What website is this?
CtxSift (ctxsift.dev) is an open-source agent skill aimed at token waste from command output and context compaction. Commands like systemd, docker, and git often produce long logs that agents pull wholesale into context; after compaction, they may need to rerun commands or reread files to recover the same conclusions. CtxSift runs a local compress–cache–recall loop: it keeps only the anchors needed for the next step and lets you search them back as sessions grow. The workflow stays lightweight with no MCP dependency; compression runs on CPU by default, with optional CUDA or LiteLLM-compatible remote models.
Key Features
- Instruct pipe or command-capture output to extract only the fragments needed for the next step
- Store compressed results with command metadata for semantic retrieval instead of reruns
- Run compression on local CPU/GPU or switch to LiteLLM-compatible remote models
- Maintain recall embeddings locally on a separate path, whether compression is local or remote
- Mark records stale when source files change, down-rank expired context, and clean up
Use Cases
- While debugging in Cursor or Claude Code, keep only systemd error line numbers and causes in context instead of full journal pages
- After multiple compaction rounds, recover why the last git clone failed without pulling a large repo again
- For local agents such as Aider or OpenCode where data should stay on-device, keep compression and retrieval on the machine
- When juggling multiple feature branches, cache structured conclusions from similar commands so agents stop re-parsing the same log formats
Who is it for?
- Developers who already use agents heavily in Cursor, Codex, Claude Code, Copilot, and similar environments
- Debugging, ops, and full-stack work where terminal output is long and sessions often hit token limits and get compacted
- Developers who want an open, locally controlled option without adding MCP servers or sandboxes just to save tokens
- May not fit: users who only run occasional short commands and rarely see long logs
- May not fit: scenarios where compliance requires retaining complete raw terminal output with no bytes dropped
How It Compares to Similar Tools?
- CtxSift focuses on lightweight distillation of command output and post-compaction state, not a team knowledge base or cross-repo index
- If log flooding and post-compaction amnesia are the main pain points, a local skill is often easier to control than stacking extra token tools
- If you need a collaborative context dashboard, or provider-side summarization is enough and you do not need searchable local cache, weigh whether stale/freshness handling matches your workflow
FAQs
Q: Is CtxSift free? Do I have to stay online?
A: The site labels it Local, Free, and Opensource. Compression and recall work on local CPU by default; with a LiteLLM remote endpoint, compression can use a hosted model while recall embeddings stay local.
Q: Do I need an MCP server?
A: No. It works as an agent skill through pipe/command capture and local caching, without MCP or multi-tool orchestration.
Q: How is this different from asking the model to summarize?
A: It caches conclusions with command metadata for later retrieval; when source files change, records are marked stale to reduce reuse of outdated context.
Q: Which agent environments are supported?
A: The site lists Cursor, Claude Code, Codex, Copilot, Aider, and others; install via standalone scripts or let an agent run the install skill—see the docs for details.
