Jabuti

Free and open sourceSelf-hostedEarly

The AI agent that works on your SEO, AEO and GEO. Not just a chat.

Jabuti sees impressions, clicks and conversions for your whole business, works on a plan you approve instead of waiting for prompts, and publishes original studies that earn citations. Open source, on your own server.

MIT license · Docker · Bring your own Anthropic API key

blog · plan
approvedMilestone: grow clicks from non-brand queries
routine runmeasure: Search Console, by device, vs control pages
mcp__dataforseo__ai_mentionswhich AI answers cite the site, and who they cite instead

Proposed experiment

Page
Hypothesis
Metric
CTR, by device
Baseline
Checks
day 3 · 14 · 28
Waiting in your inboxReviewApprove

Illustration. Nothing goes live until you approve it.

01 · Full context

Impressions, clicks and conversions, in one place.

Ranking well for a query that never converts is not a win. Each Jabuti agent sees the whole funnel for its site, and you decide which number it works to move.

Search Console

Impressions

How often you show up: visibility, before anyone clicks.

Search Console

Clicks

How often people choose you: traffic from search.

Analytics 4 key events

Conversions

What the traffic is worth: sign-ups, leads, purchases.

You choose the north star. Tell the agent which metric matters most for this site; it goes into the agent's brief, and the milestones you approve carry it as their success criterion.

And everything around the numbers

  1. Your business

    A brief per site: what you sell, what counts as a conversion, the pages that matter, audience, markets, tone, competitors and what was already tried. The agent drafts it from your site and you correct it.

  2. Your site

    It crawls the site and reads each page's text and structure: headings, links and calls to action.

  3. The results page

    Every block Google shows for a query, AI overview and People Also Ask included, keyword volume and difficulty, and the AI answers that cite you or a competitor.

  4. What happened before

    Dated notes of analyses, numbers, changes and results. Nothing is saved until you approve it.

02 · An agent that acts, not reacts

Autonomous, not unsupervised.

You do not have to keep prompting. Give the agent a goal, approve the plan, and it works toward it on a schedule, testing one measured change at a time.

  1. plan

    You approve the milestones

    Each milestone has a goal, a success criterion and a due date. The agent only sees the ones you approved, never sets their goal, and marks one done only on evidence.

  2. experiment

    Each change is an experiment

    A hypothesis, a metric and a baseline before anything ships. CTR and position are measured by device, over periods of equal length, next to control pages.

  3. publish

    You review, Jabuti ships

    The agent proposes the change on a branch of your site's repository. You review it on GitHub; Jabuti merges it when you approve, and that day starts the measurement.

  4. routine

    It keeps going without you

    Routine runs measure every day and, on their act days, propose the next change. Checks are recorded at day 3, 14 and 28 by default. Each run leaves a report.

Rules written in code, not in a prompt

  • Control pages never change
  • A page with a proposal or under measurement is locked
  • At most 3 open proposals
  • A rejected change to a page waits 30 days
  • Routine spend stops at a monthly cap
  • Unattended runs never approve anything

03 · Built to earn authority

Publish the number everyone else cites.

Journalists, blogs and AI answers need a figure to cite, and they cite whoever published it first and clearly. Jabuti's data-study skill turns your data into that figure: an original study on your site that earns links and citations.

  1. Pick a question nobody answered

    One question your niche asks, answerable with a number. The agent finds it in your Search Console queries, People Also Ask, the news and what your customers ask, and checks nobody published the same figure.

  2. Use data only you have

    Your own files (PDF, Word, Excel, slides, exports), public sources such as official statistics and open data, or a survey of your audience. Your data is the strongest: nobody else has it.

  3. Analyze it so it holds up

    A script that reruns from the raw files and gives the same numbers, with the base, the period and medians where averages mislead. A correlation is written as an association, never a cause.

  4. Publish a page made to be cited

    The main finding with its number up front, findings that make sense quoted alone, charts with their data as tables, methodology, a ready "how to cite" line, the CSV, and Article and Dataset structured data.

  5. Get it cited, keep it current

    The agent lists the journalists, blogs and sites that cite figures on the topic and drafts the e-mails; you send them. Next year it reruns the analysis on the same URL, so the links keep counting.

Your data stays yours

  • Only aggregates are published
  • Groups under 10 people or records are left out
  • No confidential business figures unless you say they are public
  • You approve every figure before anything goes out

SEO, AEO and GEO

Be found, be the answer, be cited.

People now get answers from AI overviews and chat assistants as often as from a list of links. The same agent works on all three, from the same data.

SEO

Search engine optimization

Rank and get the click on Google.

  • Finds queries and pages that gained or lost, comparing equal periods
  • Inspects why a page is not indexed
  • Checks Lighthouse, LCP, INP and CLS on the pages that matter
  • Researches keywords with volume, difficulty and intent
  • Briefs a writer subagent for new pages, in any language

AEO

Answer engine optimization

Be the answer Google shows before the links.

  • Reads the results page block by block: AI overview, People Also Ask, videos, top stories
  • Keeps the AI overview's text and every source it cites
  • Shows who answers each People Also Ask question
  • Explains a click-through rate by what competes with the page for the click

GEO

Generative engine optimization

Get cited when AI writes the answer.

  • Lists the questions Google's AI overviews and ChatGPT answer citing your domain
  • Shows who they cite instead of you, for a domain or a topic
  • Turns your own data or public data into an original study others can cite

AEO and GEO data come from the DataForSEO connector, with your own DataForSEO account. ChatGPT citation data covers the United States in English; Google's covers more places.

You stay in charge

Limits you set, not limits you hope for.

Approvals where it matters

Risky shell commands, package installs and changes to an agent's profile, skills or brief wait for you, with a diff or the exact arguments.

Ask or auto, per agent

Let an agent's commands run on their own, or approve each one. Choose per agent and per conversation.

Budgets per turn

Each agent has a cap on steps and cost per turn. Past it, the turn stops cleanly with a message, never a crash.

Real work in a sandbox

Analysis scripts run in a Linux container with Python, one workspace per agent, with no secret from the server.

Keys stay on the server

Encrypted at rest and handed only to the connectors of the agents allowed to use them. The model never sees them.

See what it did and cost

One trace line per step: tool calls, results, tokens, cost and duration. Traces never leave your machine.

Install

On your server, with one command.

Jabuti, its sandboxed shell and Caddy with automatic HTTPS, on any Ubuntu or Debian VPS. Or use the Hostinger button, AWS CloudFormation, a DigitalOcean Droplet, or your own machine.

curl -fsSL https://raw.githubusercontent.com/guilhermemauro/jabuti/main/deploy/install.sh | sh

See every install option →

FAQ

Questions people ask.

Is Jabuti free?

Yes. Jabuti is open source under the MIT license and has no paid plan. You pay for what it runs on: your server and your model provider's API usage. Some connectors, such as Serper and DataForSEO, need their own accounts.

How is this different from connecting Search Console to ChatGPT or Claude?

A connected chat answers the question you ask, when you ask it. A Jabuti agent keeps a brief and notes for each site, combines Search Console, Analytics, your site and the results page, works toward milestones you approve, proposes each change as an experiment with a baseline, and runs on a schedule. And it runs on your server.

What are AEO and GEO?

AEO, answer engine optimization, is about being the answer search engines show before the links: AI overviews, featured answers and People Also Ask. GEO, generative engine optimization, is about being cited when an AI assistant such as ChatGPT or Google's AI overviews writes the answer. Jabuti agents read the full results page, AI overview included, and can list which AI answers cite your domain and who they cite instead.

Can Jabuti tell whether ChatGPT cites my site?

Yes, through the DataForSEO connector and your DataForSEO account: it lists questions answered by Google's AI overviews or ChatGPT that cite a domain or mention a topic, with each answer's sources. ChatGPT data covers the United States in English only; Google's covers more places.

How does Jabuti help build authority?

With the data-study skill. The agent picks a question your niche asks that nobody answered with a number, analyzes your own data or public data with a script that reruns, and publishes a page made to be cited: the main finding first, a methodology, a ready citation line, the CSV and structured data. It drafts outreach e-mails to journalists and blogs for you to send, and updates the study on the same URL every year. Only aggregates are published, and you approve every figure.

Which metric does the agent optimize for?

The one you choose. An agent sees impressions and clicks from Search Console and conversions from Google Analytics 4 key events. You tell it what counts as a conversion and which metric is the north star; it goes into the agent's brief, and the milestones you approve carry it as their success criterion.

Does the agent work on its own?

Yes, within limits you set. An agent with a plan works on milestones you approve and runs routines on a schedule: it measures every day and, on its act days, proposes a change. Unattended runs never approve anything; what they need waits in your inbox, and their spend stops at a monthly cap.

How does Jabuti know a change worked?

Every published change is an experiment with a hypothesis, a metric and a baseline, checked at fixed days after it goes live (3, 14 and 28 by default). CTR and position are measured by device, over periods of equal length, next to control pages that never change.

Can the agent change my site?

Only with your approval. The agent proposes a change on a branch of your site's repository; you review it on GitHub and Jabuti merges it when you approve. A page with a proposal or under measurement is locked, and open proposals are capped.

Where do my data and API keys go?

Jabuti runs on your server. API keys and tokens are encrypted at rest and reach only the connectors of the agents allowed to use them; the model and the shell never see them. What the agent reads is sent to your model provider as part of the conversation. Traces stay on your machine.

Can I use it for several sites or clients?

Yes, that is the intended setup: one agent per site or project. Each agent has its own brief, notes, workspace and credentials scope, so the agent for one site never sees the files or keys of another.

Which AI models does it use?

Claude models (Opus, Sonnet, Haiku and Fable) with your own Anthropic API key. The loop is provider-neutral, but Anthropic is the only provider for now.

Is it ready for production?

Jabuti is early and built in public. It runs real SEO work for its author, but expect rough edges and breaking changes in the data format, handled by numbered migrations.

For developers

An agent harness you can read end to end.

Not doing SEO? Underneath, Jabuti is a general harness: several agents, each with its own model, prompt, tools, skills, MCP connectors and permissions, running real work in a sandboxed shell.

No agent framework

No LangChain, LangGraph, CrewAI, Mastra or agent SDKs. The loop, tool registry, context management, permissions, skills loader and agent switching are plain TypeScript, small enough to read in an afternoon.

Its promises are tests

Every guarantee is an obligation in the Contract (C1–C21), each a test run against a scripted fake model on every commit: every tool call gets exactly one result, shells have no secrets, skills load only when used.

Cheap long conversations

Append-only history, old tool results cleared from the model's view, summaries past a larger size, and request prefixes kept byte-identical so prompt caching holds.

Read the design and the Contract in AGENTS.md →

Project status

Early, and built in public.

It runs real SEO work for its author, but expect rough edges, breaking changes in the data format (handled by numbered migrations) and only one model provider for now. The direction is in SLICES.md, and every non-obvious choice, with its reason, is in DECISIONS.md.

Contribute

From a typo to a new connector. Small pull requests, the Contract first, and no technical debt merged. Open an issue first for anything larger than a fix.

Star or fork on GitHubContributing guide