What is agnt?
agnt is an MCP server written in Go that bridges your AI coding agent and the browser. It detects the project, manages dev processes, and runs a reverse proxy that instruments the page, so the agent can see screenshots, errors, network traffic and audits.
A submit button sits half off-screen. The README demo shows three rounds of blind terminal automation failing to fix it, then agnt fixing it in one pass because the agent can see the live page.
agnt creates a two-way channel between browser and agent. The agent receives screenshots, DOM details and errors, and you can send notes or sketches back from a floating indicator on every proxied page.
What the agent can do
- Start a proxy in front of your dev server and take screenshots, inspect elements and audit accessibility.
- Read JavaScript errors and network traffic as structured data instead of pasted text.
- Receive wireframe sketches drawn directly on your running UI.
- Run scored audits and a responsive audit across viewport sizes.
Why structured data helps
The README argues that structured data uses fewer tokens than descriptions. An error summary from proxylog replaces a typed paragraph, and consolidated stack traces turn long React error walls into short, actionable summaries.
Who it is for
Best for
- Developers in long vibe coding sessions who keep describing what the browser shows.
- Teams that want accessibility and layout checks beside a skill like an accessibility audit skill.
- Projects with several dev servers, since agnt detects the stack and registers scripts and proxies.
Not for
- Anyone who wants a single-purpose tool. It has 17 tools plus a PTY wrapper and daemon.
- Production monitoring. It targets local development and test sessions.
- Channel mode users who need stability: the README marks it beta and Claude Code only.
Use it with UiChemy
agnt debugs whatever runs in a browser. After UiChemy turns a design or an AI build into an Elementor or Gutenberg page, a proxy whose target_url points at your local WordPress site lets the agent screenshot and audit it. Click Get Started on uichemy.com.
Connect agnt MCP Server: Browser Debugging for AI Coding Agents to your AI client
npm install -g @standardbeagle/agntInstall the binary
Run
npm install -g @standardbeagle/agnt. The README also gives a pip or uv install of theagntpackage, and a source build withmake build && make install-local.Register the MCP server
Add the config above to your client's MCP settings. It runs
agnt mcp.Optional: Claude Code plugin
Inside Claude Code run
/plugin marketplace add standardbeagle/agnt, then/plugin install agnt@agnt. The README also describesagnt skills, which installs the agent skills and registers the MCP server in one command.Start a proxy
Ask the agent to run
proxy {action: "start", id: "app", target_url: "http://localhost:3000"}, then useproxywithexecto call__devtoolfunctions such asscreenshotorinspect.Optional: configure the project
Create
.agnt.kdlin the project root to auto-start scripts and proxies. The README says agnt writes it itself when the project shape is unambiguous, andagnt initconfigures a project without launching a session.
Client configuration
{
"mcpServers": {
"agnt": {
"command": "agnt",
"args": ["mcp"]
}
}
}Endpoints and ports
https://dev.standardbeagle.com/agnt/demos/vhs-spiralThe 1:40 full demo from the READMEhttps://dev.standardbeagle.com/agnt/features/dev-authDev sign-in and protected sharingWhat you can do with it
- Instrumented reverse proxyEvery proxied page gets a floating indicator and an injected
__devtoolAPI with 50+ functions. - Screenshots and DOM inspectionCapture the page, inspect elements, check visibility and read the last click's full context.
- Real-time error captureJavaScript errors reach the agent as they happen, with message, stack and timestamp.
- Sketch modeDraw rectangles, arrows, freehand shapes and wireframe elements on your UI, then send them to the agent.
- Scored audits
__devtool.auditAll()runs 8 weighted audits covering dom, css, perf, security, seo, a11y, API efficiency and loading. - Responsive auditChecks layout, overflow and accessibility across viewport widths, with a width workbench from 320 to 1920.
- Dev sign-in personasDeclare test users in
.agnt.kdland switch between a standard user and an admin without real IdP accounts. - TunnelsExpose the proxy through cloudflare or ngrok, or keep it tailnet-only with tailscale.
Tools it exposes
detect
Auto-detects the project type and available scripts.
run
Executes scripts or commands, in the background or foreground.
proc
Manages processes: status, output, stop and list.
proxy
Controls the reverse proxy: start, stop, exec and status.
proxylog
Queries logs: http, error, screenshot, sketch and panel_message.
currentpage
Shows active page sessions with grouped resources.
get_incidents
Pulls from the always-active incident inbox.
responsive_audit
Audits layout, overflow and accessibility across viewport sizes.
api_audit
Audits API efficiency: waterfall, N+1, duplicate and chatty loads.
loading_audit
Audits loading UX: spinner cascade and concurrent fragmentation.
snapshot
Visual regression with baseline and compare screenshots.
tunnel
Manages tunnels: cloudflare or ngrok for public, tailscale for tailnet-only.
devauth
Dev OIDC personas: list, switch the browser, or mint a token.
verify_change
Rechecks the session's recorded findings after a fix.
daemon
Manages the background daemon service.
watch
Streams daemon events via agnt monitor.
channel_reply
Sends messages to the developer's browser overlay, in channel mode only.
Frequently asked questions
What is the agnt MCP server?
agnt is an MCP server for development tooling. It handles project detection, process management and a reverse proxy with frontend instrumentation, so an AI coding agent can see the browser through screenshots, DOM inspection, captured errors and audits.
How do I install agnt?
Run npm install -g @standardbeagle/agnt, or use the pip or uv install of the agnt package, or build from source with make build && make install-local. Then add an mcpServers entry named agnt whose command is agnt with the argument mcp.
Which agents does agnt work with?
The README names Claude Code and Cursor as examples and says it works with other AI coding tools. The agnt run wrapper shows claude, cursor and aider, and agnt init gemini configures a project for another agent. Claude Code also gets a plugin.
What does agnt run alongside my dev server?
It detects the project and writes .agnt.kdl when the shape is clear, covering Node apps, Django, Rails, Laravel, Phoenix, docker-compose and more. Scripts and a reverse proxy then start from that file, and changes apply live.
What tools does agnt expose?
It lists 17 tools: detect, run, proc, proxy, proxylog, currentpage, get_incidents, responsive_audit, api_audit, loading_audit, snapshot, tunnel, devauth, verify_change, daemon, watch and channel_reply. The last one works in channel mode only.
What license does agnt use?
GitHub lists the repository as Apache-2.0, while the README badge shows MIT. The two sources disagree, so read the LICENSE file in the repository itself before you depend on either one inside a product or a client project.
Can agnt help with accessibility and layout bugs?
Yes. It can run __devtool.auditAccessibility(), check contrast and tab order, and run a responsive audit across viewport sizes. For a rule-based review skill alongside it, see a design critique skill or the Elementor abilities server.