10 lesser-known MCP projects worth a look

The famous MCP servers connect an assistant to products you already pay for. The interesting new ones give a coding agent abilities it never had: a memory of past sessions, a phone to test on, a map of your codebase, an up-to-date view of your dependencies, even a ruler for construction drawings. Here are ten we reviewed, what each does well and where each one asks for trust.

Memory and context for coding agents

deja-vu (4.2) solves the agent that has forgotten yesterday. Instead of asking the agent to decide what to remember, it indexes the transcripts Claude Code, Codex and Cursor already write to disk, so a fix worked out in Codex last month can surface in a Claude Code session today. The project reports 97.2% recall at 5 on the LongMemEval-S benchmark and lookups under a millisecond. The catch: search is keyword-based unless you add an embeddings endpoint, and the transcripts it reads may still contain secrets.

Repowise (4.1) builds a local map of a repository, from code and git history to ownership and test coverage, which the agent queries instead of grepping file by file. On the maintainers' own test of 43 questions about Django, an agent with Repowise made 3.8 tool calls instead of 7.2. The first full index of a large repository is slow, and the AGPL-3.0 licence matters if you plan to build it into a product.

stalemap (3.8) tackles code written for the library version the model was trained on rather than the one you pin. It compares the type definitions of two versions and hands the agent only the APIs that changed in the parts your code imports. In a pre-registered test of 66 tasks, Claude Sonnet 5.5 passed 89% of sessions with stalemap against 68% without it. It covers JavaScript and TypeScript only and raises the cost of a typical session by about 22%.

XERJ (3.6) is a local search engine for large folders of code, documents and logs, fast and largely compatible with the Elasticsearch API. Read its agent instructions before installing: by default they tell the agent to file reports to the project automatically, and the quick start runs without authentication or TLS.

Agents that can see and touch

agent-device (4.3) closes the loop on mobile work: an agent that has written a login screen can open the app on a simulator or real device, type test credentials, tap the button and check what appears. It works from compact accessibility snapshots rather than screenshots, which keeps each step cheap, and claims devices per git worktree so two agents do not fight over one simulator. It is still on 0.x releases.

Scrapling (4.4) is the scraper for sites that fight back: one Python library that goes from a simple fetch to a stealth browser session without rewriting your parsing code, and finds elements again after a site renames its classes. Its MCP server hands that to an agent. The anti-bot features put the legal and terms-of-service risk on you, and robots.txt is respected only if you switch it on.

screenpipe (3.7) records your screen and audio, keeps it on your machine and lets agents search it over MCP, so "what was the website I looked at yesterday?" has an answer. It also captures private messages and passwords on screen, sends analytics by default, and since June 2026 is under a commercial licence: free for personal use, but businesses need an Enterprise plan.

Making things

ffmpeg-skill gives an agent 42 typed FFmpeg tools for cutting, joining, captioning, removing silence and reformatting video, all processed locally. For speech-led clips with careful confirmation steps, editonny (3.4) is the narrower, more cautious alternative, though it is twelve days old and tested mainly on Macs.

Audio Studio MCP (3.7) generates instrumental music and sound effects with Stable Audio 3 on your own Mac or NVIDIA PC, then does the dull work an audio editor would: gapless loops, loudness normalised to a target, fades and export for a game engine. It needs Apple Silicon or an 8 GB NVIDIA card, makes no vocals, and the model weights need an Enterprise licence above US$1 million in revenue.

OpenTakeoff (3.9) shows how far MCP reaches beyond software. It is a free takeoff tool for construction drawings, built by a working estimator, whose MCP server lets an agent set a scale, measure areas, count symbols and price quantities across 58 tools. It refuses to measure an unscaled sheet rather than guess, and only a person can apply the approved seal. The licence moved from Apache-2.0 to AGPL-3.0, and 58 tools take a lot of an agent's context.

How to try them safely

Most of these run locally, which keeps your data on your machine but gives them broad access to it. Read the licence and default settings first, add one server at a time, and watch what the agent does with it for a few sessions before you rely on it.

Our take: the best MCP servers no longer just fetch things. They give the agent a memory, a pair of hands or a domain expert's caution, and the ones worth keeping are those that say clearly when they are unsure.