Felo Skills for OpenClaw 2.0: Install Once, Then Audit What Runs
Add Felo skills to OpenClaw 2.0 for live web answers, X signals, decks, and transcripts — installed from ClawHub, reviewed in Skill Workshop.
OpenClaw is not a chat window. It is a personal assistant that answers in the app you already use — Telegram, WhatsApp, Slack, Discord, or the native desktop and mobile apps — keeps its own memory, and keeps working on a schedule.
That range is the point. The assistant goes to the team instead of asking the team to come to it.
The trouble starts one layer down. A skill you install from someone else runs with the assistant's reach: your group chats, your files, your name on the message. One command, and instructions you have never read are acting on your behalf.
The 2.0 release turned that from an act of faith into a record you can read. Felo skills are the half that makes the record worth opening.
What the 2.0 release changed
v2026.8.1 is the release OpenClaw calls 2.0. Its notes describe skills as following one connected path — creation, validation, discovery, installation, calling — instead of scattering those steps across a folder, a command line, a settings page, and a community thread.
Skill Workshop is where that path leaves a trail. It holds proposals, checks, decisions, and applied history. When the skill came from someone else, that history is the whole argument: you can see what the skill proposes and what check ran before it was applied, rather than inferring its behavior from what it did to you.
| OpenClaw 1.x habits | In the 2.0 release | |
|---|---|---|
| Discovery | A skill found in ClawHub or in a community thread | Curated ClawHub skills sit next to vetted connectors in the plugin catalog |
| Install | A CLI command dropped files in place with limited feedback | Install, enable, disable, and remove run through the same surface, so state is visible |
| Validation | You checked the skill yourself, if at all | Skill Workshop records proposals and checks before a skill is applied |
| Permissioning | Behavior was inferred from what the skill did at run time | Approvals attach to the exact request, command, session, and person |
| Memory | Context lived in the conversation | Sessions and transcripts live in SQLite, and can be searched, imported, or removed |
| Scheduling | Recurring work was a patchwork of names | Scheduled work is consistent across the agent, Control UI, CLI, docs, and native apps |
Read the permissioning row twice. An approval in the 2.0 release is not a yes to a skill in general; it binds to the exact request, the command, the session, and the person. That is a far narrower grant than most assistants ask for, and it is why you can let a skill read a channel all week while still approving every message it sends.
Memory moved in the same direction. Sessions and transcripts living in SQLite means memory can be searched, imported, and removed — inspected like a file rather than absorbed like a mood. That matters for a research skill, because the difference between a useful assistant and a noisy one is usually what it chose to remember.
Two install routes, one review step
There are two official ways to add a Felo skill. Both end with the skill available — and not yet enabled.
The installer script, which handles the full setup in one pass:
bash <(curl -s https://raw.githubusercontent.com/Felo-Inc/felo-skills/main/scripts/openclaw-install.sh)
Or ClawHub, one command per skill, on a machine where the registry is already configured:
clawhub install felo-search
The second step is the one people skip. Open Control UI → Skills / Skill Workshop, read the proposal and the checks for the skill you just added, then enable it for the agents that should use it. Enabling everything globally because the install worked is how a research skill quietly acquires the reach of a messaging skill.
Skills are only useful once there is a delivery surface, so pair the chat app before you judge the result. Then verify from inside it.
Search for the current status of the rail strike and reply here with two sources.
An answer with two sources means the install worked. An answer with none means the skill never registered — better to learn that now than after a week of scheduled runs.
Five Felo skills for a chat-first assistant
Chosen for what people actually ask a personal assistant inside a thread. Each one returns something checkable — a date, a source, a file — rather than a confident paragraph.
Felo Search — answers worth forwarding
Research that returns dated results with sources, so a reply your assistant posts into a group carries the link that backs it. It is the same live web search behind felo.ai/search, delivered where the conversation already is.
clawhub install felo-search
Search for tonight's train options from Tokyo to Osaka and send the times with sources to this chat.
Felo X Search — what people are saying right now
When the question is what developers think today, summarizing an article about what developers think is not an answer. This reaches posts, users, and replies directly.
clawhub install felo-x-search
What are developers saying about the new release today? Pull the top posts and replies.
Felo Slides — the file lands in the thread
A file in the conversation beats a file in a folder. The deck is built from what the assistant already researched, then shared through the channel under your approval rules — and it comes back as an outline before it becomes a file, so the argument gets settled while changing it is still cheap. Same product as felo.ai/slides, pointed at a chat.
clawhub install felo-slides
Turn yesterday's research thread into a ten-slide summary deck I can send to the team.
Felo LiveDoc — answers grounded in your own documents
The bridge between research and delivery. A comparison you assembled last month stays a document the assistant can read, not a thread it has to be reminded about.
clawhub install felo-livedoc
Summarize the terms in the LiveDoc I added last week and flag only what changed.
Felo YouTube Subtitling — video into text
Someone drops a talk in the channel. The transcript comes back as text the assistant can summarize, quote, or turn into a deck.
clawhub install felo-youtube-subtitling
Pull the transcript of this talk and list the three claims that matter for our roadmap.
The full catalog, including Felo SuperAgent for images and Felo Web Fetch for structured page extraction, is on the Felo skills for OpenClaw 2.0 page.
Three recurring jobs worth scheduling
Written the way an OpenClaw owner would actually type them.
The morning brief that arrives before you ask
A short digest in Telegram at 7:30 — weather, a market signal, and anything that changed at three competitors.
Every weekday at 7:30, research these three competitors, flag only what changed, and post the brief here.
What lands back is a scheduled message that references previous runs instead of repeating them, with sources on every change. Scheduling is consistent across the agent, Control UI, CLI, and native apps in the 2.0 release, so the same job is visible from wherever you happen to be. The "only what changed" instruction is not a style preference — a brief that grows longer every week stops being read.
X monitoring with a human gate
The assistant should watch a topic all day. Nothing should be published under your name without a yes.
Watch posts about our launch this week, summarize the recurring complaints daily, and draft replies for my approval.
You get a daily summary and a queue of drafts sitting behind the gate. Because approvals attach to the request, command, session, and person, the boundary is per action rather than per skill — reading is autonomous, sending is not. That is the split worth designing for.
Research that ends in a file, not a paragraph
A colleague asks in Slack for the comparison you researched last week, and they want slides for a meeting.
Turn the vendor comparison in my LiveDoc into a deck and post it in this thread.
A deck built from stored research, shared through the channel with an approval step in front of it. The sources stay attached to the work, which is what lets the file survive the meeting rather than just start it.
Guardrails
Keep approvals attached to the request. 2.0 binds an approval to the exact request, command, session, and person. Skills that send messages or move files should route through that gate, not around it. When an install asks for a standing "always allow", that is the request to decline.
Let credentials stay out of model-visible text. Protected credentials can reach supported destinations without entering the transcript. Store the Felo key in the CLI config and reference it by name — a key pasted into a chat becomes memory, and memory is searchable.
Check what a skill wants before enabling it. Skill Workshop keeps the proposal and check history. Read it once, then decide per agent instead of enabling globally.
Review memory the way you review files. Sessions and transcripts can be searched, inspected, imported, and removed. If a research thread should not become long-term context, keep it in a document where you can see it.
FAQ
Where do I install skills in OpenClaw 2.0?
Use the ClawHub CLI or the official Felo installer script, then review and enable the skill through the Control UI. Admins can browse, enable, disable, and remove plugins from that same surface.
Do the existing Felo OpenClaw skills still work after 2.0?
Yes. The bundles and the ClawHub install path are unchanged; the 2.0 release adds lifecycle and approval tooling around them. If you installed through ClawHub before, update in place.
Can skills run on a schedule?
Yes. Scheduling is unified across the agent, Control UI, CLI, docs, and native apps in the 2.0 release, so a research skill can run as a repeating brief. Keep a manual approval on anything that publishes.
Does Felo publish on ClawHub?
Yes. Felo maintains a ClawHub publisher page and keeps the skill sources in a public repository, so you can compare the installed bundle against the source before you enable it.
What does a Felo skill need to run?
The Felo CLI with an API key stored in ~/.felo/config.json. The assistant calls the skill, the skill calls the CLI, and the key never has to be pasted into a chat.
Install one skill through ClawHub, read what Skill Workshop kept on it, then ask a question that needs current information and see whether the answer is worth forwarding. The second part is the one that changes how much you trust the first.
Explore Felo skills for OpenClaw 2.0 →
This post is also available in 简体中文, 日本語, 한국어, 繁體中文, हिन्दी, Français, العربية, Русский, اردو, Bahasa Indonesia, Deutsch, Tiếng Việt, Türkçe, Italiano, ไทย, Español, বাংলা and Português.