Do tools work on both OpenClaw and Hermes?

Yes. The catalog is not framework-specific and the install flow is identical on both. The only difference is that Hermes builds its skill index at gateway start, so a bot may need a restart before it notices a newly installed tool, while OpenClaw re-reads skills each turn.

Are tools installed per account or per bot?

Per bot. Each instance has its own set, so two bots on the same account can carry completely different tools. The instance must be running to install or read status.

What is the difference between a tool and a skill?

A tool is software installed into the bot’s container. A skill is a written procedure the agent follows, with no software involved. Integrations connect third-party SaaS accounts, and MCP is a protocol several tools speak rather than a category.

How do I call a tool once it is installed?

You do not call it by name. Ask for the outcome — convert this file, transcribe this audio, index this repository — and the agent uses the tool it has.

Do installed tools use up my instance disk?

Yes, they are real software in your container. Uninstall what you are not using from the same card if a bot is running short on disk.

← Home

Guide

Tools

A model can reason about a PDF it cannot open. Tools fix that: a catalog of open-source runtimes and MCP servers you install into a specific bot, in one click, so it can actually do the thing. Same catalog on OpenClaw and Hermes.

What a Tool Is Here

Not a plugin for our site — a program that ends up inside your bot’s own container, which the agent can then call while it works. Convert a spreadsheet, transcribe an audio file, index a repository, remember across sessions, cut a video.

Two shapes, and the card tells you which you are getting:

Installed locally. The tool is fetched into your instance at a pinned version, checksum-verified, and runs there, with your bot’s existing file and network permissions. Local execution is not the same as offline: a locally installed tool can still reach the network if that is what it does, and whatever the agent reads may end up in its model requests either way. What you get is that the program runs on your own box rather than someone else’s service.
Connected. The tool lives elsewhere and your bot talks to it over an HTTPS endpoint you supply. You pick the protocol it speaks — REST, OpenAI-compatible or MCP — and, separately, how to authenticate: none, bearer, basic, or an API-key header. Useful when you already run the service, or when it is too heavy to sit beside the agent.

Install One

Go to the dashboard and open Tools in the sidebar.
Choose which bot you are fitting out. Tools are installed per instance, not per account — two bots can have entirely different kit — so the page asks which one, and it needs to be running.
Find the tool and install it. The install runs as a background job.
Then just ask. Once a tool is installed the agent knows it is there; you do not call it by name or learn a syntax. “Turn this spreadsheet into a summary” is the whole interface.

The page is in beta and labelled as such. Installs are pinned to a specific version and verified before they run, rather than pulling whatever is newest, so a tool that worked yesterday behaves the same today.

What Is in the Catalog

It moves as things ship, so browse rather than trust a list — but the categories are stable and tell you what kind of gap each one fills:

Documents. Reading and writing the formats work actually arrives in — Office files, PDFs, conversion between them, OCR for scans. See OfficeCLI and MarkItDown.
Media. Downloading, transcribing and assembling audio and video. See yt-dlp.
Memory. Structured recall that survives a session, separate from the framework’s own memory.
Code. Repository indexing and code-aware agents, so a bot can answer about a codebase instead of a file.
Data and models. Vector stores, search indexes, local model runtimes such as Ollama.
Web. Crawling and search backends — Crawl4AI, SearXNG.

The grid is ordered by how many accounts have actually installed each tool, with GitHub star momentum as the tie-break. That is a deliberate choice: what other people here found useful is a better first signal than what is trending upstream.

OpenClaw and Hermes

The catalog is not framework-specific. Every tool installs on both, and the flow is identical — pick the bot, click install. What differs is only where the files land inside the container, which the installer handles.

One practical difference worth knowing: several tools arrive as a skill the agent reads. OpenClaw re-reads its skills each turn, so a new tool is usable immediately. Hermes builds its skill index when the gateway process starts, so if a freshly installed tool seems invisible, restart that bot from the dashboard and try again.

Removing One

Uninstall from the same card. Worth doing rather than leaving everything installed: each tool is real software in your container taking real disk, and the instance disk is finite — see freeing up disk space if a bot is getting tight.

Tools, Skills, Integrations, MCP

Four words for four different things, and mixing them up is the usual source of confusion:

Tools (this page) install software into your bot’s container.
Skills are written procedures the agent follows — no software, just instructions. See installing skills and Skill Sync.
Integrations connect third-party SaaS accounts such as Gmail or Notion. See Composio.
MCP is a protocol several of these tools speak, not a category of its own. See MCP on OpenClaw and MCP on Hermes.

Troubleshooting

“Could not read this bot’s tool status.” The instance is stopped or asleep. Pick a running one and refresh.
The install does not finish. It runs in the background and fetches real software, so give it a minute. If it fails, retry — the installer is built to be re-run rather than leave a half-state.
Installed, but the bot does not use it. On Hermes, restart the bot so its skill index picks the tool up. On either framework, ask for the outcome (“convert this to a Word document”) rather than naming the binary.
A connected tool never answers. Check the endpoint is HTTPS and reachable from outside your network, and that the auth mode matches what the service expects.

Try It

Open Tools, install OfficeCLI on a running bot, then send it a spreadsheet and ask for a summary as a Word document. It is the fastest way to feel the difference between a model that can describe a file and a bot that can open one.