Self-hosted watcher agents

Agents that look far,
so you don't have to.

Describe what to watch in plain language. Lookfar checks pages, feeds and repos on a schedule, tells you only when something real changes, and asks before it does anything on your behalf.

Get started free See how it works

Open source · Bring your own LLM keys · One docker compose up

Why Lookfar

Background AI you can actually leave running.

Alerts that cry wolf

Page monitors fire on every cookie banner and "updated 3 minutes ago". Lookfar judges relevance against your mission and stays quiet unless it matters.

AI that burns tokens on nothing

Most agents re-read the web on every run. Lookfar diffs first, with no model, and only calls one when something actually changed.

Agents you cannot trust alone

Anything with side effects waits in your inbox with the exact payload shown. Nothing is sent, posted or called without a yes.

How it works

Three steps, then forget about it.

  1. 01

    Describe the mission

    Write what to watch and what counts as important, in plain language. Add pages, feeds or GitHub repos as sources.

  2. 02

    Lookfar watches quietly

    On a schedule it fetches, diffs and triages. Unchanged pages cost nothing. Real changes get a short, sourced report.

  3. 03

    You stay in control

    Reports, alerts and questions land in one inbox. Anything outward-facing waits for your approval, from your phone if you like.

No change, no tokens

Three gates between the web and your bill.

Every run passes through cheaper checks before it reaches an expensive model, and every step shows its cost.

  1. 1. Detect No model

    Free

    Fetch, normalise, hash and diff. If nothing changed, the run ends here.

  2. 2. Triage Small model

    Fractions of a cent

    One cheap call decides whether the change matters to your mission.

  3. 3. Act Your main model

    Only when it matters

    A tool loop with a hard budget reads, searches memory and writes the report.

Starter agents

A mission is all it takes.

Competitor pricing

"Check these 4 pricing pages every Monday. Report only real price, plan or limit changes, not copy edits."

Dependency radar

"Watch Go, Postgres and pgx releases. Tell me only about breaking changes or security fixes that affect a Go + Postgres 17 app."

Opportunity scout

"Watch these grant and job boards. Alert me about remote Go roles over $150k or grants for open-source infra."

Story tracker

"Follow this topic across my feeds. Send a short "what changed since last time" each evening."

Features

Built for trust, not for demos.

Live

Self-hosted, your keys

One compose file brings up Postgres, the API, the worker and the web app. Add your own Anthropic, OpenAI or OpenRouter keys, or start on the free default.

Coming soon

Approvals inbox

Reports, alerts, questions and approvals in one place, with push and email notifications. Approvals show the exact recipient and payload.

Coming soon

Budgets and a full run log

Per-run and monthly spend caps. Every run has a step-by-step timeline with tokens, cost and duration for each step.

Coming soon

Memory that sticks

Reports and snapshots are chunked, embedded and searchable, so agents remember what changed last time.

Coming soon

MCP for your other AI tools

Connect Claude, Cursor or ChatGPT to your memory, and let any agent ask you for an approval through the same inbox.

Live

Change detection that ignores noise

Watch pages, RSS feeds and GitHub releases. Text is normalised, cookie banners and "5 minutes ago" are ignored, and every change is stored as a readable diff.

Own your stack

Your server. Your keys. Your memory.

Lookfar runs on a small VPS. Everything agents read and write stays in your own Postgres, and your other AI tools can use it over MCP.

Run it anywhere

git clone https://github.com/codebyaadi/lookfar
cd lookfar
cp .env.example .env     # set BETTER_AUTH_SECRET
docker compose --profile app up -d

Plug in your AI tools

Coming soon
{
  "mcpServers": {
    "lookfar": {
      "url": "https://your-host/mcp",
      "headers": { "Authorization": "Bearer lf_…" }
    }
  }
}
View on GitHub

FAQ

Straight answers.

Is Lookfar ready to use today?

It is early. Accounts, API tokens, and watching pages, feeds and GitHub releases with diffs work now. The agent runtime, approvals, memory and MCP are being built in that order, and every feature on this page is labelled Live or Coming soon.

What will it cost to run?

Lookfar is free and open source. You pay only for your server and for whatever LLM provider you choose. Change detection is free, a small model triages, and per-run and monthly budgets stop spend cleanly.

Which models can I use?

Any OpenAI-compatible provider (Groq, OpenRouter, OpenAI, local Ollama or LM Studio) and Anthropic natively. A free Groq → OpenRouter chain is the zero-config default, though tool-heavy agents work best with a stronger model of your own.

Can an agent do something I did not approve?

No. In v1 every outward action, such as sending an email or calling a webhook, goes through the approvals inbox, and the card shows the exact payload. Fetched web content is treated as untrusted data, and webhooks can only reach endpoints you registered.

Does it work on JavaScript-heavy pages?

Version one fetches pages over plain HTTP and extracts readable text or a CSS selector. Headless-browser sources are on the roadmap as an opt-in.

Is it multi-user?

It is built self-host first for one person or a small team. The first account becomes the admin and sign-up then closes, with invites coming later.

Put an agent on watch.

Create your account, point Lookfar at the pages that matter, and get on with your day.

Get started free