Free tools Windows power users keep installed
One-click scans. No signup required.
Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
In Visual Studio Code, an AI agent is a language model connected to tools and your development context. It can inspect a repository, edit files, run commands, read the results, and repeat that cycle until it reaches (or fails to reach) a goal. The exact models, tools, approval prompts, sandboxes, and custom-agent features you see depend on the selected harness, your account, and your organization’s policies.
This guide explains the agent loop, the tool types available in VS Code, reusable agent roles, permission controls, harness differences, and a review-first workflow you can use safely.
What an AI agent does in VS Code
Ordinary chat generates text in response to a prompt. An agent extends that interaction with environment context and executable tools. Visual Studio Code documentation defines an agent as “an AI system that uses a language model and tools to complete a goal on your behalf.”
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →A practical request might be: “Find the cause of the failing tests in this project, fix it, and run the relevant tests to verify the change.” The agent may search the codebase, open the failing files, propose or apply an edit, execute the test command, inspect its output, and decide whether another change is needed.
#1 Best Overall
The normal tool loop
- Request: You state an outcome, constraints, and any files or commands that are in scope.
- Context and reasoning: The selected model receives your prompt plus permitted workspace context and tool descriptions.
- Tool action: It selects an enabled tool, such as code search, file editing, terminal execution, or an MCP operation.
- Result: VS Code returns the tool’s output, including command output, diagnostics, or changed-file information.
- Another cycle: The model evaluates the result and may call another tool, revise its plan, or report that it needs your input.
This is not an autonomous guarantee of success. A wrong assumption, incomplete context, unavailable tool, failing dependency, or misunderstood requirement can send the loop in the wrong direction. Treat each proposed edit and command as work to review.
Which tools an agent can use
VS Code groups agent tools into three broad categories. The set exposed in your session is controlled by the harness and account policy.
Built-in tools
Built-in tools cover common development operations: searching code, reading and editing files, navigating the editor, and using the terminal. They are the foundation for repository maintenance tasks and test-driven fixes.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsMCP tools
Model Context Protocol (MCP) servers add tools that connect an agent to external data or services. An MCP tool might query an internal system, retrieve documentation, or perform an operation outside the local workspace. Review what the server can access before enabling it; an external tool is part of the agent’s effective capability set.
Extension-contributed tools
Extensions can contribute tools through VS Code’s Language Model Tools API. These tools may expose domain-specific analysis or actions that are not built into the editor.
Choosing and constraining tools
The agent generally chooses among enabled tools based on your request. You can direct it with a # tool reference when the harness supports that syntax. Enabling a tool and approving a call are different controls: enabling makes a tool available, while approval settings determine whether a particular invocation requires confirmation.
Rank #2
- Enable only the tools needed for the task.
- Use explicit tool references when you want to constrain the route.
- Inspect tool names and parameters in approval prompts, especially for terminal, file-write, and external-service calls.
A repeatable workflow for agent tasks
A precise workflow reduces accidental changes and makes failures diagnosable.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstall- Define the boundary. Name the repository, target files or directories, commands the agent may run, and files that must not change.
- Ask for an inspection plan first. For example: “List the likely causes and the files you will inspect; do not edit yet.”
- Allow narrow investigation. Permit read-only searches and the relevant test command. Confirm that the command does not expose secrets or modify production systems.
- Review the proposed patch. Check the diff for unrelated formatting, generated files, dependency changes, and altered security checks.
- Run focused verification. Ask the agent to run the smallest relevant test set, then expand to broader checks if the change warrants it.
- Integrate deliberately. Keep the work in a branch or separate commit until review is complete. An agent’s successful command output is evidence, not a code-review approval.
A useful prompt pattern
Investigate the failing tests in this repository.
1. Identify the failing test and likely cause.
2. Inspect only the files needed to confirm the cause.
3. Propose a minimal fix and wait for approval before editing.
4. After approval, make the change and run the relevant tests.
5. Report changed files, commands, and any remaining failures.This prompt separates diagnosis from modification and asks for an auditable report.
Reusable custom agents
A custom agent packages a recurring role—such as planner, code reviewer, or test triage assistant—into Markdown instructions. An optional YAML header supplies metadata and a selected tool set. Once defined, the role can be reused instead of recreated in every prompt.
What belongs in a role definition
- Purpose: The job the agent performs and the expected output.
- Scope: Repositories, directories, file types, or services it may inspect.
- Process: Ordered steps such as inspect, explain, propose, edit, and verify.
- Tool policy: Tools it may use and actions that always require confirmation.
- Output contract: A consistent report format for findings, diffs, tests, and unresolved risks.
Where files live
Locations depend on the selected harness. VS Code documentation describes .github/agents as a workspace location and ~/.copilot/agents or ~/.claude/agents as user locations for relevant agent-host sessions. Verify the current official setup page for your harness before standardizing paths: support, precedence, and controls can change.
Example reviewer definition
---
description: Review a change for correctness and risk
# Configure tools according to the harness that loads this file.
---
You are a read-first code reviewer.
Rules:
- Inspect the diff and related tests before making claims.
- Do not edit files or run destructive commands.
- Report findings by severity, with file and line references.
- Distinguish confirmed defects from questions.
- End with missing tests and a short verification plan.The YAML keys accepted by a file are harness-specific. Keep the Markdown instructions portable, and consult the harness documentation for supported metadata and tool-selection syntax.
Approval and sandboxing are separate safety controls
Approval prompts
Approval controls whether an action needs your confirmation. Prompts can show the tool name and input parameters before edits, terminal commands, or external-service calls. A confirmation means you reviewed that invocation; it does not prove the command is harmless or the model’s interpretation is correct.
Sandboxing
Sandboxing restricts the filesystem and network resources available to terminal commands, including after a command has been approved. It limits where a process can read, write, or connect.
How to use both
- Use a narrow sandbox for repository-local work and deny unnecessary network access.
- Require approval for writes, package installation, database operations, and commands involving credentials.
- Keep secrets outside the workspace and avoid pasting tokens into prompts or tool parameters.
- Inspect the final diff and tool history even when every call was approved.
Sandboxing does not replace review, and approval does not remove the need to understand an action’s side effects.
Harness, account, and organization policy
VS Code can host different agent harnesses, including Copilot, Claude, and Codex integrations. “Available models, tools, and customizations depend on the selected harness, your account, and your organization’s policies.” Therefore, two developers can see different model pickers, MCP controls, handoff features, or custom-agent locations on the same VS Code version.
Questions to ask before comparing harnesses
- Which harness is enabled for this VS Code session and account?
- Which built-in, MCP, and extension tools does it expose?
- How are tool enablement, approval, and sandbox settings administered?
- Which customization format, handoffs, and role locations are supported?
- Where is the model hosted, and where do the tools execute?
Model hosting and tool execution are separate considerations. Do not assume that choosing a hosted model determines where repository files or tool calls are processed. Check the provider and organization policy for your deployment.
Reviewing changes and recovering from mistakes
Before accepting a change
- Read the diff, not only the agent’s summary.
- Check for modified configuration, lockfiles, generated artifacts, and deleted tests.
- Confirm that commands ran against the intended workspace and environment.
- Run tests independently when the change affects security, data handling, or deployment.
If the agent takes a wrong turn
- Stop the current run or deny the next tool call.
- Revert unreviewed edits with your normal version-control workflow.
- Restart with a smaller scope and an inspection-only prompt.
- Provide the exact error output and identify the incorrect assumption.
- Ask for a patch and verification plan before permitting another write.
Using an agent for visual QA without giving it a browser harness
An agent can prepare visual-test instructions, inspect generated assets, or call an enabled external tool. If your harness has no browser tool, you can still capture a URL with a screenshot API and bring the result into your review workflow. Keep the URL, viewport, authentication, and handling of private pages explicit.
Or skip the browser setup
ScreenshotNeo provides a website screenshot API and MCP server. A GET request returns PNG, JPEG, WebP, or PDF. It accepts cookie and consent banners before capture and removes more than 60 known consent platforms, newsletter popups, and chat widgets; each cleanup step can be disabled. Bot checks or CAPTCHAs, blank pages, timeouts, failed loads, and cache hits are not billed, and response headers identify the page verdict and billing status. Its MCP server exposes take_screenshot, get_page_info, and capture_pdf for AI-agent clients.
curl -G "https://api.screenshotneo.com/v1/shot" -d access_key=YOUR_API_KEY --data-urlencode url=https://stripe.com -o shot.webpimport requests
r = requests.get("https://api.screenshotneo.com/v1/shot", params={"access_key": "YOUR_API_KEY", "url": "https://stripe.com"}, timeout=90)
open("shot.webp", "wb").write(r.content)const q = new URLSearchParams({ access_key: 'YOUR_API_KEY', url: 'https://stripe.com' });
const res = await fetch(`https://api.screenshotneo.com/v1/shot?${q}`);See the ScreenshotNeo API documentation for options such as full-page capture with lazy images, CSS-selector elements, device presets, dark mode, custom CSS or JavaScript, waits, blocked resources, headers, cookies, geolocation, PDF settings, signed links, asynchronous webhooks, bulk capture, caching, and usage reporting. The Free plan includes 1,000 screenshots per month without a card; paid plans start at $5 for 3,000. Create a free ScreenshotNeo account.
Common failure modes
The agent cannot see a file
The file may be outside the workspace, excluded by policy, or unavailable to the selected harness. Open the correct folder, confirm workspace trust and access settings, and provide a narrower path.
A tool is missing
The tool may not be enabled, supported by the harness, or allowed by organization policy. Check the tool list and extension or MCP configuration rather than repeatedly prompting.
A command is blocked
Approval or sandbox policy may prohibit it. Read the proposed command, determine the minimum required permission, and adjust policy only if the risk is acceptable.
Tests pass locally but the agent reports failure
Compare working directory, environment variables, dependency versions, and test command. Ask for the complete output and rerun the command yourself.
The agent edits too much
Revert, request an inspection-only pass, specify exact files, and require a minimal patch with a diff summary before approving edits.
FAQ
Does an agent run continuously after I send one prompt?
No. It performs additional tool cycles only as needed and as permitted by the harness, approvals, and available tools.
Best Value
Can I use MCP tools without allowing file edits?
Usually you can configure a read-focused tool set, but the exact controls are harness- and organization-dependent. Verify the server’s capabilities and approval settings.
Are custom-agent files portable between Copilot, Claude, and Codex?
The Markdown role instructions may be portable, but YAML keys, file locations, handoffs, and tool-selection syntax can differ. Validate each harness separately.
Should I trust a green test run as proof the patch is correct?
No. Tests cover only what they assert. Review the diff, behavior, security implications, and untested edge cases before merging.
Frequently Asked Questions
Does an agent run continuously after I send one prompt?
No. It performs additional tool cycles only as needed and as permitted by the harness, approvals, and available tools.
Can I use MCP tools without allowing file edits?
Usually you can configure a read-focused tool set, but the exact controls are harness- and organization-dependent. Verify the server’s capabilities and approval settings.
Are custom-agent files portable between Copilot, Claude, and Codex?
The Markdown role instructions may be portable, but YAML keys, file locations, handoffs, and tool-selection syntax can differ. Validate each harness separately.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Should I trust a green test run as proof the patch is correct?
No. Tests cover only what they assert. Review the diff, behavior, security implications, and untested edge cases before merging.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
