Local-first default
code stays on the machine unless the user explicitly opts into a cloud provider.
open source · local-first
Local code context for AI agents, priced honestly.
A local MCP server and CLI that indexes a repo, stores team memory locally, and gives coding agents bounded files, symbols, and notes before they edit.
Claude Code · one command
claude mcp add codebase-context -- npx -y codebase-context20frozen tasks
100retrieval rows
99officially scored
66×indexing cost delta
narrow case study · not a leaderboard
problem
Agents can drown in context
- Agents often start by opening half the repo before they know what matters.
- Memory turns into noise when every question gets the same notes.
- Search tools can sound confident even when they did not retrieve enough code to edit safely.
- Most benchmark writeups bury the install pain, indexing time, hardware assumptions, token bill, and failed setup paths.
product
What the tool gives the agent
- Local search that combines text matches with semantic retrieval
- AST-aligned chunks tied to real source structure
- Symbol references, imports, and project metadata
- Team memory and conventions stored locally
- CLI workflows for indexing, maps, and memory
- MCP server interface for coding agents
engineering
The tool had to fit normal dev machines
CPU by default
the tool has to run on a normal dev machine before any optional cloud profile is considered.
Framework-neutral core
specialized analyzers self-register instead of leaking framework rules into shared code.
Claim discipline
the write-up says what the run showed and blocks claims the evidence cannot support.
cost
Context is not free
A higher retrieval score is not enough if the tool takes forever to install, needs a hosted service, burns the token budget, or fails before the evaluator can judge the result.
- How long does install and first setup take?
- How long does the first index take?
- How fast is a normal query after indexing?
- How much disk and cache footprint remains?
- How many tokens reach the model?
- What did the benchmark orchestration itself cost?
- What failed before quality scoring?
evidence
One honest run
what the run showed
Close retrieval. Very different indexing cost.
jCodeMunch
27.1%recall
11smedian index
Codebase Context
25.7%recall
750smedian index
1.4 pp recall gap. 66× indexing cost difference.
100 rows · 20 frozen open-source tasks · narrow case study, not a winner claim
limits
What I will not claim
- It does not show that codebase-context writes better patches.
- It does not show that developers ship faster.
- It does not prove codebase-context is generally superior.
- It does not prove repeat stability because this run has one repeat per task and lane.
- It does not prove provider-dollar or normalized clean-install cost.
- The current evidence supports a narrow retrieval-overlap engineering note, not a winner table.
roadmap
Before stronger claims
- 1Normalize cold install, dependency cache, and index reuse before making cost-efficiency claims.
- 2Decide whether provider-dollar spend is out of scope or explicitly measured.
- 3Collect enough repeated judged runs per tool/task before stability or confidence claims.
- 4Improve Codebase Context cold indexing and exact code-range matching, then rerun the same frozen tasks.
- 5Keep GrepAI in the appendix unless the owner reopens a capped readiness run.
local-first · open source
MCP server and CLI · runs fully offline