All agents

Website Manager AI

A senior web team in one agent — it connects to your website's GitHub repo, learns your design system and brand voice, then makes any edit or builds any new page from a plain-English instruction, shipped as a reviewable Pull Request.

View in marketplace Open dashboard

The Website Manager AI (agent #16, Marketing) manages your existing website through GitHub. You connect a repo and branch, click Learn site, and it builds a Site Profile — framework, design tokens, component library, page patterns, and brand voice. From then on you describe changes in plain English ("change the homepage hero headline", "create a /pricing page") in the Website workspace at /dashboard/website; it plans which files to edit or what page to create, applies the change consistently with your components and voice, and opens ONE GitHub Pull Request per task. Nothing goes live until you merge — your host's CI/preview build is the final gate. It pairs with the SEO Agent: SEO recommends pages, the Website Manager builds them.

What it does

This agent removes the developer bottleneck from routine website work. Learning is deliberate and bounded: it reads the repo tree, auto-detects a monorepo site subfolder and the framework (Astro, Next.js, and React are detected explicitly), then studies a curated set of up to 12 key files — package/Tailwind/framework config, global styles, a layout, top components, and a few pages — plus a names-only inventory of up to 80 components and 60 pages. From that it produces a structured Site Profile (design tokens, key components with usage notes, the page pattern, brand voice, nav, and a plain-English summary) that is injected into every subsequent task, so it behaves like it already knows your site. One profile is kept per organization and can be re-learned any time.

Execution is a plan-then-apply loop with guardrails. For each instruction it first plans: edit (1–5 existing files, chosen from the real inventories) or create (a new page path plus an existing similar page to use as the template). Edits must change only what was asked — an edit is discarded if the rewritten file comes back empty, unchanged, or suspiciously resized (under 40% or over 3× the original length). New pages reuse your actual Layout and components so they're indistinguishable from existing ones, with SEO structure (meta title/description, a single H1), and can include an on-brand hero image generated with first-party Workers AI (FLUX.1-schnell) saved under public/generated/. Everything is committed to a fresh web/<action>-<timestamp> branch and opened as one Pull Request — never a direct push to your deploy branch — and each task is recorded in a change history with its PR link and status.

How it works

Your Business
1
Connect the repo
In the Website workspace you set the GitHub repo (owner/name), base branch, and optionally your live site URL (for the preview pane). This is saved to your org's site profile; GitHub itself is connected once under /dashboard/tools.
2
Learn the site
Learn site streams progress while learnSite reads the repo tree, detects the monorepo prefix and framework, studies up to 12 curated files (configs, global CSS, layout, components, pages), and has Claude produce a strict-JSON Site Profile: design_tokens, key_components, page_pattern, brand_voice, nav, and a plain-English summary — plus component/page inventories for planning.
3
Instruct in plain English
You type an instruction (or click a page/component in the site browser to prefill context like "On /pricing …"). The task runs in a held-open streaming request (/api/website/task) so multi-file edits and PR creation complete reliably, with live progress.
4
Plan
A planning pass decides action 'edit' (1–5 existing files picked from the real page/component inventories) or 'create' (a new page path, an existing similar page as the template, and whether it needs an image).
5
Apply with guardrails
Edits rewrite only what the instruction asks, preserving all other code — a result is rejected if it comes back empty, unchanged, or under 40% / over 3× the source length. New pages are generated from your template page + Site Profile with SEO structure (meta, single H1), optionally with a Workers AI (FLUX.1-schnell) hero image at public/generated/.
6
Open ONE Pull Request
All files are committed to a new web/<edit|create>-<timestamp> branch and opened as a single PR with the instruction, the changed-file list, and a summary. It never pushes to your base branch — your CI/preview build runs and you merge to ship.
7
Record the change
Every task (kind, instruction, PR URL/number, branch, status) is saved to the change history shown in the workspace, so you always have an audit trail — and Git gives you rollback.
Outcomes delivered

Setting it up — owner / admin

  1. 1
    Subscribe and open the workspace
    Subscribe from the marketplace, then open the Website workspace at /dashboard/website — this agent is driven from its own console, not a run schedule.
  2. 2
    Connect GitHub
    Connect GitHub under /dashboard/tools with repo scope. The agent reads files and opens PRs with that connection; it needs nothing else.
  3. 3
    Point it at your site
    Enter the repo (owner/name), the base branch (e.g. main), and optionally your live site URL so the workspace can show page previews. Monorepos are fine — the site subfolder is auto-detected.
  4. 4
    Click Learn site
    One click builds the Site Profile (framework, design tokens, components, page patterns, brand voice). You'll see the learned summary, brand voice, and component count in the workspace; re-learn any time the site changes significantly.
  5. 5
    Make the first change
    Type something small — "change the homepage hero headline to …" — and watch it stream progress and return a PR link. Review the diff, let your host's preview build run, and merge.
  6. 6
    Pair it with the SEO Agent
    If you run the SEO Agent, its Page ideas hand off with one click: the workspace opens pre-filled with the page title, slug, target keywords, rationale, and suggested sections — the Website Manager builds it on-brand.

Using it day to day — your team

  1. 1
    Browse your site inside the workspace
    The workspace lists your real pages and key components (nav, footer) from the learned inventory. Click one to see its section/link outline and a screenshot preview of the live page.
  2. 2
    Describe the change
    Write plain English — no code. Selecting a page first anchors the instruction ("On /pricing …") so the agent targets the right file.
  3. 3
    Watch it work
    Progress streams live: planning, editing or creating files, generating an image if needed, then opening the Pull Request.
  4. 4
    Review and merge the PR
    Open the PR link, check the diff and your host's preview build, and merge to publish. Nothing changes on the live site until you do.
  5. 5
    Track the history
    The Changes list shows every past task with its instruction, PR link, and status (open/failed), so the whole team can see what was shipped and when.

Use cases

Plain-English edits
"Change the homepage hero headline", "add a testimonial section to /pricing", "update the footer links" — the agent locates the right files (up to 5 per task) and opens a PR with just that change.
Build SEO-recommended pages
The SEO Agent's Page ideas hand off with one click; the workspace opens pre-filled with title, slug, keywords, and sections, and the Website Manager builds the page from your real components.
New landing pages
"Create a landing page for our new feature" — it copies the structure of an existing page as a template, reuses your Layout and components, writes conversion-minded SEO-structured copy, and can generate an on-brand hero image.
Safe delegation to non-technical teammates
Marketing can ship site changes without repo access or code knowledge — every change is still a normal PR your developers (or CI) gate before it goes live.

What to expect

  • A living Site Profile of your website — framework, design tokens, key components, page patterns, and brand voice — the agent reasons from on every task
  • Any edit made by plain-English instruction, delivered as a single reviewable GitHub Pull Request
  • New pages built from your real components and Layout, on-brand and SEO-structured (meta, single H1), optionally with generated imagery
  • Zero direct pushes to your live branch — branches, PRs, CI preview builds, and Git rollback by construction
  • A complete change history in the workspace: instruction, PR link, branch, and status for every task

Metrics to watch

  • Task success rate — changes reaching an open PR vs failing (visible per task in the Changes history)
  • PR merge rate and time-to-merge — how often the agent's diffs ship as-is
  • Edit precision — diffs touching only what was asked (the engine rejects suspicious rewrites, but review confirms)
  • New-page consistency — created pages matching your design system without manual rework
  • SEO handoff throughput — Page ideas from the SEO Agent turned into built, merged pages
  • Learning freshness — re-learn after big redesigns so the Site Profile matches the current site

FAQ

Can it break my live site?
It can't push to your live branch — every task commits to a new web/<action>-<timestamp> branch and opens a Pull Request. Your CI/preview build runs first and you review the diff before merging; edits are also bounds-checked (rejected if the rewritten file is empty, unchanged, or drastically resized).
Will new pages look like the rest of my site?
Yes — it learns your design tokens, components, page pattern, and brand voice first, then generates new pages from an existing page as the template, reusing your actual Layout and components.
Do I need to know code?
No. You describe the change in plain English in the Website workspace; the agent picks the files from its learned inventory and makes the edits. Developers can still review every PR.
What frameworks and repo layouts are supported?
Any GitHub-hosted site — Astro, Next.js, and React are detected explicitly, and monorepos work because the site subfolder is auto-detected during learning. Only GitHub (with repo scope) is required.
How does it work with the SEO Agent, and does it run on a schedule?
It's on-demand, driven from the Website workspace — there's no run schedule. The SEO Agent's Page ideas deep-link into the workspace with the brief pre-filled (title, slug, keywords, rationale, sections), so its recommendations become built pages in one click.