FELO CONNECT
Wiring live search into your coding agent
Working in Codex or Claude Code and need to check something current? Until now that meant opening a browser tab, searching, and pasting the answer back in. Felo's open-source felo-skills hands that loop to the agent itself — search, research, and slide generation, all behind one API key.
The Gap
Search always lived outside the agent
Codex and Claude Code are strong coding assistants, but by default neither can reach information newer than its training cutoff. Checking a library's latest version, a pricing change, or a documentation update meant stopping, switching to a browser, searching, then pasting the result back into the agent — a round trip that broke concentration every time.
Felo Inc.'s open-source felo-skills repository is built to remove exactly that round trip. Released under the MIT license, it ships 12 skills — including real-time search and slide generation — across four agent environments: Claude Code, OpenAI Codex, OpenClaw, and Hermes Agent. That's the substance behind what Felo markets as "Felo Connect."
By The Numbers
felo-skills, by the numbers
How It Works
It's not an MCP server —
it's a skill-triggered script
Rather than a tool that stays resident in the agent, a bash/node script fires on trigger keywords and calls Felo's REST API directly.
The published felo-search skill definition shows the implementation is simple at its core: a bash script at ~/.agents/skills/felo-search/scripts/search.sh posts JSON to https://openapi.felo.ai/v2/chat and receives back an answer and a query_analysis. Authentication is a bearer token via the FELO_API_KEY environment variable. In other words, Felo Connect isn't a new MCP server — it's a thin wrapper that calls an existing REST API through each agent's own skill/plugin mechanism (Claude Code Skills, ClawHub, and similar).
The slide-generation skill, felo-slides, is a step more involved: it creates a task with POST /v2/ppts, then polls GET /v2/tasks/{task_id}/historical at a default 10-second interval for up to 1,800 seconds (30 minutes), returning a ppt_url and a live-doc URL on completion. If a run times out, the --task-id flag resumes the same task instead of creating a duplicate.
Get an API key
Create a Felo account and issue an API key from the settings page.
Install the skill
Run npm install -g felo-ai, or copy the skill folder from the repo into your agent's skills directory (e.g. ~/.claude/skills/).
Set the key
Run felo config set FELO_API_KEY <key> or set the environment variable; it's stored in ~/.felo/config.json.
Just ask
A prompt like "look up the latest X" triggers the keyword match, and the skill fires automatically and returns the result in the conversation.
Who Benefits
Who this actually helps
Engineers
Check a library's latest version or a pricing change without leaving Codex or Claude Code. Fewer context switches, fewer broken trains of thought.
PMs and planners
Hand felo-slides a research prompt, a URL, or internal docs and get a first-draft deck back. The first pass of a briefing document shows up faster.
Both
Tools like felo-x-search let you check social reaction or competitor moves without leaving the coding agent's conversation — research to deliverable in one session.
Landscape
How it differs from search-only APIs
The market for agent-facing search APIs is already competitive. Exa prices around $7 per 1,000 requests and leads on complex queries; Tavily, acquired by Nebius in February 2026, remains a straightforward basic-search API; Perplexity and others are now building search directly into their inference APIs.
Where most of these compete on search accuracy and speed alone, Felo differentiates by bundling search, slide generation, mind maps, X search, and page generation behind one API key. For someone working inside a coding agent, the practical value isn't just getting a search result — it's turning that result into a deliverable without leaving the same session.
| Before | With Felo Connect |
|---|---|
| Fresh info requires a separate browser tab and manual copy-paste | felo-search searches and summarizes inside the agent |
| Findings get typed up by hand into Slack or a doc | felo-slides auto-generates a PPT, polling for up to 30 minutes |
| Each agent has its own, inconsistent way to search | The same 12 skills work across Codex, Claude Code, OpenClaw, Hermes |
| Compared feature-by-feature against Exa/Tavily on search alone | Search, slides, X intel, and pages bundled under one key |
A coding agent no longer needs
to open another tab.
Caveats
What optimism glosses over
Only the skills layer — the calling scripts — is open source. The core search, summarization, and slide-generation processing runs on Felo's own, closed openapi.felo.ai backend. That means part of what you ask your coding agent, and any attached material, gets sent to Felo's servers, so teams need to decide in advance what's acceptable to send before rolling this out internally.
The implementation is also a skill/plugin, not an MCP server, so toolchains that only support MCP can't use it as-is without a plugin/skill-compatible layer such as Claude Code Skills or ClawHub. If your stack assumes MCP, it's worth comparing against an MCP-native search server such as Exa's. And openapi.felo.ai's own pricing doesn't appear to be published anywhere we could confirm — there's no comparable number yet to set against Exa's roughly $7 per 1,000 requests.
Next
What to do next
The safe first step is trying felo-search alone on a personal account, to gauge free-tier limits and answer quality. Before wider rollout, scope the API key narrowly and check what data it's allowed to send against Felo's terms of use. Teams already committed to an MCP-only toolchain should weigh existing MCP-native search options like Exa against what felo-slides and Felo's other output-generating skills specifically add.