Pensieve 0.6: Local Code Search and MCP Context for AI Agents

Pensieve is a free, open-source local search tool that gives coding agents relevant code, documents and earlier conversations through MCP. Its latest releases add guided setup and visibility into agent search activity, making it easier to judge whether a local index is worth maintaining.

Version 0.6.0 arrived on October 11 with a guided setup flow. The project reached 0.6.3 on October 12, following an update that shows how agents use its search tools. These are practical additions to an existing project, with a useful question behind them: can better retrieval reduce the work an agent spends finding the right context?

How Pensieve gives coding agents context

Pensieve indexes selected files, repositories and supported agent conversations on your computer. You can search through a Mac panel, or connect a coding agent through the Model Context Protocol. Results include relevant passages and file locations, so the agent can inspect a smaller set of candidates.

The interesting part is the mix of sources. A function may explain what the code does today, while an earlier agent session records why someone changed it. Searching both gives a developer a way to recover work that would otherwise be scattered across terminals and chat histories.

Pensieve supports semantic queries alongside exact phrases and filters. Claude Code, Codex, Cursor and other MCP clients can connect. Optional instructions and skills encourage agents to search the index before falling back to their usual tools.

What changed in the latest releases

The release history shows several changes across October 11 and 12. Version 0.6.0 walks users through choosing how they work, allowing folder discovery and selecting sources. Agents can carry out the same setup through MCP.

Version 0.6.2 adds a Usage tab showing calls, timings, empty searches, errors and recent activity. That helps answer a basic adoption question: is your agent using the integration, and is it finding anything? Version 0.6.3 fixes interactions with author cards in the Map and an error shown when AI insights are off.

The downloadable Mac app requires Apple Silicon and macOS 14 or later. The repository also documents a source-installed backend and browser Map for Linux and Intel Macs. The native search panel remains a Mac feature. Check GitHub for the current version: the product page still displayed 0.5.9 during this source check.

What the Pensieve benchmark measures

The benchmark harness compares Claude Code's usual search tools with Pensieve plus an instructions block. The reported evaluation used Claude Sonnet on 72 held-out code-finding tasks across four repositories, repeated twice for 144 runs per setup.

The maintainer reports that Pensieve found the correct file in 131 runs, compared with 123 for the baseline. Average tool calls fell from 3.44 to 1.97, and reported lookup cost fell from $0.050 to $0.041. The evaluation used Pensieve 0.5.6, so these are earlier results carried into the current documentation.

Those numbers describe finding a file. They do not establish a reduction in the total cost of implementing a feature or fixing a bug. The comparison also includes instructions that change how the agent searches. Claude's prompt cache affects the reported costs, and the documentation advises comparing runs within the same batch.

A useful trial would repeat representative searches from your own repository, including exact error messages and vague descriptions. Record whether the agent finds the right place, then see whether that improves the task you actually care about. The published scripts provide a starting point.

Local search still needs deliberate access choices

Indexing and search run locally. Retrieved text can still reach a connected cloud agent, and optional summaries use either local models or the selected Claude Code/Codex cloud provider.

The versioned documentation says automatic stdio scope selects a named scope containing the repository, otherwise that repository and its worktrees and sessions. Outside a repository, or over HTTP without a named scope, the connection can access the whole index. Set a narrow scope before exposing sensitive material.

Start with one repository and a few relevant sources. Check the Usage tab, inspect the returned passages and compare the results with your existing workflow. Pensieve's value will depend on how much useful context your agent currently misses and whether maintaining the index saves enough repeated searching to justify it.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool teams turn technical details into clear documentation and launch stories, including what a benchmark shows and where a product fits into everyday work.

Book a 30-minute call.