP server with 15 tools. This post covers what changed and how to upgrade.

One honest note up front: Cora Code is still early. The repo sits at 14 stars and 4 forks on GitHub. v0.8 is a real step forward, but this is a young tool. If you try it, expect rough edges and file issues.

The rename: Cora CLI becomes Cora Code

The crate was originally called cora-cli. As of v0.8.0 (issue #338), it is cora-code. The binary is still cora, so your muscle memory and shell aliases keep working. What changed is the package name, the repository URL, and the docs.

Why rename? The old name implied a generic CLI tool. Cora Code is specifically a code review and search tool. The new name says what it does. It also lines up with the MCP server story: when an editor connects to Cora Code as an MCP server, the name should describe the capability, not the interface.

The repository moved to github.com/codecoradev/cora-code. If you cloned the old repo, update your remote:


git remote set-url origin https://github.com/codecoradev/cora-code.git

The crate on crates.io is published as cora-code. If you installed the old cora-cli crate, uninstall it and install the new one:


cargo uninstall cora-cli
cargo install cora-code

Brain Mode: three search engines, one ranked list

Brain Mode is the headline feature of v0.8. It is a hybrid code search engine that runs three retrieval strategies in parallel and fuses their results.

The three engines:

1. FTS5 (SQLite full-text search) handles exact matches. Search for ThrottleGuard and FTS5 finds every occurrence. Fast, precise, oblivious to meaning.

2. Vector KNN (usearch HNSW, 256 dimensions) handles semantic matches. Search for "function that retries failed HTTP requests" and vector search finds the retry loop, even if none of those words appear in the code.

3. Graph BFS handles structural relationships. Given a seed symbol, BFS traverses the code graph to find callers, callees, trait implementations, and re-exports.

Each engine produces a ranked list. Brain Mode fuses them with Reciprocal Rank Fusion (RRF, k=60). RRF uses only rank positions, not raw scores, so it sidesteps the problem of comparing BM25 relevance against cosine similarity against hop distance. A code chunk that appears near the top of two or three lists gets a higher fused score than one that appears in only one list.

All indexes are local and the embedding model runs through your configured provider (Cora Code is BYOK). Query latency stays under 50ms on a moderate codebase. Brain Mode powers the cora review and cora scan commands, sending the LLM fused search results instead of a flat file list.

MCP server: 15 tools for editor integration

v0.8 ships an MCP (Model Context Protocol) server. Any MCP-compatible editor or agent can talk to Cora Code directly. The server exposes 15 tools covering code search, review, scanning, and project initialization.

The MCP server runs Brain Mode under the hood. When your editor asks "find the rate limiter", the server runs FTS5, vector KNN, and graph BFS, fuses the results, and returns them to the editor's LLM. The LLM gets better context than raw grep would provide.

To start the MCP server:


cora mcp

This starts a stdio MCP server. Point your editor's MCP client at cora mcp and the 15 tools become available, covering search, review, scan, branch analysis, init, index management, and configuration.

Other changes in v0.8

Beyond Brain Mode and the MCP server, v0.8 includes a pre-commit hook. Run cora init to install a Git pre-commit hook that runs cora review on staged changes. Reviews happen before the commit lands, not after.

Cora Code now ships 12 deterministic rules that run without an LLM. These catch common issues like unused imports, naming violations, and missing docs on public items, fast and deterministically. LLM-based review runs on top of these for deeper analysis.

cora scan can emit SARIF 2.1.0 JSON for CI integration. Pipe it into GitHub Code Scanning or any SARIF-compatible tool.

Multi-LLM support is in. Bring your own key for OpenAI, Anthropic, Google, or local models via OpenAI-compatible endpoints. The review quality depends on the model you choose.

The four main commands are cora review (review a diff or PR), cora scan (scan the whole project, emit SARIF), cora branch (review changes on a branch vs. main), and cora init (set up config and hooks).

Upgrade guide

From cora-cli to cora-code:


cargo uninstall cora-cli
cargo install cora-code

Your existing config at ~/.config/cora/config.toml still works. The config format did not change in v0.8.

First run after upgrade:


cora init          # set up config and pre-commit hook
cora scan          # build the index and run a full scan

cora scan builds the FTS5 index, the vector index, and the code graph on first run. Subsequent runs are incremental. The index lives in .cora/ inside your project.

One version note: Cargo.toml reports 0.8.2. GitHub Releases shows 0.8.1 as the latest tag. The 0.8.2 bump was a patch for a packaging issue and is functionally identical to 0.8.1 for end users. Install whatever cargo install cora-code gives you.

What is next

Cora Code is 14 stars. That is the honest number. The project has a clear direction but a small user base. Here is what is planned for the next releases:

  • More deterministic rules (target: 20 by v0.9).
  • MCP tool expansion based on real editor usage.
  • Incremental index updates on file watch, so you skip re-running `cora scan` after every change.
  • Performance work on large monorepos (100k+ files).

If you try Cora Code and hit a bug, file an issue at github.com/codecoradev/cora-code/issues. The project is MIT licensed and contributions are welcome.