DeepSeek Harness keeps the agent loop in a composable plugin tree. Felo skills drop into that tree as plain SKILL.md bundles, so your local dsh session can search the open web, pull a page into context, and hand back slides without a browser round-trip.
npx @deepseek-ai/dsh web # Web UI at 127.0.0.1:3080
dsh mounts every capability as a plugin: tools, models, session storage, the Web UI. Skills follow the same idea at a smaller scale — a folder with a SKILL.md that describes when to use it, what to run, and what to return. The harness decides how to compose the agent; the skill decides how one job gets done.
That split matters when the job involves the outside world. A coding agent can read your repository all day and still not know what the library changelog says this week, what a pricing page shows right now, or what a conference talk covered last month. Felo skills close that gap with live search, page extraction, YouTube transcripts, and slide output.
Nothing here needs a custom plugin build. Copy a skill folder into a discovered root, reload the session, and the skill tool can call it like any other capability in the scope tree.
Most people start with one research skill in a project folder, then move the keepers to the user layer and add output skills.
Copy felo-search into <projectRoot>/.dsh/skills and ask a question the repository cannot answer.
Release streams and maintainer threads are where version details land first.
The session already holds the findings; the last step is the artefact.
Felo skills are portable bundles. Install the CLI once for credentials, clone the skill repository, and copy the folders you want into a root that dsh already watches.
The CLI owns authentication for every Felo skill. One key covers search, page fetching, transcripts, and slide generation.
npm install -g felo-ai
felo config set FELO_API_KEY your-api-key-hereEvery skill is a self-contained folder with a SKILL.md at its root, so you can take only what you need.
git clone https://github.com/Felo-Inc/felo-skills.git
cd felo-skillsUse the project root while you are evaluating a skill, then move the keepers to the user layer so every session sees them.
# Project-scoped (recommended first)
mkdir -p .dsh/skills
cp -r felo-skills/felo-search felo-skills/felo-slides .dsh/skills/
# Or user-scoped for every project
mkdir -p ~/.agents/skills
cp -r felo-skills/felo-search ~/.agents/skills/dsh discovers direct children only — a nested path such as skills/research/felo-search will not load. Keep each bundle one level below the root.
The skill provider watches existing roots, so a new bundle usually appears after the next catalog read. Start a fresh session if the skill tool does not list it yet.
List the skills you can call, then use felo-search to compare the
current DeepSeek API pricing tiers and cite the sources you used.Pick by the gap they close. Each one is a folder you can read before you trust it — which matters in a harness that runs with your normal user permissions.
A dsh session can read the repository all day and still not know what the vendor changed this week. Search returns dated answers with sources, so the agent can act on what shipped rather than what the lockfile remembers.
Search for the breaking changes in the SDK we just upgraded and list the ones that touch this module.cp -r felo-skills/felo-search .dsh/skills/Release walkthroughs and conference demos live on video, not in the docs. The transcript arrives as text the session can quote while it edits files.
Pull the transcript from this release stream and list every API the presenter says changed.cp -r felo-skills/felo-youtube-subtitling .dsh/skills/Regressions often surface on X before anyone files an issue. This reads posts and replies, so a debugging session can check whether other people hit the same wall.
Check X for reports about this version's connection handling and summarize what people are hitting.cp -r felo-skills/felo-x-search .dsh/skills/The harness owns the investigation; a deck is how the result leaves the terminal. One ask turns the session's findings into slides the platform team can review.
Turn the upgrade checklist into six slides with one risk per slide.cp -r felo-skills/felo-slides .dsh/skills/A side project needs a mark before it needs a launch. SuperAgent drafts logo directions and product-image variants in the same session that writes the code.
Design three logo directions for this CLI and export the one that still reads at 32 px.cp -r felo-skills/felo-superagent .dsh/skills/Twelve skills, one copy command each. Match the skill to the next job you want dsh to run: research, extraction, slides, transcripts, social monitoring, or a long-running agent workflow.
Add on dsh
Live web research with sources you can cite
cp -r felo-skills/felo-search .dsh/skills/Turn pages into structured, reusable data
cp -r felo-skills/felo-web-fetch .dsh/skills/Answers grounded in your own documents
cp -r felo-skills/felo-livedoc .dsh/skills/Transcripts from talks, demos, and reviews
cp -r felo-skills/felo-youtube-subtitling .dsh/skills/Track posts and replies on X
cp -r felo-skills/felo-x-search .dsh/skills/Generate a deck from a single prompt
cp -r felo-skills/felo-slides .dsh/skills/Turn notes and URLs into a deck outline
cp -r felo-skills/felo-content-to-slides .dsh/skills/Long-running workflows that keep their memory
cp -r felo-skills/felo-superagent .dsh/skills/Capture pages as snapshots and Markdown
cp -r felo-skills/doc-screenshot-agent .dsh/skills/Draft X posts from your own material
cp -r felo-skills/felo-twitter-writer .dsh/skills/Generate a landing page from a brief
cp -r felo-skills/felo-landingpage .dsh/skills/Apple product research before you buy
cp -r felo-skills/apple-buy-advisor .dsh/skills/Swap .dsh/skills/ for ~/.agents/skills/ when the skill should load in every project on this machine.
These are written for a local harness with filesystem access: the skill does the outside-world work, and the harness keeps ownership of the code and the session.
A repository pins an older SDK. The harness can read the lockfile but not the vendor's current migration notes.
Fetch the vendor migration guide, search for known issues in the new major version, and write the upgrade checklist for this repo.A cited checklist that names the files to touch, the config keys that changed, and the sources behind each claim.
A dsh GitHub review session has the diff, but not the upstream advisory or the library discussion that explains the regression.
Check whether this dependency bump lands on a version with an open CVE, then comment with the evidence.A review comment that links the advisory and the release note, instead of a vague 'this looks risky'.
You spent a session comparing vendors or approaches and now need to present the decision to people who will not read raw Markdown.
Summarize the comparison into six slides: options, costs, risks, recommendation.A deck draft generated from the same sources, ready to edit rather than rebuild.
dsh is explicit that the preview moves fast. The same caution applies to the capabilities you add to it.
A skill can instruct the model to run commands. The Felo bundles are Markdown plus small scripts, so review them the way you would review a plugin.
Store the Felo API key in the CLI config, not in a repo-local skill folder that might get committed with your code.
Developer-preview harnesses change quickly. Record the felo-skills commit you copied so a later failure has a known-good starting point.
Start in .dsh/skills so one repository can try a skill without changing every session on the machine.
Clone the repository, copy two folders, and ask your dsh session a question it could not answer from the codebase alone.
Felo AI is an independent skill publisher. DeepSeek Harness is maintained by DeepSeek AI.
Start with a skill, then compose the exact models, live data, knowledge, and output tools your DeepSeek Harness workflow needs — the same key every Felo skill already uses.
Use cost-efficient LLM access when your agent needs custom reasoning or generation.
Give agents current web and social signals instead of relying on stale context.
Ground work in reusable knowledge, then turn it into shareable deliverables.
Extend the workflow from research to visual and office-ready results.