Recomby GEO - Open-Source GEO Skills for AI Agents

TL;DR

TL;DR: recomby-geo is an open-source GEO workflow package for AI agents. It bundles domain-specific skills, schemas, and an orchestrated 7-stage pipeline for Claude Code and Codex, but it draws a hard line around human expertise: business experts must fill the brief, and production cannot proceed until that human-in-the-loop step is complete.

Source and Accuracy Notes

• Official repo: recomby-ai/recomby-geo
• Install and usage guide: INSTALL.md
• English README: README.en.md
• Orchestrator spec: plugins/recomby-geo/orchestrator/run.md
• This article is based on the official English README, install instructions, orchestrator spec, and packaged command/schema structure as of June 4, 2026.

What Is recomby-geo?

recomby-geo is not a general SEO toolkit and not a hosted SaaS. It is better understood as a domain package for agentic GEO work.

The repo’s thesis is very specific: if you want an “AI employee” in the GEO layer, you do not get there by throwing a foundation model at content production. You need four parts working together:

• business expert
• agent runtime
• office/data CLI layer
• GEO skills

This repository focuses on the fourth part while documenting the surrounding stack. The main asset is the plugin package under plugins/recomby-geo/, which includes commands, schemas, orchestrator docs, and six vendored GEO-oriented skills.

The repo is also unusually explicit that GEO should remain collaborative. In its own framing, GEO is a strong testbed for human-plus-agent work because search engines punish low-value AI content and reward actual business insight. That leads directly to the project’s hardest rule: required expert input cannot be auto-filled by agent.

Repo-Specific Setup Workflow

Step 1: Choose runtime path

The repo supports two agent environments from the same source tree:

• Claude Code via plugins/recomby-geo/plugin.json
• OpenAI Codex CLI via plugins/recomby-geo/.codex-plugin/plugin.json

That dual-manifest arrangement is one of the more useful details in the repo. Skills are shared, orchestration logic is shared, but each runtime gets the loading model it expects.

Step 2: Install into Claude Code

Official plugin install path for Claude Code is:

/plugin marketplace add recomby-ai/recomby-geo
/plugin install recomby-geo

In that runtime, the primary interface is a set of seven slash commands.

Step 3: Install into Codex CLI

For Codex CLI, the repo documents marketplace installation as:

codex plugin marketplace add ./
# or
codex plugin marketplace add recomby-ai/recomby-geo

Codex path does not use raw slash commands in same way. Instead, the geo-pipeline skill acts as entry point and routes work to the stage procedures.

Step 4: Prepare client workspace

The orchestrator document defines a strict per-client directory convention:

mkdir -p clients/<your-project>/inputs

Then place source material into inputs/, such as PDFs, URLs, notes, decks, or raw client documents.

That directory convention is not cosmetic. The whole pipeline depends on a file-based state model across stages.

Step 5: Run stages in order

The repo is very strict about sequence. The official flow is:

/01-intake
/02-audit
/03-gap
/04-content-brief
[business expert fills required slots]
/05-production
/06-distribution
/07-reaudit

In Claude Code, those are literal slash commands. In Codex, natural-language triggers invoke the geo-pipeline skill which then routes to the same procedures.

Step 6: Respect hard human gate

This is most important design detail in repository.

The install guide and orchestrator both state that /05-production must refuse to run unless brief status is ready-for-production. That state is only reached after expert-required slots are substantively filled.

This is not a minor validation check. It is the core product philosophy:

agent prepares structure
expert supplies real business insight
production is blocked until that happens

Step 7: Validate schemas throughout

Repo documents four JSON Schemas under plugins/recomby-geo/schemas/ and says every stage validates output before next stage proceeds.

The documented dependency is:

python3
pip install jsonschema

That matters because without schema contracts, multi-stage content pipelines degrade into prompt chains with weak guarantees. recomby-geo is trying to be stricter than that.

Deeper Analysis

Strongest part is orchestration discipline

Many “agent workflow” repos are mostly prompt assets plus vague diagrams. recomby-geo is stronger because it defines:

• explicit stage order
• exact files read and written by each stage
• JSON schema contracts
• append-only publish logging
• a monthly re-audit cycle

That gives operators something reproducible, not only a set of content ideas.

Local-first angle is practical

The repo repeatedly emphasizes local-first deployment. For GEO work, that can matter more than usual because inputs may include client strategy docs, internal notes, unpublished data, and sensitive commercial positioning.

The project argues this is better suited than cloud SaaS GEO tools when:

• data cannot leave local environment
• local models are preferred
• compliance posture is strict
• teams want auditable open-source building blocks

Whether every team needs that is debatable. But for agencies and consultants handling sensitive client material, it is a credible differentiator.

The repo is more about operating model than raw skill count

Yes, it ships six GEO-oriented skills and seven commands. But the more interesting part is the operating model it enforces. The project is not pretending agent can replace subject-matter expert. It is codifying where expert knowledge must enter pipeline and refusing to continue without it.

That is a healthier design than most AI-content tooling, which often treats human review as optional afterthought.

Limits are visible too

This repo is alpha. The workflow is document-heavy, file-structure-heavy, and best suited to users already comfortable with Claude Code or Codex plugin ecosystems. It is not beginner SEO software.

Also, because skills are partly vendored from upstream open-source work, long-term maintenance quality will depend on how actively Recomby keeps those pieces coherent.

Practical Evaluation Checklist

Workflow fit:
  [ ] You already use Claude Code or Codex for structured workflows
  [ ] You want a repeatable GEO process, not a one-shot prompt pack

Human process fit:
  [ ] A business expert is available to fill brief slots
  [ ] Your team accepts hard pauses when required human input is missing

Data fit:
  [ ] Client material should stay local-first
  [ ] File-based per-client workspaces are acceptable

Operational fit:
  [ ] You are comfortable with plugin installation and schema validation
  [ ] You want explicit stage artifacts and audit trail

Security Notes

• The repo’s local-first posture reduces need to ship raw client materials into third-party GEO SaaS platforms.
• Per-client folders help isolate artifacts, but they also mean local filesystem hygiene matters. The orchestrator explicitly says client data should not be version-controlled.
• Because workflow depends on external skills and agent runtimes, you should review what each runtime transmits if hosted models are enabled.
• Hard human gate before production is not only quality control. It is a safety feature against generating authoritative-looking GEO content with missing domain truth.

FAQ

Q: Is recomby-geo a hosted GEO platform?
A: No. It is an open-source plugin and workflow package for Claude Code and Codex environments.

Q: Can the agent finish everything automatically?
A: By design, no. The repo enforces a business-expert checkpoint before production can continue.

Q: What is the main deliverable in repo?
A: The core deliverable is plugins/recomby-geo/ with commands, schemas, orchestrator rules, and GEO skills, plus architecture docs for compatible agents and office CLIs.

Q: Why does schema validation matter here?
A: Because this is a multi-stage pipeline with handoffs. Schemas keep artifacts machine-checkable instead of relying on loose prompt outputs.

Q: Is this useful outside GEO?
A: Some orchestration ideas are reusable, but the content, commands, and skills are intentionally specialized around GEO workflows.

Conclusion

recomby-geo is interesting less because it promises magic and more because it rejects magic in a few key places. It uses agents for structure, repetition, and orchestration, but it refuses to fake domain expertise where domain expertise is the whole point.

If you want a local-first, human-in-the-loop GEO operating model inside Claude Code or Codex, this repo is one of the clearer open-source examples available now.

Related reading: AgentMBox, MCP Playground, and WeSight.