Agent Reach gives tellnova agents internet access

tellnova agents write code, manage git worktrees, and ship pull requests without supervision. What they could not do until now is look things up. Ask one to summarize a YouTube tutorial, pull reactions from X, or read the Reddit thread about the library it is patching, and it hits a wall — paid APIs, login screens, and server IPs that bot filters reject on sight.

Agent Reach solves that side of the job. It is a free, open-source tool (MIT license, Python 3.10+) that installs the most reliable access path for a dozen platforms and registers it as a skill your agent can call. This post walks through wiring it into tellnova's local runner, so every background session on your machine can do its own research.

What Agent Reach installs

Agent Reach is an installer and configuration layer, not another walled-garden API. For each platform it picks a primary backend, keeps fallbacks warm, and health-checks the result — when a platform changes its blocking or API, the maintainers switch paths and your agents never notice.

What works immediately after install, with zero configuration:

• Reading any webpage as clean text through Jina Reader
• YouTube subtitles and search via yt-dlp
• RSS and Atom feeds through feedparser
• Public GitHub repos and search through the gh CLI
• Bilibili search and video detail
• Free semantic web search — Exa, auto-configured over MCP, no API key

A second tier unlocks with a one-time setup, typically exporting a cookie from your browser and handing it to the agent: X search and timelines, Reddit posts and comments, XiaoHongShu, and LinkedIn profiles. Xueqiu quotes and Xiaoyuzhou podcast transcripts configure the same way — you tell the agent "help me set up X" and it walks the steps. Cookies stay on your machine; Agent Reach uploads nothing.

Why background agents need web access

A foreground agent can ask you to open a tab. A tellnova session cannot — it runs detached, in its own git worktree, while you do something else entirely. Every research gap becomes either a blocked task or a guess.

The gaps show up at the worst moments. The agent hits a regression introduced upstream and needs the maintainer's thread about it. It wants the migration notes from a conference talk before rewriting an API. It should read real user complaints before redesigning a flow. Without network reach it improvises; with it, the session works from evidence and cites where each finding came from.

Isolated worktree plus live research is the combination that makes background agents trustworthy for open-ended work.

Install Agent Reach inside tellnova

Agent Reach ships an agent-native install path: you hand your agent one sentence, it fetches the instructions, and runs them with shell access.

Start a new tellnova thread on This machine and paste this as the task:

Install Agent Reach: https://raw.githubusercontent.com/alfredalpino/Agent-Reach-english/main/docs/install.md

The session pip-installs the agent-reach CLI (bundling yt-dlp and feedparser), adds system dependencies such as the gh CLI and mcporter, configures Exa search, and registers a SKILL.md so the capability shows up as a skill. To review every system package before it lands, ask the agent to run the install with the --safe flag instead.

One tellnova-specific detail matters here. tellnova keeps its pi agent configuration in its own directory — ~/.nova/pi-agent — pointed at with PI_CODING_AGENT_DIR, deliberately isolated from any ~/.pi setup you use in a terminal. A skill registered there rides along with every local tellnova session without touching your personal pi config. If the installer asks which agent to configure for, that directory is the answer.

Finish with a health check:

agent-reach doctor

doctor prints a per-channel report: what works with no setup, what still needs a cookie, and the exact fix for anything broken.

Put it to work in a session

With the skill registered, research is just another part of a task. Shapes that work well:

• "Read this X thread about our router bug, extract the repro, and patch it: <link>"
• "Summarize this YouTube talk, then open a PR adding the pattern to our docs: <link>"
• "Search Reddit for reports of this traceback and triage the top three: <error>"
• "Read the last two weeks of this RSS feed and draft notes for anything we missed: <feed>"

The agent reads the skill, calls the right backend, and folds what it finds into the workflow it already has — edits land on an isolated branch, you review the diff, and one click ships a pull request. Nothing about the git side changes; the agent simply arrives with sources.

Maintenance and boundaries

Backends rot. Agent Reach's answer is a one-line update you can hand to any tellnova session:

Update Agent Reach: https://raw.githubusercontent.com/alfredalpino/Agent-Reach-english/main/docs/update.md

Two boundaries are worth knowing up front:

• The install is per machine and lives in tellnova's local runner. E2B cloud threads run in fresh sandboxes and do not inherit it.
• Cookie-backed channels inherit whatever those cookies can do. tellnova sessions execute locally with your user permissions — install on machines you trust, and disconnect channels you stop using.

Start a research-capable session

The setup costs one task. Paste the install sentence into a thread, run agent-reach doctor, and your next background session can cite its sources instead of guessing.

Grab tellnova if you have not yet, and pair it with the provider guides: download the desktop app, use your ChatGPT subscription in tellnova, or stack free tokens with OmniRoute and Codex. The source for both tools is public: tellnova on GitHub and Agent Reach on GitHub.