siftr
CLI and MCP server giving coding agents Jev-powered semantic search, focused reads and list picking in about a second for about a cent.
Overview
siftr gives AI coding agents, and people, four read-only tools that answer 'where is it?' and 'which part matters?' without reading the whole codebase: semantic_search when the agent does not know the names it is looking for, focused_read to keep the relevant part of a large file, pick_relevant to choose the few items that matter from a long list such as test files, and an experimental filter_output for long logs. It runs on Jev through OpenRouter, and siftr agents install adds it as an MCP server to Claude Code, Codex, Cursor, Claude Desktop, Gemini CLI, Windsurf, Kiro, opencode, pi and others, backing up configs first.
On a frozen 202-issue split of SWE-bench Lite that it was never tuned on, the README reports 82% right-file-in-top-5 for search versus 52% for BM25, 92% of edited lines kept while cutting 59% of a file for read, and 81% for pick versus 38% for BM25. Search takes about 2 s on a 4,000-file repo and costs 1-2 cents. It is honest that grep is better when you know the exact words and that filter lost to grep on logs. Code is sent to a third-party API.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | How likely is this file or chunk to be where the described behaviour lives?(inferred) | — |
| Yes / no | Does this part of the file matter for the task?(inferred) | — |
| Yes / no | Is this list item (e.g. test file) relevant?(inferred) | — |
How it works
- Agent calls a siftr MCP tool with a question
- siftr gathers candidate files, chunks or list items
- Jev scores candidates in parallel
- Return ranked paths or trimmed file content with scores
Strengths
- 82% top-5 file accuracy on held-out SWE-bench Lite per the README
- About 2 s and 1-2 cents per search
- Installs into many agents automatically
Caveats
- Code is sent to a third-party API
- filter_output is experimental
- grep wins for exact-keyword queries
Getting started
curl -LsSf https://raw.githubusercontent.com/Bentlybro/siftr/main/install.sh | sh, then siftr setup and siftr agents install.At a glance
- Type
- Benchmark
- Status
- beta
- Language
- Python
- Runtime
- CLI + MCP server
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Fast, cheap judgment for AI coding agents: semantic search, focused reads and list picking in ~2s. CLI + MCP server on TypeSafe Jev. Benchmarked on SWE-bench.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide in siftr?
Which files, chunks or list items are relevant to the agent's question.
How accurate is search?
The README reports 82% right file in the top 5 on a held-out SWE-bench Lite split.
Which key do I need?
An OpenRouter API key.
Related projects
Local long-term memory for coding agents with an optional Jev reranker that orders recalled memories.
ProjectproductionTypeScriptMCP server giving coding agents fast typed decisions (ship gates, risk checks, file triage, page checks) as accept, review or escalate.
MCP serverbetaTypeScriptSingle MCP tool that indexes a repo and has Jev rank a shortlist so a coding agent reads the right files and lines instead of grepping.
MCP serverbetaPythonMCP server and semantic index that tags shadcn-format components once and matches a plain-language brief to them deterministically.
MCP serverbetaTypeScriptDeterministic code-graph context for coding agents on Postgres, AGE and pgvector, with optional Jev or local re-selection of symbols.
MCP serverbetaPythonCLI and MCP server where Jev scores code fragments for a behaviour question and returns exact excerpts with line numbers.
MCP serverbetaTypeScript