OpenTabs
Let your AI call real web APIs through your logged-in browser session — no screenshots, no DOM scraping, no API keys or OAuth setup.
The README asserts tools-off-by-default, three permission levels (Off/Ask/Auto), permission reset on plugin update, full audit logging, and opt-out anonymous telemetry, but none of this is backed by code in the repository. The core design lets an AI invoke arbitrary site APIs through the user's authenticated browser session ('if you're logged in, your AI can use it'), which is intrinsically broad; least_privilege is at best thin. SECURITY.md is present with a defined scope (positive), but the contact email [email protected] is coupled to a 'Ralph' project credited in the README, weakening attribution; publisher is unverified. Deducted: all security claims are assertions with no permission-model or audit-log implementation evidence.
package., CI workflows, and README are mutually consistent (Node 22, Chrome, package name, license all align); CI covers unit, E2E, Windows, and docs checks. Dependencies are installable from npm with lockfile-hashed caching. Deducted: no error messages or failure behaviors are shown anywhere in the provided files.
Audience is clear (AI users with MCP clients); environment requirements are explicit (Node 22+, Chrome); CI runs on both Linux and Windows. Deducted: capability boundaries live almost entirely in external docs, with only promotional figures ('100+ plugins, ~2,000 tools') in-repo; the 'point your AI at any website and it discovers APIs' claim has no in-repo discussion of trigger precision or misuse risk.
Full MIT license text consistent across LICENSE, package., and badge — full marks. Directory layout, script naming, and contribution workflow are orderly; install notes are concrete including extension loading steps. Deducted: no CHANGELOG, no maintainer roster, no known-limitations section, examples rely on GIFs and external links; the copyright year '2026-present' is questionable. Unverified publisher itself is not deducted, but maintenance evidence is genuinely thin.
Differentiation is clear — real API calls instead of screenshot/DOM scraping, reusing the browser session to skip key/OAuth setup — giving real marginal value. Deducted: output usability is demonstrated only by GIFs with no in-repo output-format documentation; cost-benefit (install/maintenance overhead of local server + extension + plugin system vs. gains) is unquantified.
Facts and inference are reasonably separated, and the third-party non-affiliation disclaimer is explicit. Deducted: most key claims (audit log, permission reset, AI plugin review, '100+ plugins') point outside the repository and cannot be verified statically; README and package. corroborate on name/license, but security claims lack any second source.
- This tool lets an AI call arbitrary site APIs through your logged-in browser session — effectively handing authenticated account capabilities to a model. Review each plugin's source and required tool permissions before enabling; keep the default Off.
- All security claims in the README (audit log, permission reset, plugin review) lack implementation evidence within the repository; this is a low-confidence static review and should not be read as verified safety.
- Prefer Ask mode over Auto, and be cautious about the 'point your AI at any website to auto-discover APIs' feature triggering unauthorized actions.
- Publisher identity is unverified, and the security contact email is associated with a third-party 'Ralph' project; patch-timeliness commitments (48h/7d) rest on an unverified party.
What does this agent do, and when should you use it?
OpenTabs is an open-source tool (GitHub repo opentabs-dev/opentabs, MIT licensed) built on the idea that browser automation clicks buttons while OpenTabs calls APIs. It consists of three parts: an npm-distributed CLI (@opentabs-dev/cli), a Chrome extension, and a local server; the extension bridges your browser to the server so your AI can call web APIs using your authenticated session. It ships an MCP server that works with any MCP client, with a CLI mode as an alternative. The ecosystem includes 100+ plugins and roughly 2,000 tools covering Slack, Discord, GitHub, Jira, Notion, Figma, AWS, Stripe, and more, plus built-in browser tools for screenshots, clicking, typing, and network capture. Security-wise, every tool starts off, permissions are Off/Ask/Auto per plugin or per tool, permissions reset on plugin updates, everything runs locally with no cloud, and there is a full audit log plus opt-out anonymous telemetry.
The flow: install the CLI and run opentabs start to launch a local server; load the Chrome extension from ~/.opentabs/extension via chrome://extensions/ (Developer mode → Load unpacked); the extension bridges your browser session to the server, letting AI make real API calls to services you're logged into. Install plugins with opentabs plugin install <plugin-name> to get that service's tools immediately, no restart needed. Built-in browser tools work on any tab without a plugin. Developers can scaffold a plugin in one command and publish to npm, or point their AI at any website so it discovers the APIs and builds the plugin for them.
- Developers using an MCP client (e.g. Claude Code) who want their AI to act in Slack or Discord — sending messages, adding reactions — as their logged-in self
- Teams that want AI operating GitHub, Jira, Notion, or Figma via one-command plugin installs that work immediately
- Individual users whose target services have no convenient API keys or OAuth flows, who can reuse their browser login instead
- Users who need AI to screenshot, click, type into, or capture network traffic on any web page for debugging or analysis (built-in browser tools, no plugin required)
- Developers building plugins for niche sites, using the one-command scaffold and publishing to npm for others to install
What are this agent's strengths and limitations?
- Real API calls instead of screenshots and DOM scraping — more reliable and precise
- Reuses your browser login, so no per-service API keys or OAuth setup
- Granular security: everything off by default, three permission levels, permissions reset on plugin updates, AI-reviewed plugin source
- Fully local with no cloud dependency, full audit log, and opt-out anonymous telemetry
- Hard dependency on Chrome and Node.js 22+; it cannot run without a local browser
- Relies on third-party services' private APIs and login sessions — upstream changes or anti-automation measures can break it, and the project is not affiliated with or endorsed by any third-party service
- All tools start off, so every plugin requires review and enabling — an initial setup cost
- Ecosystem quality depends on community plugins; for uncovered services you must build a plugin yourself or have the AI generate one
How do you install or deploy this agent?
Prerequisites: Node.js 22+ and Chrome. Install the CLI: npm install -g @opentabs-dev/cli. Start it: opentabs start. Load the extension: open chrome://extensions/, enable Developer mode, click Load unpacked, and select ~/.opentabs/extension. Install a plugin: opentabs plugin install <plugin-name>. Full walkthrough at opentabs.dev/docs/quick-start.
How do you use this agent?
With the extension connected to the local server, add the MCP server to any MCP client and your AI can invoke around 2,000 tools; use CLI mode if you prefer not to use MCP. Tools execute nothing until you explicitly enable them (Off/Ask/Auto levels, per-plugin or per-tool), and your AI reviews plugin source before you enable one. To contribute: clone the repo, run npm install && npm run build, use npm run dev for development (tsc watch + MCP server + extension) and npm run check for build + type-check + lint + knip + test.
How does this agent compare with similar options?
The README explicitly contrasts traditional browser automation (screenshots, clicking, DOM scraping) with OpenTabs's API-call approach: browser automation clicks buttons, OpenTabs calls APIs. No other specific competitor is named in the source.