M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
R

Repowise

Community
Codebase intelligence layer: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions for AI agents and humans.
How to install ↓ Compare GitHub source repository ↗
Category
Dev Tools #38 of 438
Stars
★ 7.5k Very popular
Transport
stdio (local process)
Runtime
Python
Credentials
Optional API key
License
AGPL-3.0
Last commit
Tools
10
66FMRS · C

Repowise is a feature-rich code intelligence MCP server with a comprehensive toolset and deep analysis capabilities. It's particularly strong for teams wanting to cut AI token usage, gain code health and change risk insights, and auto-generate documentation. However, be mindful of index time, privacy configuration options, and the reliance on LLM providers for optional prose. Overall, it provides clear value for medium-to-large codebases and AI-powered development workflows.

Strongest · Documentation 16/20 Weakest · Reliability 9/20

Reliability
9/20
Security and permissions
13/20
Maintenance
15/20
Documentation
16/20
Setup experience
13/20
Why each score
Reliability 9/20
The manifest and README consistently identify a PyPI package `repowise`, started via `repowise mcp` (manifest encodes positional `mcp`), stdio transport, and a required pre-index step (`repowise init`). The README names ten task-shaped tools with clear signatures. However, the manifest marks `repo_path` as required while README quickstart examples (`repowise mcp`, `.mcp.json` `args: ["mcp"]`) omit it; no CI workflows or committed tests were supplied to verify the init/tool-list handshake or error paths; failure behavior on missing or uninitialized repos is undocumented. The conservative static cap is 12, and uncertainty lowers the score.
Security and permissions 13/20
No red-line behavior (malware, token theft, covert exfiltration, destructive defaults, fabricated maintainer/tests, real secrets in examples) is evident. The README discloses privacy boundaries: self-hosted analysis stays local, telemetry is anonymous and opt-out (`repowise telemetry disable`, `DO_NOT_TRACK`), API keys are user-supplied env vars with claimed zero vendor retention, optional LLM generation is gated behind cost estimates and approval, and raw source is said to be transient. Yet no code was reviewed to confirm least privilege, the MCP server's own network access, or key storage; default telemetry is still an outbound call unless disabled; tool-level permission/confirmation matrix is not fully specified. Hence 13.
Maintenance 15/20
The repository is active and not archived, has a clear license (AGPL-3.0), a current PyPI version (0.17.1), a referenced CI workflow, contributing guide, docs site, Discord, and a security/compliance note. Deduction: no commit history, release cadence, issue-response latency, or dependency-update evidence was supplied; stars are treated as discovery only. 92 open issues also signal possible response backlog. Score 15.
Documentation 16/20
Documentation is unusually layered: quickstart, per-client config examples (Claude Code, Codex, VS Code), MCP tool reference with signatures, CLI reference, costs, health/refactoring guides, benchmark methodology with disclosed limitations, privacy/security docs, and a `doctor` troubleshooting command. Deductions: the README contradicts the manifest on the required `repo_path` argument; many referenced docs (BENCHMARKS.md, MCP_TOOLS.md, etc.) are links not supplied, so they cannot be verified; marketing benchmark claims cannot be checked statically. Score 16.
Setup experience 13/20
Installation is a single `pip install repowise`, followed by `repowise init` and client wiring; README contains copy-paste `.mcp.json`, Claude Code CLI, Codex config, and a VS Code extension path, plus a Windows pip note. Under the static cap (15), the setup path is mostly clear. Deductions: README examples omit the manifest-required absolute `repo_path`, so the exact examples may fail on first run; no CI run or committed test was provided to prove the install-connect path works; the extra indexing prerequisite and path/transport flags are only partially explained.

Static review · not runListed 2026-08-07

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesUses the network

Best for

  • Developers who want deep codebase insights locally without uploading code to the cloud.
  • Teams needing pre-merge change-risk scoring and code health metrics.
  • Those looking to provide structured context to AI coding agents to save tokens and improve accuracy.

Not for

  • Users wanting a fully managed cloud service without any local installation (use hosted repowise.dev).
  • Slow index times may be a concern for large repos (e.g., 366s for Django), so not ideal if you need instant setup.
  • Non-technical users unfamiliar with command-line tools and MCP configuration.

Required permissions

  • Read access to the specified repository's file system (mandatory).
  • Optional: send code snippets to LLM providers for prose generation (if AI-generated docs enabled).
  • Optional: write access to Git or GitHub for PR bot comments if installed.
  • Read access to Git history for analysis.

Risks and side effects

  • Privacy: while local-first, enabling LLM prose or using hosted versions sends code to third parties.
  • Performance: indexing large repositories can take minutes, slowing initial setup.
  • Reliance on Python environment and pip; potential dependency conflicts.
  • Accuracy: heuristic-based scores and generated docs may be imperfect, especially for less-supported languages.

Setup

Before you start

Runtime：Python

ANTHROPIC_API_KEY optionalsecret Optional, for LLM-written wiki pages; get it from the Anthropic console.
OPENAI_API_KEY optionalsecret Optional, for LLM-written wiki pages; get it from the OpenAI platform.
GEMINI_API_KEY optionalsecret Optional, for LLM-written wiki pages; get it from Google AI Studio.
Other optional settings (1)
DO_NOT_TRACK optional Set to 1 to disable anonymous usage telemetry.
  1. Install with pip install repowise, then index your repo: cd /path/to/your/repo && repowise init (prompts for mode; use --no-prose -y for keyless structural docs).
  2. Start the MCP server: repowise mcp (run from the repo directory), then configure your client:
  • Claude Code: claude mcp add repowise -- repowise mcp or plugin /plugin install repowise@repowise.
  • Codex CLI: add [mcp_servers.repowise] command = "repowise" args = ["mcp"] to ~/.codex/config.toml.
  • Other MCP clients: configure command repowise mcp with the repo path.
  1. First real call: ask your agent to use get_overview or get_context.

Check that it works

Run repowise init first, then add the server with 'claude mcp add repowise -- repowise mcp'; the client's tool list should show the ten tools (get_overview, get_context, etc.), and asking the agent to run get_overview on the repo confirms the connection works.

Troubleshooting

  1. Index failures: ensure the repo path is correct, run `repowise init --path /absolute/path` and check error output.
  2. MCP connection issues: confirm `repowise mcp` is running in the repo directory and client config matches.
  3. Outdated index: run `repowise update` or `repowise watch` to keep in sync.
  4. No search results: ensure wiki is generated (`repowise generate`) and embedder configured for semantic search (Ollama works keyless).
  5. For security, run `repowise doctor` to check setup and API keys.

Things to try

Once connected, you can ask your AI assistant things like:

  • Use repowise get_overview to summarize this repo
  • Use get_context for src/auth.py to show hotspots, callers and governing decisions
  • Score the defect risk of main..HEAD with get_change_risk and list the tests it touches
  • Call get_health to show the lowest-scoring files and their refactoring plans

Tools 10

get_overview read-only
Provides architecture summary, module map, entry points, and git health. The first call on any unfamiliar codebase.
get_answer read-only
Hybrid retrieval (full-text plus vector) with PageRank bias and 1-hop graph expansion returns a cited answer with a calibrated retrieval_quality score.
get_context read-only
Generates triage cards for files, modules, or symbols: summary, signatures, hotspot bit, governing decisions, and symbol IDs. Batch multiple targets in one call.
get_symbol read-only
Returns source for one indexed symbol with exact line bounds. Cheaper and safer than Read plus offset math.
search_codebase read-only
Semantic search over the generated wiki, filterable by kind (implementation/test/config/doc), tagging each result's search method.
get_risk read-only
Analyzes hotspots, dependents, co-change partners, ownership, test gaps, and bug history. Pass changed_files for PR mode to get a directive block.
get_change_risk read-only
Scores pre-merge defect risk for a commit or range based on diff shape, ranked as a percentile against recent commits, plus tests coverage proves it touches.
get_why read-only
Retrieves architectural decisions, evidence spans, and supersession lineage. Falls back to git archaeology when no decisions exist.
Show 2 more tools
get_dead_code read-only
Returns unreachable code by confidence tier with cleanup-impact estimates, and cross-repo consumer detection in workspace mode.
get_health read-only
Returns per-file marker scores across three signals: defect risk, maintainability, and performance. Includes coverage, trends, self-check, and structured refactoring plans.

Use cases

Onboard to a new codebase: use get_overview to get architecture and entry points.
Pre-merge risk assessment: use get_risk and get_change_risk to identify high-risk files and impact.
Discover dead code and refactoring opportunities: use get_dead_code and get_health for concrete plans.
Generate documentation automatically: index and search the wiki via MCP.
Reduce AI agent token consumption: leverage precomputed context and command distillation (`repowise distill`).

Supported clients

Claude Code
Codex
Cursor
VS Code

Listed from the project's documentation, not tested by this site.

Overview

Repowise is an open-source code intelligence engine that provides AI coding agents with deep repository insights via MCP: a single index builds dependency graphs, git history, generated wiki docs, architectural decision records, and defect-validated code health scores. It reduces token consumption by up to 97% for context loading, offers concrete refactoring plans, and supports multi-repo workspaces. Self-hosted and local-first, your code never leaves your machine unless you opt into LLM-generated prose.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 62.9k · Tools 2 Compare with this →
Compare with other MCP servers

Source revision ee4df5cdf1fd Data synced 2026-10-12 Read the FMRS scoring method