Claude AI Agents,
Built for Real
Business Outcomes
I build Claude integrations for support, document search and API workflows. We define what the agent can do, test it on representative tasks and plan how your team will monitor it.
Six agent shapes for practical workflows
The scope can include evaluations, access controls, logging and an operational runbook. We agree on the controls before implementation.
Customer Support Agents
Search support documentation, draft replies and call approved tools. Set escalation rules and require human approval for refunds or other sensitive actions.
Sales & Outreach Agents
Research accounts, draft outreach and qualify incoming inquiries against your criteria. Validate CRM data and review messages before sending.
RAG Pipelines
Hybrid search over your docs, tickets, and Notion. Reranking, citation tracking, freshness scoring. Long-context Claude does the heavy synthesis.
Code Review & Coding Agents
Review pull requests against project rules, suggest tests and draft changes. Engineers review the output and run checks before merging.
Workflow & Ops Agents
Multi-step workflows that span tools: read invoice from email, validate against PO, post to ERP, notify Slack, log audit trail. Fail-loud, retry-safe.
Voice Agents
Connect voice interfaces to scheduling, triage or FAQ workflows. Measure response latency and accuracy in the target environment before rollout.
Why I bet on Claude for production agents
The choice depends on your tasks, integration requirements and evaluation results. These are the capabilities I assess during discovery.
Structured tool integrations
Tool schemas define how an agent calls your APIs. I validate inputs and permissions in application code, and test both successful calls and failures before granting write access.
Working with longer context
Longer context can help with document review and code analysis. Model limits, retrieval quality, latency and cost still need testing on your actual material.
MCP-native integrations
Model Context Protocol gives the agent typed access to your tools without bespoke glue. Build one MCP server, plug into Claude Desktop, Claude Code, your custom app. Standards beat sprawl.
Explicit safety controls
Access restrictions, data minimization, audit logs and approval steps are part of the application design. Compliance requirements need a separate review of the full deployment.
Extended thinking for hard agent loops
For multi-step tasks, I compare reasoning settings using task completion, latency and cost. Tool errors still need retries, limits and a path back to a human.
Measure operating costs
I compare models, caching and batch processing against your workload. Cost estimates include token usage, retrieval, hosting and monitoring, then are checked against pilot traffic.
Four phases. Reviewable deliverables.
A typical delivery plan is shown below. We agree on scope, dates and acceptance checks after discovery.
-
DiscoverDiscovery
Write the agent's job description
I review the workflow with the people who run it, document permitted actions and define examples we can use to evaluate the agent. We agree on success criteria before building.
DeliverablesAgent spec doc Eval set v0 Risk register Architecture sketch -
DesignPrototype
Prototype the smallest useful version
Your team reviews a small working prototype, its tool permissions and the first evaluation results. This is where we change or stop approaches that do not work on your tasks.
DeliverablesClickable prototype Tool schemas Prompt v1 First eval scores -
BuildBuild and test
Build and test the agreed workflow
Implement the integrations and controls the project needs, then test task quality, tool failures, permissions and operating costs before release.
DeliverablesProduction code Eval suite Runbook Dashboards -
DeployRelease
Release with a rollback plan
Roll out at a pace that suits the workflow and review quality, latency, cost and escalations. Ongoing monitoring and support can be scoped separately.
DeliverablesFeature flag On-call runbook Cost dashboard Monthly reviews
Real Claude work you can click and verify
Explore the tools on this site and inspect the linked public repositories to see the scope of the work.
Agent Skills Marketplace
50+ ready-to-use Claude agent skills
A live marketplace of Claude agent skills — each one a full, copy-pasteable prompt for development, DevOps, data analysis, security testing, and content work. Built, categorized, and maintained on this site.
- Claude skills
- 50+ Claude skills
- categories
- 6 categories
- on this site
- Live on this site
AI Prompt Library
850+ reusable AI prompts
A searchable prompt library with categories, usage notes and one-click copy. Adapt the prompts to your task and check the generated output before using it.
- prompts
- 850+ prompts
- languages
- 6 languages
- on this site
- Live on this site
nexa
Multi-agent AI platform, powered by Claude
An open, public codebase: a multi-agent platform built on Claude that coordinates specialized agents through a shared task pipeline. Read the source, fork it, and run it yourself.
- agent platform
- Multi agent platform
- powered
- Claude powered
- on GitHub
- Public on GitHub
my-ai-crew
An AI agent team, powered by Claude Code
A public repo where a crew of Claude Code agents split up and run an automated content workflow end to end. Open to read, fork, and adapt.
- content crew
- Agents content crew
- Code
- Claude Code
- on GitHub
- Public on GitHub
Explore my Claude-related work
Review public repositories, practical resources and writing before we discuss a project.
Claude integration work
I build against Anthropic’s public developer tools and document how agents use your data and APIs. Project scope and provider terms are reviewed with your team before implementation.
Explore Anthropic developer docsLocal-first Claude integrator agent for business ops
Framework for safer bash execution in AI coding agents
Open source
Public repositories you can read, fork, and run — real Claude agent code, not screenshots.
Published in 6 languages
The blog supports English, Dutch, German, Spanish, Portuguese and French. Translation availability varies by article.
Read the blogTechnical articles and notes
Read articles on Claude, MCP, RAG and agent workflows. Check examples against your own stack and requirements.
Read the blogPortfolio work
Browse selected software, web and AI work. Ask for the scope and my role in any example relevant to your project.
Explore projectsThe questions smart buyers ask
If a question is missing here, ask it in the contact form. Honest answers, no vendor-speak.
How long does a Claude AI agent project take?
The schedule depends on integrations, data quality, approval requirements and how much evaluation is needed. A focused pilot can be shorter than a multi-tool production workflow. After discovery, I provide a proposed timeline and milestones for your specific scope.
What does a Claude agent project cost?
Cost depends on the workflow, integrations, data preparation, evaluation and support needs. After reviewing the requirements, I provide a written scope and price. Provider usage and hosting costs are estimated separately so you can assess the ongoing cost.
Do you build with MCP (Model Context Protocol)?
Yes, when MCP is a good fit for the tools and security model. I can build an MCP server for approved internal APIs or use a direct integration where that is simpler. Permissions, validation and auditability remain application responsibilities.
How do you handle data privacy and compliance?
We review the selected provider’s data terms and your retention and access requirements before implementation. Options include data minimization, tenant isolation, logging and human approval. An AI integration alone does not establish regulatory compliance or certification.
Can Claude work with my existing stack?
Often. I work with Laravel, Node.js, Python, databases and cloud APIs. We first check the permissions, data access and reliability of each integration; a documented API helps, but may still require custom work.
What is RAG and do I need it?
Retrieval-augmented generation brings relevant source material into an answer, such as approved documentation or support records. It can help when answers must reflect changing or private information. Whether you need it depends on the task; we test retrieval quality and citations before committing to an architecture.
What happens if the agent gets something wrong in production?
Agents can make mistakes. We define how to detect them, when to stop an action and when to hand the task to a person. Depending on the deployment, controls may include evaluations, logs, approval steps, staged release and a rollback procedure.
Do you offer ongoing support after launch?
Yes. We can scope monitoring, evaluation updates, model changes and incident support after launch. The support terms and response expectations are agreed in the proposal.
Let's ship the agent
your team can evaluate
Start with a discovery call to discuss the workflow, data and permissions. A written proposal sets out scope, delivery and acceptance checks.
