Wisegit
About
MCP server that extracts decision intent from git history and protects intentional code from AI modification.
Details
- Author
- Sandip124
- Downloads
- 298
- Categories
- Other, Developer Tools
Jump to
- Zero config setup with a single command
- SQLite database (no Docker or external services)
- Team support via .wisegit/ directory tracked by git
- AI-era commit origin detection (human, AI reviewed, AI unreviewed)
- Theory holder tracking per function (healthy / fragile / critical)
- HTML codebase health report (wisegit report)
- Grounded in 12 published software engineering papers
Setting up with Highlight
This MCP is not yet compatible with Highlight’s one-click setup. However, you can still use it with Highlight by following these steps:
- Download and install Highlight from highlightai.com/download
- Navigate to the plugins tab and select "Add Custom Plugin"
-
Configure the plugin with the settings below
Plugin Name
WisegitCommand (node, npx, python, etc.)Please refer to the README for specific instructions on how to obtain API keys or other required environment variables.
- Enable "Start Automatically" if you want the plugin to start when Highlight launches
From the repository
Run npx @sandip124/wisegit setup in any repository—zero configuration. After indexing, AI agents call MCP tools such as get_file_decisions to retrieve decision manifests, and create_override to override frozen functions with user approval in Claude Code.
get_file_decisions
Get the decision manifest for a file — shows freeze scores, intent history, and recovery levels for all tracked functions. Call this BEFORE editing any file.
get_freeze_score
Get the freeze score and signal breakdown for a specific function. Shows why a function is frozen/stable/open.
search_decisions
Search past decisions by keyword. Finds intent history across the entire repository.
create_override
Request an override for a FROZEN or STABLE function so it can be modified. The user MUST approve this action. Provide a clear reason why the override is needed. The override is recorded in the audit trail permanently.
get_function_history
Get the full chronological decision timeline for a specific function. Shows every event: creation, changes, overrides, issue enrichments.
get_theory_gaps
Get functions in a file where the original decision rationale may be unrecoverable (primary author inactive, timeline gaps). These need extra caution.
get_branch_context
Get branch merge history — what branches were merged and what they changed. Useful for understanding migrations and cross-platform decisions.
extract_intent
Extract decision intent for a function's NOISE commits using the host LLM (no Ollama needed). Call this when a function has events with LOW or missing intent to recover the 'why' behind changes.
Claude Desktop / Cursor
Paste into your MCP client config file to install this server.
{
"mcpServers": {
"wisegit": {
"wisegit": {
"command": "npx",
"args": [
"-y",
"@sandip124/wisegit",
"serve"
]
}
}
}
}McpServers
{
"wisegit": {
"command": "npx",
"args": [
"-y",
"@sandip124/wisegit",
"serve"
]
}
}WiseGit is a local MCP server that reconstructs the "why" behind code decisions from git history — so AI agents respect what was intentional, not just what compiles.
The problem: LLMs treat all code equally. A manually-tested race condition fix (sleep(350)) and a broken stub look identical — both are just text. When Claude Code sees "find bugs," it removes the sleep. Production incident. The proof of intention exists in git history, but nobody extracts it.
What wisegit does:
1. Indexes your git history — walks every commit, parses diffs at the AST level using Tree-sitter (C#, TypeScript, JavaScript, Python, Go, Rust), classifies commits (STRUCTURED/DESCRIPTIVE/NOISE), and extracts decision intent.
2. Computes freeze scores — each function gets a 0–1 score derived from 7 signal categories: git history (reverts, age, contributor count), issue enrichment (Won't Fix/By Design from GitHub/GitLab), code structure (magic numbers, defensive patterns), test signals, structural importance (PageRank from call graph), Naur theory signals (institutional knowledge loss), and Aranda signals (timeline gaps, forgotten patterns).
3. Serves decision manifests via MCP — before any file edit, the AI agent calls get_file_decisions and sees which functions are FROZEN (don't touch without override), STABLE (proceed with caution), or OPEN (safe to modify). Theory holders show who on the team understands each function.
4. Enables seamless overrides — when the AI needs to modify frozen code, it calls create_override with a reason. The user approves with one click in Claude Code. The override is recorded in an append-only audit trail shared with the team via git.
8 MCP tools:
- get_file_decisions — decision manifest with freeze scores, theory holders, override status
- get_freeze_score — detailed score breakdown for a function
- get_function_history — full chronological decision timeline
- get_theory_gaps — functions with no active theory holders (Naur death)
- get_branch_context — branch merge history and migration context
- search_decisions — keyword search across all past decisions
- create_override — override frozen functions (user approves in UI)
- extract_intent — extract intent for NOISE commits using host LLM via MCP sampling
Key features:
- Zero config — npx @sandip124/wisegit setup in any repo, that's it
- SQLite database (no Docker, no PostgreSQL, no external services)
- Team support via .wisegit/ directory tracked by git (JSONL files merge cleanly)
- AI-era commit origin detection (HUMAN / AI_REVIEWED / AI_UNREVIEWED)
- Theory holder tracking per function (healthy / fragile / critical risk levels)
- HTML codebase health report (wisegit report)
- Grounded in 12 published software engineering papers
Academic foundation: Naur (1985) theory building, Aranda & Venolia (2009) bug repository analysis, Kim et al. (2007) fault prediction from cached history, Ying et al. (2004) co-change mining, Giger et al. (2011) AST-level change analysis, Aryani et al. (2014) domain-based coupling, and 6 more.
Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
