Turn most website into a command line interface, so AI agents like Claude Code, Codex, and Antigravity can browse without burning tokens on raw HTML or screenshots.
A typical page is tens of thousands of tokens of markup. The signal on it fits in a few hundred. only-cli fetches the page, distills it into a compact text view with numbered actions, and lets an agent drive the site by number:
```
$ oc open news.ycombinator.com
# Hacker News
[1] Show HN: I built a tiny CSV toolkit
[2] 312 comments
...
actions: do <n> | raw <url>
$ oc do 1
```
No per-site adapters required, no browser extension, no daemon. One generic distillation engine, three runtime dependencies, and a hard token budget on everything it prints.
Requires Node 20+. Requests go through [impers](https://github.com/lexiforest/impers) impersonating Chrome; if impers is unavailable the tool falls back to native fetch.
## For AI agents
The fastest setup is one line in your agent's instructions file (CLAUDE.md, AGENTS.md, or equivalent):
> When you need content from a web page, run `npx @only-cli/oc open <url>` instead of fetching raw HTML. Run `npx @only-cli/oc --help` once to learn the commands.
Claude Code users can install the skill instead: copy `skills/only-cli/` into your project's `.claude/skills/` directory (or `~/.claude/skills/` to enable it everywhere). A skill costs almost no tokens until the agent actually invokes it, which fits how this whole project thinks. The same skill also installs through the [skills.sh](https://skills.sh) directory into Claude Code, Cursor, Codex, Copilot, and others:
No setup at all also works: `npx @only-cli/oc` runs without a global install, and the tool teaches its own command surface through `--help`, the `actions:` line at the bottom of every render, and error messages that name the next command to run.
oc open <url> fetch and render a page with numbered actions
oc raw <url> distilled markdown of the whole page
oc do <n> activate a numbered element (v0.2)
oc fill <n> <text> type into a numbered input (v0.2)
oc submit [n] submit a form (v0.2)
```
Flags: `--budget <tokens>` (default 500), `--json`, `--html` (raw as cleaned HTML instead of markdown), `--verbose`/`-v` (metrics on stderr: tokens saved vs the page HTML, HTTP status and client identity, timing, transfer size, memory; alias `--stats`, or export `OC_VERBOSE=1`). Metrics are off by default because they cost tokens too; agents should pass `--verbose` only when running verbosely.
only-cli works on any mostly-static website with no per-site setup: news sites, blogs, documentation, forums, search engines. One generic engine distills whatever HTML comes back. On top of that, `clis/` ships tuned command shortcuts for:
The engine also renders Atom and RSS feeds as regular pages. That is how Stack Overflow works: the site serves every HTML page a Cloudflare challenge, but publishes full question and answer bodies under `/feeds`, so `oc open stackoverflow.com/feeds/question/11227809` returns the question and its top answers in about 500 tokens. The same trick applies to any site that gates its pages but leaves its feeds open.
Not supported yet: pages that only render with JavaScript (a headless fallback is planned for v0.3), sites behind logins (sessions land in v0.2), and sites with hard bot challenges that do not expose feeds. Adding a site shortcut is a small JSON file; see [CONTRIBUTING.md](CONTRIBUTING.md).
Measured against live sites in [only-cli/benchmarks](https://github.com/only-cli/benchmarks) (only-cli 0.2.0-beta.1, 2026-08-18). What each method hands the agent per page view, across six real tasks (an article, a news front page, a Reddit discussion, a web search, a GitHub repository search, a LinkedIn company page):
The compact view reads all six pages for less than half the tokens of its cheapest rival, a single-screenshot floor, and 92x fewer than raw HTML. The nearest rivals are floors, not full reads: the computer-use rows price a single 1024x768 screenshot, one look at the top of the page, and Browser Use's state message drops most page text. Among methods that deliver the page content, the gap is 8x to Jina Reader and 13x to Playwright MCP's accessibility snapshot. oc was also the only cleaner to return real content on all six tasks: lynx and the naive fetcher hit a DuckDuckGo challenge, and Reddit served Jina its block page.
The end-to-end agent benchmark runs Claude Code headless (`claude -p` on `claude-sonnet-5`) on three live tasks, one web tool per session, and reads success, turns, tokens, and cost from its JSON output:
Each session gets a skill documenting its tool, so every condition runs at its best. oc finished all three tasks at the lowest cost of any full-success condition, and oc and lynx were the only tools whose answers were real content on every task: Jina Reader and Playwright MCP answered the Reddit task by reporting that Reddit blocks them, while raw curl burned its whole 13-turn budget there, roughly 400k tokens and twenty cents, and returned nothing. Totals include Claude Code's own per-session overhead, so compare rows, not absolutes. The benchmark repo has per-task numbers, methodology, and instructions for adding other tools and models.
The same three tasks through OpenAI's Codex CLI (`codex exec` on gpt-5.6-sol) tell the same story with sharper edges: oc finished 3/3 on 110,551 tokens and 8 tool calls, while raw curl needed 46 tool calls and 978,663 tokens, 872k of them ground out on the Reddit task alone, and Jina Reader and Playwright MCP again answered Reddit by reporting the block. Codex's session overhead is much smaller than Claude Code's, so its totals are not comparable to the rows above; the full codex table is in the benchmark repo.
Early. v0.1 covers static pages, budget-aware rendering, and offline tests. Sessions and actions land in v0.2, a lazy headless fallback for script-heavy pages in v0.3. The design principles and how to contribute are in [CONTRIBUTING.md](CONTRIBUTING.md).
