A Claude Code agency for the setup that ignores your rules
Your rules file grew, and Claude Code follows half of it. Developers click through prompts or switch them off. We rebuild the setup so the rules that must hold are enforced, and every engineer works from the same configuration.
For teams already using Claude Code, with mixed results
- Free 30-minute call
- Senior engineers only
- You own every file
Teams we’ve built for
Definition
What is a Claude Code agency?
A Claude Code agency is an engineering firm that configures and extends Anthropic’s Claude Code for a client’s codebase and team. It writes the CLAUDE.md and project rules, builds hooks, skills, subagents and MCP servers, sets permissions and a sandboxed environment, and trains developers to work with the tool safely.
Teams look for this service under several names: a Claude Code development agency, a Claude Code consultant, or Claude Code experts. The job is the same. Make the tool behave predictably on a real codebase.
- Vendor
- Anthropic
- Runs in
- Terminal, IDE, desktop, web, CI
- License
- Proprietary, commercial terms
What goes wrong
Three ways a Claude Code setup stops working
Each of these is described in Anthropic’s own documentation. None of them is fixed by a longer prompt.
- 01
Rules that are read, not enforced
CLAUDE.md is loaded as context, not as configuration. Anthropic’s documentation says long files reduce adherence, so the rule that matters gets lost among the ones that do not.
What we do: We shorten the file and move must-hold rules into hooks and permission rules.
- 02
Prompts approved without reading
After enough approval prompts, developers stop reading them or turn on bypass mode on their own laptops. Anthropic says that mode belongs only in containers or virtual machines.
What we do: We set allowlists, deny rules and a sandbox, so fewer prompts appear and each one matters.
- 03
MCP servers nobody reviewed
Each developer adds their own servers. Anthropic does not security-audit them, and a server that fetches outside content can carry injected instructions into a session that holds your code.
What we do: We review every server, keep the approved list in the repository and scope its tools.
What we put in place
Rules that guide, controls that enforce
Guidance belongs in files Claude Code reads. Anything that must hold belongs in hooks, permission rules and the sandbox.
A short CLAUDE.md, scoped rules
A root file with build commands and the conventions that differ from defaults. Path-scoped rules load only when matching files are read. Both are reviewed like code.
Hooks for rules that must hold
Scripts that fire on fixed events: block edits to protected paths, format after each change, run your checks before a task ends. They do not depend on the model.
Permissions and a sandbox
Allow, ask and deny rules checked into the repository, plus filesystem and network isolation for shell commands. Managed settings keep the policy from being overridden locally.
A reviewed list of MCP servers
Every server is read before it is approved. The approved set lives in the repository with tool permissions scoped, so each developer connects to the same systems.
How the repair runs
From a free call to a setup that holds
One agreed deliverable first. You see progress every week and keep everything we write.
- 01Call
Free 30-minute call
You show us the current setup and where it fails. We agree the scope and price of a paid pilot before any work starts.
- 02Audit
Read the configuration
We read your CLAUDE.md, rules, settings, hooks and MCP servers, and watch real sessions. You get a written list of findings, ordered by risk.
- 03Pilot
Fix one thing properly
The pilot has one agreed deliverable, such as enforced permissions on one repository. You get a weekly demo and a written update while it runs.
- 04Handover
Docs, runbooks and training
Your team receives the documentation, the runbooks and the training. Every file is in your repository and your accounts.
When it fits
Where Claude Code is the right tool
We use Claude Code every day. That is also how we know where it does not belong.
The platformClaude Code
Claude Code is Anthropic’s agentic coding tool. It reads a codebase, edits files, runs commands and works with git and external tools, from the terminal, IDE extensions, a desktop app, the browser and an SDK.
Where it is strong
- Works across many files and runs commands to check its work
- Rules, hooks, skills and subagents are plain files in git
- Permission rules and managed settings enforced by the client
Where it stops
- CLAUDE.md is context, not enforcement. Long files lose adherence
- Output quality drops as the context window fills
- Anthropic does not security-audit third-party MCP servers
Claude Code is the right tool when
- Your code is in git, with tests or a build the agent can run
- Engineers review every change before it merges
- You want one shared setup, versioned in the repository
We would say no when
- Nothing can check the output: no tests, no build, no review
- Your policy forbids sending source code to a hosted model
- You want non-engineers to build apps without developers
If Claude Code is the wrong tool for your case, we say so on the call.
Two ways to work
A scoped build, or an engineer on your team
Commission the repair as a project, or hire Claude Code developers who fix it from inside your team.
We repair the setup and hand it over
A paid pilot with one agreed deliverable, then further scoped work if you want it. We run it, demo it weekly and document it.
- One agreed deliverable per stage
- Weekly demo and written update
- Docs, runbooks and training at handover
- All work in your repository
A senior engineer joins your team
An engineer who uses Claude Code daily works inside your Slack and repository, fixing the setup alongside your developers for as long as you need.
- Can join within 5 working days of the call
- Works in English, in any timezone
- Senior engineers only
- NDA signed on request
FAQ
Questions about fixing a setup
Short answers. The call covers the rest.
What does a Claude Code agency do?
It configures and extends Claude Code for your codebase and team. The work covers CLAUDE.md and project rules, hooks, skills, subagents, MCP servers, permissions, a sandboxed environment and developer training. Everything is delivered as files in your repository.
Why does Claude Code ignore parts of our CLAUDE.md?
Anthropic’s documentation says CLAUDE.md is delivered as context, not enforced configuration, and that long files reduce adherence. The fix is a shorter file, path-scoped rules, and hooks or permission rules for anything that must hold every time.
Is it safe to give Claude Code access to our repository?
Permission rules, deny lists, a sandbox and managed settings limit what it can read, write and reach. Anthropic states that no system is immune to every attack, so human review before merge stays in place.
Can you work with the setup we already have?
Yes. We start by reading your existing rules, settings, hooks and MCP servers, and keep what works. Most repairs are edits to files you already own, reviewed through your normal pull request process.
How much does it cost?
Scope and price are agreed on the free call, before any work starts. The first step is a paid pilot with one agreed deliverable. Claude Code usage itself is billed by Anthropic or your cloud provider to your own account.
Who owns the configuration and the code?
You do. The code, the IP and the docs are yours, and everything lives in your repository and accounts. We sign an NDA on request.
How is this different from your Software factory service?
Software factory sets up a delivery line from ticket to deploy. This service is narrower. It configures Claude Code itself: rules, hooks, skills, MCP servers, permissions and team rollout, for developers who drive the tool directly.
Bring the setup. Leave with a plan.
You describe what breaks. We say what we would change first, and we agree the scope and price of a pilot before any work starts.
Book a Free 30-Minute Call
We map one workflow that eats your team’s week and show you what an engineer and a few AI agents could take off it. No slide deck, no sales pitch. Just a working session.
Don’t want a call? Email [email protected]
Book a free call- One workflow mapped with you, live
- A clear first step for your team
- Reply within one working day
We worked with Seif on the React Native app for our vehicle-appraisal platform. What stood out most was how much ownership he takes: he thinks a feature through, raises edge cases before they turn into bugs, and delivers something that actually works. And then there is one more thing why I wanted to work with Seif after our very first call: Exceptional clear and consistent communication. It is a pleasure to work with him and his team!
Michael EmaschowFounder, RepairCheck
Seif SgayerFounder ·View LinkedInFree Strategy Call
30 minutes · Google Meet · Free
No packages, no sales pitch. You leave with a clear first step.
- 30 min
- Google Meet
- Calendly
- No commitment
- The plan is yours to keep
- Built for teams at Tabeebi, RepairCheck and MasterPilot
HorizonLux is an independent company. Claude Code is a trademark of its owner, which does not sponsor or endorse this page.