Integrate Anthropic Computer Use for Interface Automation

Need to automate a legacy CRM with no API? Or test a desktop interface for visual bugs that unit tests miss? We integrate Anthropic Computer Use — a solution where Claude controls a computer via screenshots and actions, adapting to any interface changes. Unlike AutoHotkey or PyAutoGUI, the agent doesn't break when a button shifts by 5 pixels. Our experience: 5+ years in AI/ML and 20+ automation projects, so we deliver a reliable turnkey solution.

How Computer Use Works

Claude receives a screenshot and returns a command: click, type, scroll. Implementation in Python with the anthropic library:

import anthropic
import base64
import subprocess
from pathlib import Path

client = anthropic.Anthropic()

def get_screenshot() -> str:
    """Takes a screenshot and returns base64"""
    import pyautogui
    screenshot = pyautogui.screenshot()
    screenshot.save("/tmp/screen.png")
    return base64.standard_b64encode(Path("/tmp/screen.png").read_bytes()).decode()

def execute_action(action: dict) -> str:
    """Executes an action from Computer Use"""
    import pyautogui

    action_type = action["type"]

    if action_type == "screenshot":
        return get_screenshot()

    elif action_type == "mouse_move":
        pyautogui.moveTo(action["coordinate"][0], action["coordinate"][1])
        return "moved"

    elif action_type == "left_click":
        pyautogui.click(action["coordinate"][0], action["coordinate"][1])
        return "clicked"

    elif action_type == "double_click":
        pyautogui.doubleClick(action["coordinate"][0], action["coordinate"][1])
        return "double_clicked"

    elif action_type == "type":
        pyautogui.write(action["text"], interval=0.05)
        return "typed"

    elif action_type == "key":
        pyautogui.press(action["key"])
        return "key_pressed"

    elif action_type == "scroll":
        direction = 1 if action["direction"] == "up" else -1
        pyautogui.scroll(direction * action.get("amount", 3), x=action["coordinate"][0], y=action["coordinate"][1])
        return "scrolled"

    return "unknown_action"

def run_computer_use_agent(task: str, max_iterations: int = 30) -> str:
    """Runs the Computer Use agent to complete a task"""

    tools = [{
        "type": "computer_20250124",
        "name": "computer",
        "display_width_px": 1920,
        "display_height_px": 1080,
        "display_number": 1,
    }]

    messages = [{"role": "user", "content": task}]

    for iteration in range(max_iterations):
        response = client.messages.create(
            model="claude-opus-4-5",
            max_tokens=4096,
            tools=tools,
            messages=messages,
        )

        if response.stop_reason == "end_turn":
            return next((b.text for b in response.content if hasattr(b, "text")), "Done")

        # Process tool_use blocks
        tool_results = []
        for block in response.content:
            if block.type == "tool_use" and block.name == "computer":
                action = block.input
                result = execute_action(action)

                # If a screenshot was requested, add the image
                if action["type"] == "screenshot":
                    tool_results.append({
                        "type": "tool_result",
                        "tool_use_id": block.id,
                        "content": [{
                            "type": "image",
                            "source": {"type": "base64", "media_type": "image/png", "data": result}
                        }]
                    })
                else:
                    # After an action, take a screenshot so Claude sees the result
                    import time
                    time.sleep(0.5)  # wait for UI animation
                    screenshot = get_screenshot()
                    tool_results.append({
                        "type": "tool_result",
                        "tool_use_id": block.id,
                        "content": [{
                            "type": "image",
                            "source": {"type": "base64", "media_type": "image/png", "data": screenshot}
                        }]
                    })

        messages.append({"role": "assistant", "content": response.content})
        messages.append({"role": "user", "content": tool_results})

    return "Max iterations reached"

Why Computer Use Is More Reliable Than Traditional RPA

Classic RPA tools (Selenium, Playwright) require stable selectors and cannot adapt to element shifts. Computer Use analyzes context: if a button changes position, the agent "sees" it in the screenshot and adjusts accordingly. This is critical for interfaces that are updated or customized. Computer Use works by analyzing screenshots and returning commands in natural language, making it resilient to UI changes.

How We Implement the Integration

We design the architecture, set up a Docker sandbox with a virtual display, and write the agent for your specific task.

Secure Sandbox via Docker

FROM ubuntu:22.04

RUN apt-get update && apt-get install -y \
    xvfb x11vnc \
    python3-pip \
    chromium-browser \
    libreoffice \
    && rm -rf /var/lib/apt/lists/*

RUN pip3 install anthropic pyautogui pillow

# Virtual display
ENV DISPLAY=:99
CMD ["bash", "-c", "Xvfb :99 -screen 0 1920x1080x24 & python3 /app/agent.py"]

import docker

def run_in_sandbox(task: str) -> str:
    """Runs a Computer Use task inside a Docker container"""
    container = docker.from_env().containers.run(
        "computer-use-sandbox",
        command=f'python3 -c "from agent import run_computer_use_agent; print(run_computer_use_agent({repr(task)}))"',
        remove=True,
        mem_limit="2g",
        cpu_period=100000,
        cpu_quota=50000,  # 50% CPU
        network_disabled=True,  # disable network if not needed
        volumes={"/tmp/output": {"bind": "/output", "mode": "rw"}},
    )
    return container.decode()

Typical Use Cases for Computer Use

• Working with legacy systems without APIs — old 1C configurations, Excel macros, corporate ERPs with only desktop interfaces.
• UI testing — the agent clicks through the interface like a real user, catching visual bugs that unit tests miss.
• Code-free RPA — gathering data from multiple web interfaces, moving data between systems without APIs.

Comparison with Alternatives

What's Included in the Work

• Task analysis and agent architecture design
• Development and setup of the Docker sandbox with virtual display
• Integration with your system (data exchange, logging)
• Testing on real scenarios and optimization
• Documentation and training for your team
• Post-deployment support (1 month)

Work Process

1. Analysis — we break down the task, define scope and success metrics.
2. Design — we select the model, set up the sandbox, and write a prototype.
3. Implementation — we develop the agent, add logging and error handling.
4. Testing — we run it on test data and fix bugs.
5. Deployment — we deploy in your infrastructure (on-prem or cloud).
6. Support — monitoring and adjustments based on feedback.

Estimated Timelines

• Basic Computer Use agent with pyautogui: 2–3 days
• Docker sandbox with virtual display: 2–3 days
• Specific legacy system automation task: 1–2 weeks
• Agent action monitoring and logging: 3–5 days

Timelines are refined after project evaluation. The cost is calculated individually — contact us for an initial consultation.

Common Challenges and Solutions

• Latency: Each action requires a screenshot and an LLM call — typically 300–800 ms per step. Solution: cache repeated steps, run independent agent branches in parallel, and pre-save states that the agent visits frequently.
• Hallucinations: The agent may misidentify a button or enter data in the wrong field. Solution: validate by comparing before/after screenshots, require explicit confirmation for key steps, and limit the agent's view to only the relevant window.
• Security: The agent has direct OS access. Solution: strict Docker sandbox with network disabled, resource limits (CPU 50%, RAM 2 GB), a whitelist of allowed applications, and a full audit log of every action.

Our stack for Computer Use includes Python 3.11+, Playwright for browser management, Docker with Xvfb virtual display, and for on-premises scenarios — a VNC server for real-time agent monitoring. Every action is logged in structured JSON for later auditing and incident analysis. Evaluate your project: contact us — we'll propose architecture and precise timelines. We guarantee quality and transparency at every stage.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.