Four Repos That Upgrade Your Claude Setup: anydoc, herdr, DeepSeek Harness and OmniRoute

Four open source GitHub repositories, each fixing one specific thing that goes wrong when you use Claude for real work. None of them is a prompt pack. Each is a tool you install once, and each one solves a problem you have probably already hit: office files that turn to mush, a terminal full of tangled agents, a model you cannot swap, and a usage limit that stops you mid task.

The star counts, licences and figures below were read from the GitHub API and each repository's README on 5 September 2026. They move. Check the repo, not this page, before you quote one.

• firecrawl/anydoc, MIT, over 20,000 stars. Converts Word, PowerPoint, Excel, PDF and more into clean Markdown.
• herdrdev/herdr, Apache 2.0, over 35,000 stars. A terminal workspace that runs your coding agents in their own panes and tells you which one is stuck.
• deepseek-ai/deepseek-harness, MIT, over 212,000 stars. An agent runtime where every part, including the model, is a plugin you can swap.
• diegosouzapw/OmniRoute, MIT, over 61,000 stars. A free AI gateway that reroutes to another model when Claude runs out of credits or errors.

Paste a PowerPoint or a Word document into an AI model and the formatting gets in the way. Tables lose their columns, headings become body text, slide order gets scrambled. The model is reading the junk that wrapped the content, not the content.

anydoc, built by the Firecrawl team in Rust, strips that layer off. It reads Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV and text based PDF, and writes GitHub Flavored Markdown with one consistent output whatever went in. The README's own claim is single digit milliseconds per document, and every format goes through the same document model and the same Markdown serializer, so a table from Excel and a table from Word come out looking identical.

1. 1
   Install the agent skill so Claude Code can call it on its own: npx skills add firecrawl/anydoc
2. 2
   Or use it directly from the command line: npx @firecrawl/anydoc report.docx, or with -o to write a file.
3. 3
   Or import it: npm install @firecrawl/anydoc for Node, pip install firecrawl-anydoc for Python, cargo add anydoc for Rust. There is also a WebAssembly build that runs in the browser so files never leave your machine.

One honest limit. anydoc reads text based PDFs locally but does no OCR. A scanned PDF fails with a NeedsOcr error unless you opt into the hosted Firecrawl Parse route, which sends that document off your machine. If your PDFs are scans, know that before you point it at a folder.

Once you run more than one coding agent at a time, a normal terminal turns into a wall of interleaved output. You cannot tell which agent is waiting on you, which one finished, and which one has been stuck for ten minutes.

herdr is a single Rust binary that gives each agent its own pane inside a persistent workspace. Every pane is marked working, blocked or idle, so the stuck one is visible without scrolling. It is a background server, so you can close the laptop or drop the network and the agents keep running; you reattach later from any terminal, including over SSH. It runs the tools you already use: Claude Code, Codex, Cursor, OpenCode and others. It does not wrap or replace them, it owns their terminals.

1. 1
   Install: curl -fsSL https://herdr.dev/install.sh | sh, or brew install herdr on a Mac.
2. 2
   Start it where the work lives by running herdr in your project folder, then launch your agents in separate panes.
3. 3
   Detach with ctrl+b q. Run herdr again to reattach. The agents never stopped.

The part people miss: herdr exposes a socket API that agents themselves can drive, so one agent can spawn a pane, prompt another agent, and wait until that agent is genuinely blocked. That is the beginning of agent to agent orchestration without writing an orchestrator.

Most agent frameworks bolt you to one model. When it starts giving weak answers you rebuild the whole setup around a different one, and every tool, memory and session you wired up comes with it.

DeepSeek Harness, released by DeepSeek on 13 August 2026, is built on an everything is a plugin architecture. The model adapter is a plugin. So are the tools, the skills, the sessions, the sandboxes, the storage, the scheduler, the agent loop itself and the user interface. Swapping the model means swapping one plugin, and the rest of your assistant stays exactly as you built it.

1. 1
   Try it with one command, no clone: npx @deepseek-ai/dsh web. It opens a local web UI at 127.0.0.1:3080.
2. 2
   Or run from source: clone deepseek-ai/deepseek-harness, then pnpm install, pnpm run build and pnpm dsh web.
3. 3
   Browse community plugins under the dsh-plugin topic on GitHub. Add that topic to your own plugin repo so others can find it.

The README says it plainly: the project is a developer preview and there will be compatibility breaking changes. Build on it knowing that, and read the safety notice before you run it