codescope

---

Why This Exists

Most AI coding assistants still embed every file as a vector, nearest-neighbor a chunk, and pray it's relevant. When you ask "if I change User::email, what breaks?" they read 40 files and burn 150,000 tokens guessing.

That's not a code intelligence problem. It's an architecture problem. Vectors can't do graph traversal. Fuzzy search can't tell you who calls whom.

Codescope solves it the right way: parse the code into a knowledge graph — functions, calls, imports, type hierarchies, decisions, all of it — and let agents walk the graph instead of flipping through files.

Question: "Who calls parse_config transitively within 3 hops?"

Traditional RAG:        Codescope:
─────────────────       ─────────────────
~150K tokens            ~1-2K tokens
~12 seconds             ~3 ms (end-to-end)
Fuzzy text match        Deterministic edge walk
Guess confidence        Actual answer

---

Install

Platform	Command
Linux / macOS	curl -fsSL https://raw.githubusercontent.com/onur-gokyildiz-bhi/codescope/main/install.sh | bash
Windows	irm https://raw.githubusercontent.com/onur-gokyildiz-bhi/codescope/main/install.ps1 | iex
Homebrew	brew install onur-gokyildiz-bhi/codescope/codescope
Claude Code plugin	/plugin marketplace add onur-gokyildiz-bhi/codescope then /plugin install codescope@codescope
Build from source	cargo install --git https://github.com/onur-gokyildiz-bhi/codescope

Already installed? codescope --version to check. Update in-place with codescope upgrade.

Pre-built binaries: x86_64-unknown-linux-gnu, aarch64-unknown-linux-gnu, aarch64-apple-darwin, x86_64-pc-windows-msvc.

---

Quick Start

# 1. Bring the bundled SurrealDB server up (idempotent)
codescope start

# 2. In your project — writes .mcp.json and indexes your code
cd your-project
codescope init

# That's it. Claude Code, Cursor, Codex — any MCP-compatible
# agent in this project now has codescope wired in.

Target a different agent:

codescope init --agent cursor          # .cursor/mcp.json
codescope init --agent gemini-cli      # ~/.gemini/settings.json
codescope init --agent vscode-copilot  # .vscode/mcp.json
codescope init --agent codex           # ~/.codex/config.toml
codescope init --agent windsurf        # ~/.codeium/windsurf/mcp_config.json
codescope init --agent kiro            # .kiro/settings/mcp.json
codescope init --agent cline           # .vscode/cline_mcp_settings.json
codescope init --agent antigravity     # global + GEMINI.md nudge

Daemon mode (MCP + Web UI in one process):

codescope init --daemon   # port 9877 — per-repo routing at /mcp/<repo>
# Web UI: http://localhost:9877/

LSP mode (editor-agnostic — VS Code, Zed, Neovim, Helix, IntelliJ):

codescope lsp
# Go-to-def, Find References, Hover, Workspace Symbols — all graph-backed.

Daily operation:

codescope status            # surreal server state
codescope gain              # cumulative token savings
codescope insight           # per-repo + hourly activity
codescope session           # last 5 MCP sessions with tails
codescope upgrade           # in-place self-update
codescope repair --repo <n> # drop + re-index a corrupted repo
codescope hook install      # PreToolUse bash-suggest for Claude Code
codescope doctor            # diagnose setup (+ --fix)

---

What Your Agent Gets

A structured MCP surface your agent programs against, instead of scrolling output.

Code navigation & impact

• search(mode) — fuzzy / exact / file / cross_type / neighborhood / backlinks
• find_callers / find_callees — 1-hop call graph
• impact_analysis — transitive BFS blast radius
• type_hierarchy — inheritance chains
• context_bundle — file overview with delta-mode caching (97% savings on repeat visits)

Knowledge — memory that survives sessions

• knowledge(action) — save / search / link / lint; scopes project / global / both
• memory(action) — save / search / pin
• capture_insight — record decisions in real time
• manage_adr — Architecture Decision Records