# MCP Rack Discover MCP servers built by Alexei Udall, a curated selection from other makers, and practical guides. Clear setup details, honest project status and open data for agents. Maintainer-authored metadata. Inclusion is not an independent security audit. Confirm current versions and permissions with the source. Generated: 2026-09-07T17:38:20.541Z Build time, not a verification date. ## Owner-built projects # gimp-agent-mcp https://mcprack.dev/servers/gimp-agent-mcp/ An MCP server that hands AI agents the whole of GIMP 3 rather than a hand-picked slice. Every one of GIMP's roughly 1,000 Procedure Database functions and 207 GEGL filters is reachable through runtime introspection, so nothing goes stale when GIMP updates. The agent can see its work: a PNG render after each step, before/after/diff comparisons, and pixel measurement instead of guessing. Recipes turn repeatable jobs (stickers, icon sets, watermarks, contact sheets, social cards from a layout manifest) into one call, and a batch runner applies them to whole folders. Status: beta Transport: stdio License: Apache-2.0 Source: https://github.com/SarutobiSasuke8/gimp-agent-mcp - Tools, reusable recipes and bundled skills for layered assets, sprite sheets and batch jobs - Runtime introspection of the PDB and GEGL, with typed argument descriptions and enum choices read from GIMP at call time - Render, measure, snapshot and compare so the agent verifies each edit before moving on - Non-destructive layer effects, selections, masks, text, vector paths and AI subject cut-outs - Works inside the GIMP window you already have open: Filters, Development, Start Agent Bridge Install: claude mcp add gimp -- uvx gimp-agent-mcp serve # jobscout-mcp https://mcprack.dev/servers/jobscout-mcp/ A multi-source job discovery engine exposed as MCP tools. It searches several boards, normalises each listing into one record, deduplicates across sources on a shared key space, and keeps provenance so an agent can say where a role came from. It deliberately stops at discovery: the judgement about whether a role fits belongs to the person, or to a private layer that never leaves their machine. Status: beta Transport: stdio License: Apache-2.0 Source: https://github.com/SarutobiSasuke8/jobscout-mcp - Several providers behind one normalised record - Deterministic deduplication with provenance on every listing - AI and Web3 taxonomy for the lanes the author actually works in - Listed in the official MCP Registry and published on npm Install: claude mcp add jobscout -- npx -y @sarutobi-sasuke/jobscout-mcp # website-content-mcp https://mcprack.dev/servers/website-content-mcp/ Exposes a website's content to agents as structured, read-only tools: page inventory, headings, body text and metadata, scoped to sites the operator owns. It is the open-source foundation for managed web intelligence work, and runs live as read-only endpoints for the author's own sites. Status: beta Transport: stdio License: MIT Source: https://github.com/SarutobiSasuke8/website-content-mcp - Read-only by design, scoped per site - Structured page, heading and metadata output rather than raw HTML - Published on npm; runs as hosted endpoints for owned sites Install: claude mcp add website-content -- npx -y @sarutobi-sasuke/website-content-mcp # agent-handoff-mcp https://mcprack.dev/servers/agent-handoff-mcp/ A vendor-neutral way for one agent to hand work to another with a record of what was handed over, to whom, and what happened next. Handoffs are bounded messages with lifecycle state; a handoff can reference a repository or vault artefact without gaining any authority to read or change it. Status: alpha Transport: stdio License: Apache-2.0 Source: https://github.com/SarutobiSasuke8/agent-handoff-mcp - Lifecycle state and receipts for every handoff - Works across Claude, Codex and any stdio MCP client - Content-free evidence path: identities and artefacts are referenced, never copied Install: Consult source. # credential-broker-mcp https://mcprack.dev/servers/credential-broker-mcp/ A policy-gated broker that lets an agent exercise a credential without holding it. The agent asks for an operation; the broker checks scope and policy, performs the call, logs it, and returns the result. Secrets stay on the operator's side. Status: alpha Transport: stdio License: Apache-2.0 Source: https://github.com/SarutobiSasuke8/credential-broker-mcp - Scoped grants per agent and per operation - Every request audited - Local-first; no secret ever enters an agent's context Install: Consult source. # obsidian-github-mcp https://mcprack.dev/servers/obsidian-github-mcp/ Agents do not write into the vault. They open proposals against the GitHub repository that backs it, within the paths and rules the operator allows, and a human merges. White-label, so any vault-in-GitHub setup can use it. Status: alpha Transport: stdio License: Apache-2.0 Source: https://github.com/SarutobiSasuke8/obsidian-github-mcp - Proposal-first: every change is a reviewable pull request - Path and rule permissions per agent - Works with any Obsidian vault kept in a GitHub repository Install: Consult source. # source-pack-mcp https://mcprack.dev/servers/source-pack-mcp/ Turns a research question into a source pack an agent or a person can check: the facts found, the quotes and numbers with their primary links, and a coverage map of what was and was not found. Built for research that has to survive a second reader. Status: alpha Transport: stdio License: MIT Source: https://github.com/SarutobiSasuke8/source-pack-mcp - Facts, quotes and numbers each carry a primary link - Coverage map shows the gaps as well as the finds - Fetcher hardened against server-side request forgery Install: Consult source. # meeting-context-router https://mcprack.dev/servers/meeting-context-router/ Takes the output of meeting tools (transcripts, summaries, action items) and routes it as proposals into the CRM and knowledge systems that should hold it, rather than writing directly. A bounded stdio MCP surface sits on top so agents can ask what a meeting changed. Status: alpha Transport: stdio License: Apache-2.0 Source: https://github.com/SarutobiSasuke8/meeting-context-router - Proposal-first: nothing is written to a system of record without review - Bounded tool surface for agents - Designed around Fireflies-style meeting exports Install: Consult source. # mcp-dashboard https://mcprack.dev/servers/mcp-dashboard/ A local dashboard that reads the MCP configuration of the agent clients on your machine and shows what each server costs you in RAM and context tokens against how much you actually use it, with working on/off switches. Not an MCP server itself: a control surface for the fleet. Status: preview Transport: local web UI License: MIT Source: https://github.com/SarutobiSasuke8/mcp-dashboard - Reads Claude Code, Codex, Gemini CLI and Cursor configurations - RAM, context tokens and real usage per server - On/off switches that write back safely Install: Consult source. # sports-oracle-teneo-agent https://mcprack.dev/servers/sports-oracle-teneo-agent/ An experimental agent that exposes live sports data through both a REST interface and an MCP transport, built to explore how a data provider looks from inside an agent network. Status: experimental Transport: REST + MCP License: MIT Source: https://github.com/SarutobiSasuke8/sports-oracle-teneo-agent - REST and MCP transports on the same data - Eight sports covered - Experimental; requires a separately configured provider API credential Install: Consult source. ## Curated picks # Asana MCP Maker: Asana https://developers.asana.com/docs/using-asanas-model-control-protocol-mcp-server Projects and tasks as tools, so status updates write themselves. Use: Closing the loop on tasks after a working session # Canva MCP Maker: Canva https://www.canva.dev/docs/mcp/ Designs, brand templates and exports as tools. Social cards without opening the editor. Use: Daily social assets from a brand template # Figma MCP Server Maker: Figma https://developers.figma.com/docs/figma-mcp-server/ Design context, variables and screenshots straight from the file. Design to code in both directions. Use: Reading a design system before building a page # Fireflies MCP Maker: Fireflies.ai https://docs.fireflies.ai/getting-started/mcp-configuration Meeting transcripts, summaries and action items as tools. Every call becomes searchable context. Use: Pulling decisions and owners out of a call transcript # HubSpot MCP Maker: HubSpot https://developers.hubspot.com/mcp CRM objects, campaigns and content analytics as tools. The sales side of the rack. Use: Pipeline questions answered from the CRM, not from memory # Notion MCP Maker: Notion https://developers.notion.com/docs/mcp Pages and databases as tools. Where durable notes go when they leave the chat. Use: Turning session output into pages a team can find # PixelLab MCP Maker: PixelLab https://github.com/pixellab-code/pixellab-mcp Pixel-art characters, tilesets and animations generated as tools. Game assets without leaving the editor. Use: Sprites and tilesets for small game builds # TinyFish Maker: TinyFish https://docs.tinyfish.ai/ Web search, page fetch and browser automation for agents. The eyes on the open web. Use: Current facts and page content before answering ## Journal # Why a smaller MCP rack can be more useful 2026-09-07 A useful collection helps you make a decision. Finding another hundred servers is rarely the hard part. Knowing which one fits a specific job, what it can touch, and where its limits sit takes more care. ## Start with a job For each candidate, write the task in one sentence. Read design variables from a Figma file. Find job listings with their original sources. Edit an image and inspect the result. If the task is still vague, the selection is too early. Then look for the smallest tool surface that serves that job. A long tool list can mean flexibility, but it does not tell you whether the tools behave predictably in your client. ## Make the recommendation inspectable A listing should give you a reason to investigate and a direct route to the maker's documentation. Project status, transport, package source and limitations matter more than a badge with an unexplained score. MCP Rack separates projects built by its owner from the curated rack of other makers' tools. That distinction is intentional: authorship and personal use are different kinds of evidence. Neither is an independent security audit. ## Curation creates a maintenance obligation A smaller list still needs attention. A recommendation can become stale when access requirements change, a package moves, or a feature disappears. It is better to narrow a claim or remove a recommendation than preserve the appearance of completeness. There is already an [official MCP Registry](https://modelcontextprotocol.io/registry/about) for broad discovery and standardized server metadata. An independent collection earns its place by adding context to a choice. ## Keep commercial exposure distinct If paid placements are introduced, they should be visibly disclosed and kept separate from the decision to recommend a server. Paying for a page should never buy an assurance that the software is safe or suitable. The aim here is a collection you can understand, inspect and return to. Read the [listing standards](/standards/) or explore the [curated rack](/rack/). # Before you connect an MCP server 2026-09-07 Connecting a server gives your assistant another way to read information or take action. Before making it part of daily work, get clear on what runs, where it runs, and what it can access. ## 1. Start at the source Follow the maker's documentation from the listing. Check the package name and repository owner carefully. Similar names are not evidence that two packages have the same maintainer. Record the version you tested so the result remains useful later. ## 2. Understand the connection With stdio, the client launches a local process and exchanges messages through its standard input and output. The runtime, dependencies and environment are part of the setup. A local process can still make network requests or access files according to its permissions. A remote HTTP server runs elsewhere. Its URL, authentication and data handling matter. Use the connection method documented by the server and supported by your client. The [MCP transport documentation](https://modelcontextprotocol.io/specification/latest/basic/transports) is the primary reference for the protocol details. ## 3. Scope access to the job Use a test folder, a test account or a restricted token where the provider supports it. Identify which tools only read and which can write, publish or delete. Do not infer that every tool is read-only because the first example was a search. Keep credentials in the client's supported secret or environment configuration. Avoid putting them in chat transcripts, shared screenshots or public issue reports. ## 4. Run one small proof Pick a harmless task with a result you can check independently. Read a known document. Fetch a page whose content you can compare. Render an image into a disposable output file. Check both the result and any side effects. For a tool that writes, confirm the approval behavior and recovery path before using important data. A successful connection only proves the connection works; it does not prove every operation is appropriate. ## 5. Keep the useful context Write down the client, version, permissions, successful task and any limitation you found. That short record makes an eventual upgrade much easier to assess. Explore the [server collection](/servers/) for setup links. For automated discovery of this site, use the [agent guide](/agents/). # gimp-agent-mcp 0.2: the whole of GIMP 3, for AI agents 2026-09-05 Every colour the agent asked for came out lighter than it asked. `#2b2f5a` rendered as `#7277a0`. The agent could not tell, because until 0.2 it had no way to read a pixel back. That bug, and the fact that the tool could now catch it, is the whole point of this release. ## What it is gimp-agent-mcp is an MCP server that hands an AI agent the whole of GIMP 3 rather than a hand-picked slice. GIMP's Procedure Database has roughly 1,000 procedures and GEGL has 207 filters. Instead of wrapping a few dozen by hand, the server reads them from GIMP at call time: `gimp_pdb_search`, `gimp_pdb_describe`, `gimp_pdb_call`, and the same three verbs for filters. Argument types, enum choices and defaults come from the running GIMP, so nothing goes stale when GIMP updates. It runs on Windows first, because CPython on Windows has no Unix sockets and every earlier GIMP server assumed it did. The bridge is a GIMP plug-in listening on loopback TCP with a per-install token. macOS and Linux paths are implemented but not yet exercised on real machines. ## What 0.2 added 0.1 could open, edit and export. 0.2 is the detailed-work release: - **Sight.** `gimp_render` returns a PNG of the current state: the whole image, one layer, or a region at full resolution. `gimp_snapshot` and `gimp_render_compare` give before, after and a pixel diff side by side. - **Measurement.** `gimp_measure` reads the colour at a point, the bounding box of visible pixels, a per-channel histogram and the dominant colours. The agent stops guessing from a downscaled render. - **Selections and masks.** Rectangle, ellipse, by colour, by alpha, from a path, grow, shrink, feather, border. Layer masks from any of those, plus raw mask pixel writes. - **Text and paths.** Text layers with real fonts, and vector paths that can be stroked, filled or turned into selections. - **Non-destructive effects.** GEGL filters can be appended as GIMP 3 layer effects and edited afterwards with `gimp_layer_effect`. - **AI cut-outs.** `gimp_remove_background` runs a segmentation model and writes the result back as an editable layer mask. - **Recipes.** Repeatable jobs written once as Python that runs inside GIMP: Telegram stickers, web-optimised exports to a size budget, icon sets, watermarks, contact sheets, sprite-sheet slicing, and `compose`, which builds a card or banner from a layout manifest. `gimp_batch_recipe` runs one over a folder. 33 tools and 8 recipes in total. A CI job installs the real GIMP 3.2.4 on a Windows runner and runs 23 live checks on every push. ## The bug The repo banner was built through the bridge itself, and it came out wrong. GEGL's `Color.set_rgba` takes linear RGB, and the server was passing sRGB hex values straight in. Every colour rendered lighter than asked. 0.2.1 routes colours through GEGL's sRGB string parser and converts reported colours back; the smoke test now asserts an exact `#2b2f5a` round trip. Two lessons. The product's own output is the best test fixture. And an agent that can measure a pixel finds this class of bug in one call, where an agent that only sees a render does not find it at all. ## 0.2.2 to 0.2.4, same day - `gimp_help(topic)`: in-band documentation with worked examples, because tool descriptions say what a tool does and not how to combine six of them. - The bridge serves several clients at once with no threads in the plug-in process. Threads inside GIMP crashed PyGObject intermittently; sockets are now served by GLib IO watches on the plug-in's main loop. - Agents work in the GIMP window you already have open: **Filters, Development, Start Agent Bridge**. Every agent edit lands in your layer stack as an undoable step. - 0.2.4 publishes the server to the official MCP Registry from the release workflow, with GitHub OIDC and no stored secrets. ## Install ```bash claude mcp add gimp -- uvx gimp-agent-mcp serve ``` Then `uv run gimp-agent-mcp install-plugin` once, restart GIMP, and start the bridge from the Filters menu or let the agent launch a headless instance. Requires GIMP 3.0 or newer and Python 3.11+. ## What is next 0.3 is macOS and Linux verification (native, Flatpak, Snap), layer-effect reordering, guided mask refinement, a render overlay that draws selection bounds and measurement points so the agent can see coordinates, and recipe parameters exposed as JSON Schema so clients can render forms. The roadmap is in the repository. # MCP Rack: one page for every server I have shipped 2026-09-05 Ten MCP servers, two package registries, one official registry, and no single place that said what each one was for. Searching my own handle on GitHub returned the list without the context. That is the gap this site closes. ## What is here Every server I have published has a page: the package and version, the transport, the licence, whether it is in the official MCP Registry, and a one-line install for Claude Code that any stdio client can reuse. The board on the front page is sortable and searchable. Release notes are written here first and go out on the RSS feed. The rack is the short list of servers built by other people that I run alongside mine. One card each, with what I use it for. A server gets a card by being wired into my tools; it loses the card when it stops being used. Eight today. ## What is not here A directory. The official MCP Registry is the authority for what exists, and the aggregators that read it (Glama, PulseMCP, mcp.so, LobeHub) already compete on coverage. Publishing to the registry is the discoverability move that matters; a curated page on top of it is worth more than a bigger list. ## For agents The same facts are machine-readable. `servers.json` carries the board with install commands. `llms.txt` summarises the site. Every page has structured data. If an assistant is asked what I have published, it has a source to cite. ## How it is built Astro, static, no client framework, no tracking. Fonts are self-hosted. The build fails on em dashes and marketing filler, the same gate astraeus.ie runs. Source is on GitHub under MIT, and the next release note will be the next time a server changes.