iloveAgents
An open-source browser platform running 160+ ready-to-use AI agents: bring your own key, no sign-up, with multi-provider battles, chained workflows, and scheduled automations.
- Source repo
- AditthyaSS/iloveAgents
- Stars
- ★ 150
- Last updated
- 4d ago
- License
- MIT
- Primary language
- JavaScript
- FA score
- 57/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A developer choosing between LLMs uses Battle Mode to fire one prompt at GPT-4o, Claude Sonnet, and Gemini Flash simultaneously and judge outputs side by side.
- Not a fit if
- Teams that need a platform to custody LLM keys for them — interactive runs require your own API keys
- Users who want Workflows or Scheduled Automations without a Supabase project, which those features require
- Source review
- 57/100 · Major gaps
What does this agent do, and when should you use it?
iloveAgents is an open-source React web platform (React 18 + Vite 6 + Tailwind CSS 3) hosting 160+ AI agents across engineering, education, marketing, healthcare, security and more, usable directly at iloveagents.vercel.app or self-deployed to Vercel. It uses a BYOK model: for standard interactive runs, API calls go straight from your browser to OpenAI, Anthropic, or Google Gemini, and the key never touches a server. The platform is config-driven — a registry (src/agents/registry.js) auto-collects definition files in src/agents/definitions/ via import.meta.glob, and a single llmAdapter.js unifies all providers behind one runAgent() interface, so adding an agent means dropping in one JavaScript object. Beyond individual agents it offers Battle Mode (same prompt fired simultaneously to GPT-4o, Claude Sonnet, and Gemini Flash for side-by-side comparison), an AI Workflow Builder (up to 5 chained agents where each step's output feeds the next), and Scheduled Automations (hourly/daily/weekly runs with keys encrypted via Supabase pgsodium and email alerts via Resend). It is MIT-licensed, community-built under GSSoC 2026, and contributing a new agent takes about 5 minutes.
The platform runs AI agents in the browser: you pick a provider and paste an API key in ApiKeyBar.jsx, AgentRunner.jsx builds the input form and prompt from the agent config, calls the unified runAgent() function in llmAdapter.js to send requests directly to OpenAI (GPT-4o, GPT-4o-mini), Anthropic (Claude Opus, Claude Sonnet), or Google Gemini (Gemini 2.5 Flash), and OutputRenderer.jsx renders responses as markdown/text/JSON, with ScorecardOutput.jsx providing a visual scorecard for JSON output. Battle Mode (/battle) fires the same prompt to all three providers at once, loading each column independently so one failure doesn't break the others, and lets you pick a winner. The Workflow Builder (/workflows routes) lets you chain up to 5 agents sequentially via WorkflowRunner.jsx with per-step status (waiting → running → done/failed), stopping at failures with a retry button; workflows publish to a community library with live usage counters. Scheduled Automations (/automations) schedule any agent on an hourly/daily/weekly cadence: with explicit opt-in, the key is encrypted at rest in the Supabase user_secrets table via pgsodium, Vercel Cron (/api/cron/tick) triggers execution, runs and error logs are recorded in automation_runs, reports are emailed through Resend, and Supabase RLS restricts access to the authenticated owner.
- A developer choosing between LLMs uses Battle Mode to fire one prompt at GPT-4o, Claude Sonnet, and Gemini Flash simultaneously and judge outputs side by side.
- A content creator chains a research agent → PDF summarizer → LinkedIn post writer in the Workflow Builder and gets a finished post from a single input.
- An operations person sets up a Scheduled Automation for competitor monitoring or weekly reports, with results and outputs emailed automatically via Resend.
- A user who needs a one-off task — meeting-notes summary, code review, SQL generation — opens the site, pastes an API key, and runs the matching agent with no sign-up or local install.
- An open-source contributor follows the 3-step guide to add a new domain-specific agent by creating one JS file in src/agents/definitions/ and opening a PR.
How do you install or deploy this agent?
Standard path is clone and run locally (Node.js 18+, npm or yarn):
git clone https://github.com/AditthyaSS/iloveAgents.git
cd iloveAgents
npm install
cp .env.example .env.local
npm run dev.env.local is for Supabase-backed features (Workflows, Scheduled Automations):
VITE_SUPABASE_URL=your_supabase_url
VITE_SUPABASE_ANON_KEY=your_supabase_anon_keyAI provider API keys are not configured in .env; they are entered at runtime in the UI and never stored. Alternatively, fork the repo and import into Vercel (or any static host) — the included vercel.json handles SPA routing; set VITE_SUPABASE_URL and VITE_SUPABASE_ANON_KEY in the host environment to enable Workflows and Automations.
How do you use this agent?
- Start the dev server and open http://localhost:5173 (or the hosted demo at iloveagents.vercel.app).
- In the API Key bar, select a provider (OpenAI / Anthropic / Gemini — switchable at runtime, no restart) and paste your runtime key; for standard runs it stays entirely in your browser.
- Pick an agent from the homepage grid, fill its input form, and run; output renders as markdown/text/JSON.
- To compare models, click "Enter Battle Mode" or visit
/battle, enter the prompt once, and all three providers generate simultaneously — pick the winner.
- To chain agents, visit
/workflows/build, arrange up to 5 agents, and save; run at/workflows/:id/runwith one input, retrying any failed step.
- For recurring tasks, visit
/automations, pick an agent, configure inputs, and set an hourly/daily/weekly schedule; note the server cron ticks once daily, so hourly fidelity requires keeping a browser tab open with the in-browser heartbeat.
What are this agent's strengths and limitations?
- Privacy-first BYOK architecture: interactive runs send keys and requests directly from the browser to the provider — no middleman server, storage, or tracking.
- Genuine multi-provider support: OpenAI, Anthropic, and Gemini are switchable at runtime, and Battle Mode compares all three side by side on the same prompt.
- Config-driven, near-zero extension cost: adding an agent is one JS definition file in src/agents/definitions/, auto-collected by the registry — the README says about 5 minutes per contribution.
- Complete feature gradient: single-agent runs, up-to-5-step workflow pipelines, and scheduled automations with encrypted key storage and email notifications cover increasing automation depth.
- Workflow Builder and Scheduled Automations hard-require Supabase (database plus RLS setup), so a pure static deployment cannot use them.
- The server cron ticks only once daily, so hourly schedule fidelity requires leaving a browser tab open with the in-browser heartbeat — weakening unattended reliability.
- You must supply paid model API keys (GPT-4o, Claude, Gemini); there is no free/local model path, so costs scale with usage.
- Server-side automations require explicit consent to store your API key encrypted (pgsodium) in Supabase — a trust hurdle for key-sensitive users.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| iloveAgents This agent | 57 · Major gaps | Web appFree + model costs | ★ 150 | 4d ago | JavaScript | OpenAI API · Claude API |
| Okou | 56 · Major gaps | Hosted serviceFreemium | ★ 1.2k | today | TypeScript | OpenAI API · Claude API |
| COMPASS Skills | 85 · Good | CLIFree | ★ 762 | 6d ago | Python | Codex · Claude Code |
| Skales | 36 · Major gaps | Desktop appFreemium | ★ 2.2k | 2d ago | — | ChatGPT · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
README and SECURITY.md detail a BYOK model: keys stay in the browser, direct provider calls, explicit opt-in plus pgsodium encryption for server-side storage, and documented external effects (Resend email, Vercel cron). But these are claims, not verifiable code — llmAdapter.js and useApiKey.js are referenced yet absent from the evidence. All dependencies use floating caret ranges with no lockfile or audit guidance, so dependency_security is thin; there is no key-rotation/recovery path for stored automation keys (rollback deduction); maintainer attribution is an unverified personal claim.
A thorough troubleshooting guide (401/429/timeouts) and workflow retry-with-diagnostics support failure_messages; CI runs tests and builds on conventional npm dependencies. However, the docs contradict themselves: Repository Highlights claims OpenRouter support while the provider table lists only three providers, and the README says 'No .env files' then requires .env.local — self_consistency is deducted.
The browser-based, multi-provider, backend-optional scenario is clearly described and categories are broad. But the '160+ agents' list points to AGENTS.md, which is not in the evidence, so capability boundaries cannot be verified; no agent definition source is provided, so trigger precision rests only on a config example; prerequisites say Node v18+ while CI pins Node 22 — an environment-fit inconsistency.
Install documentation is excellent (quick start, env vars, troubleshooting, deploy guide) and the LICENSE is full MIT text, earning full marks for install_notes and license. But there is no CHANGELOG and the version is static at 1.0.0; no known-limitations section beyond one aside about cron frequency; CONTRIBUTING.md, MAINTAINERS.md, and AGENTS.md are all referenced but absent from the evidence, so information architecture and maintenance responsibility score only on references.
Output rendering is specified (markdown/text/JSON, scorecard component), and Battle Mode, workflows, and automations provide some differentiation; BYOK in-browser mode removes deployment cost. Yet no value claim is backed by execution evidence and agent quality cannot be confirmed statically, so marginal_value and cost_benefit sit at a middling level.
SECURITY.md names the two files that handle keys, and README's architecture is file-level specific; facts are mostly separated from inference. But key claims (no key persistence, no tracking, 160+ agents, pgsodium encryption, RLS) point to source and database definitions not present in the evidence, corroboration across sources is impossible, and contradictions like OpenRouter remain — hence low scores.
- Key security claims (no key persistence, no tracking, pgsodium encryption, RLS) cannot be verified from the provided evidence; audit llmAdapter.js and useApiKey.js yourself before use.
- Documentation is self-contradictory: Highlights claims OpenRouter support but the provider table omits it; the README says no .env is needed yet requires .env.local.
- Automations store an encrypted copy of your API key server-side — the largest trust leap in this review; confirm the opt-in flow and a key-rotation path before enabling.
- No changelog or known-limitations list, and the version is static at 1.0.0; upgrade at your own risk.