What it does
PASR provides source context to coding agents through a select_context tool. The token budget applies to the returned context, including labels, headings, separators and redaction. Whole computed spans are included only when the final output fits, so spans are not necessarily complete functions. Each selection produces a receipt that records every span's file:line, token count, retrieval signals, and the reason it was kept or dropped. Receipts are saved under .pasr/receipts on a best-effort basis, and cover only the context PASR delivers.
An agent would reach for PASR when it needs bounded, traceable source from a repository, alongside its native grep and file reads. Receipts expose heuristic query classifications, keyword coverage and routing advice. These are not calibrated correctness probabilities and do not prove that all necessary evidence was retrieved, so the source should still be checked. Without an agent, the pasr explain command prints a selection receipt for a project.
No daemon, vector database or manual index is needed. The package is installed through uv from PyPI, or from a GitHub release wheel, or from a source checkout for developers. The model decides when to use PASR tools rather than native tools. The pasr doctor command runs a diagnostic in a disposable fixture and never writes to the project; it reports runtime versions and check outcomes, not source or paths. Smaller selected context does not by itself establish cheaper or more accurate answers.
Limitations and requirements
Requires uv (with uvx on the PATH) and Python 3.10 or later, unless uv provisions Python. Offline selection needs installed dependencies and cached tokenizer data; optional model weights are separate. Initial setup and tokenizer downloads can require network access. Returned source can be forwarded by the client to a cloud model. Default redaction is a no-op, not automatic secret detection. Receipt persistence is best-effort, and PASR may write receipts and cache files in the workspace. MCP clients may require approval of the server; in unattended Codex runs the server must be explicitly pre-approved, and the MCP process is not sandboxed by the host shell. A doctor pass does not verify client registration, approval settings or answers.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.Apheironn/pasr
Local, token-budgeted code context with file:line provenance and selection receipts.
v0.4.0As shown on registry.modelcontextprotocol.io when we last read it