One local memory server that Claude Code, Codex, Cursor, and
other agents share. It lives in a single SQLite file on your
machine, uses only the Python standard library, and makes no
network calls.
An agent looks something up only when it is missing context. It
must give a reason for each lookup, and it gets back at most
three memories, so a normal turn makes no lookup at all. A
correction you give an agent becomes a memory only after you
repeat it in a second session.
A local Rust tool that sits between a coding agent and your repo.
It caps how much a single file read or search can return. It also refuses an edit if the file changed since the agent last read it.
It indexes the workspace with tree-sitter, supports
ripgrep-compatible search, and tags every piece of context with
its source. The code, the index, and any optional vectors never
leave your disk.
My own coding agent CLI. After every write it checks the file for syntax errors. If the edit broke the syntax, it rolls the edit back before the model is told it worked.
When the context window fills, older turns get compacted on your machine instead of rewritten by an LLM. The dropped text stays around as citations you can open. An allowlist blocks untrusted
tool output by default.
It runs on WebKit, the engine already inside macOS, so it ships
no engine of its own. A tab you leave alone for 30 minutes goes
to sleep and frees its memory.
SearchX applies the uBlock Origin filter lists inside the WebKit
network layer, so the browser never requests a blocked ad. The
WebKit blocker cannot run filters like :has-text or :xpath, so
SearchX runs those itself. The new tab redraws any picture you
pick as pixel art.
I built this while job hunting on a visa. It answers “does this
company actually sponsor?” by joining H-1B records,
company profiles, salary data, and live job listings.
Celery and Redis workers pull from applicant tracking systems,
drop duplicates, and match listings to your resume with TF-IDF.
It also generates PDF reports.