open sourceLocal context engineContext engine for AI coding agents
uv tool install scubieeScubiee indexes your repository on your machine and hands your coding agent focused, ranked context, so it spends tokens writing code, not finding it.
The problem
Before your agent can write code, it spends context searching, opening files, and following connections.
01
The agent searches broadly before it knows where to work.
02
One useful file leads to the related code the change depends on.
session.ts
createSession()
middleware.ts
requireAuth()
user.ts
type User
token.ts
verify(jwt)
store.ts
readSession()
auth.ts
no references
Repeated searching
Broad grep and rg passes return far more matches than the task needs, and the agent reads through them.
Redundant file reads
Whole files land in context just to locate one function.
Connections by hand
Imports and callers get traced one tool call at a time before the first edit.
How it works
Three simple steps, plus a refresh loop: Scubiee builds a local index of your code and its relationships, then answers each agent query with focused context.
Build a local repository index on your machine.
scubiee init .Build · scubiee init / sync
Your repository
source files on disk
Merkle scan
Parse & graph
Tree-sitter AST
imports · calls
Chunk & embed
symbol chunks
CodeRankEmbed
Query · every agent call
Coding agent
Cursor, Claude Code,
Codex, OpenCode…
map over MCP
find · focus
gate / status
Retrieve & fuse
BM25 · semantic · graph
RRF fusion
Local index
~/.scubiee/
Both lanes share one local index. The build lane runs when you index or sync; the query lane runs on every agent call.
Capabilities
Scubiee gives coding agents focused repository context before they start writing.
Semantic retrieval
Search by meaning and surface the files that matter to the task, ranked, with the relevant code inline.
Illustrative query · results ranked by fused relevance
Code relationships
Follow relationships between files without opening the whole repository.
Focused context
Return focused context instead of large, token-heavy file dumps.
Incremental sync
Refresh changed files incrementally as your codebase evolves.
Local-first
Keep your repository index and source code on your machine.
Agent-native
Connect to the coding tools you already use through MCP. CUDA, DirectML, Apple MLX, or CPU: the right embedding profile is detected at setup.
MCP tools
Scubiee exposes one tool, map, with two configs. Pick by what you know; act on the first good answer and stop.
agent session · scubiee MCP
recommended flowstatus(root=".")once per session→ managed: true · ok: true
map find(query="resource management")don't know where→ ranked locations + relevant code inline
map focus(names="ResourceManager")or: know the name→ full body + callers/callees + siblings
editact on the first good answer→ agent's native edit tools
scubiee sync .if needed→ refresh changed files
map find“Where is X handled?”map focus“Show me this handler and how it's wired.”gate / status“Is this repo ready?”Integrations
Connect Scubiee to compatible coding agents through the Model Context Protocol. One command wires each of the 13 supported tools; reload MCP and you're done.
Cursor
scubiee connect --cursor
Claude Code
scubiee connect --claude-code
Codex
scubiee connect --codex
Kiroper project
scubiee connect --kiro
Windsurf
scubiee connect --windsurf
GitHub Copilotper project
scubiee connect --copilot
Clineper project
scubiee connect --cline
Roo Codeper project
scubiee connect --roo-code
Continue
scubiee connect --continue
Zed
scubiee connect --zed
OpenCode
scubiee connect --opencode
Amp
scubiee connect --amp
Pi
scubiee connect --pi
Any MCP client
if it supports MCP
Compatibility is provided through MCP and does not imply endorsement by, or partnership with, any agent provider. Tools marked (per project) read workspace-local MCP config, so run connect inside each project.
Before and after
Give your agent a shorter path from a question to the code it needs.
Native search only
Scubiee MCP
More room left for actual coding.
7 → 4 stepsfrom question to edit
Illustrative workflow for the same task, “Add resource management.” It shows the shape of the change, not measured data. Measured results are below.
Measured · three-turn experiment
What we measured when the same coding tasks ran with and without Scubiee.
Context used for the same task
Bars are a visual aid, scaled to the measured 60–70% range.
In our three-turn coding-agent experiment, Scubiee used about 60–70% fewer context tokens than the comparison workflow, and about 30% fewer tokens overall on average in some of the agent harnesses we tested. These are experiment results, not guarantees. Results vary by agent, repository, and task.
Local-first
Scubiee builds and uses its repository index locally, so your source code does not need to leave your computer.
Your machine
Network, one time: package from PyPI · embedding model download at setup
Your repository
source stays in place
<repo>/.scubiee/id.json
scubiee CLI
setup · init · connect · sync
Scubiee engine
background daemon
127.0.0.1:8765
Local index
~/.scubiee/projects/<id>
vectors · graph · chunks
scubiee-mcp
spawned by your IDE
forwards map calls
Coding agent
Cursor, Claude Code,
Codex, and more
Index and retrieval run locally
Parsing, embedding, search, and ranking happen in the engine on your computer.
Data lives on your disk
The index sits under ~/.scubiee. Wiping a repo removes its enrollment and index, never your source code.
Smaller context, same agent
Your coding agent still sends the context it chooses to its model provider. Scubiee keeps that context focused.
Ready when you are
Less searching. Better context. More building. Free, open source, and running on your machine.
Four commands
Python 3.11+ · uv
01Install
uv tool install scubiee02Set up your machine
scubiee setup03Connect your agent
scubiee connect --cursor04Index your repo
scubiee init .connect wires your IDE once; init indexes each repo. Swap --cursor for your tool on the install page.