Connect a coding agent
You fix findings in your editor, but they live in the console. Connect your coding agent to Dwaar and it reads your findings, products and scans from chat, explains them against the code in front of it, fixes them, and asks for a rescan to prove the fix. It can also triage what it can't fix, with a reason you can check.
The connection is an MCP server plus a plugin of skills. The agent signs in as you, through your browser, and can never do more than your role allows.
The fastest path
Paste this into your agent:
Connect me to Dwaar (docs.dwaar.ai)
The agent reads llms.txt and install.md, finds the install command for itself, and asks before it runs it. Then it opens your browser to sign in. No account is needed to start: you create one on the way.
Prefer to do it by hand? The commands for each agent are below. Signed in to the console already? Its Installation hub has the MCP commands filled in with your console's address: see MCP server.
What you get
| Part | What it adds |
|---|---|
| MCP server | Dwaar's tools inside the agent: search, findings, resources, components, scans, connections and more. The shared address is https://console.stg.dwaar.ai/mcp, for the console at https://console.stg.dwaar.ai. A dedicated tenant has its own. |
| Skills | Instructions that teach the agent to connect and sign in, and to review the repository you are in: its findings, the fixes, and a rescan. |
Installing the plugin gives you both. An agent without plugins can take the MCP server alone, which gives it every tool but none of the skills.
Install it in your agent
Each agent has a plugin command, an MCP only command, and a way to sign in. Use the plugin unless your agent cannot take it.
Claude Code
claude plugin marketplace add DwaarAI/dwaar-agent
claude plugin install dwaar@dwaar
MCP only:
claude mcp add --transport http dwaar https://console.stg.dwaar.ai/mcp
After the plugin install, Claude Code calls the server plugin:dwaar:dwaar. To sign in, run /mcp in Claude Code, choose plugin:dwaar:dwaar, then Authenticate. From a shell, claude mcp login plugin:dwaar:dwaar does the same.
After the MCP only install, the server is called dwaar: choose dwaar in /mcp, or run claude mcp login dwaar.
Codex
codex plugin marketplace add DwaarAI/dwaar-agent
codex plugin add dwaar@dwaar
This works in the CLI and in the IDE extension.
MCP only:
codex mcp add dwaar --url https://console.stg.dwaar.ai/mcp
To sign in, run codex mcp login dwaar.
Cursor
Cursor installs plugins from a team marketplace, so a team admin adds Dwaar once. In the Cursor dashboard, open Plugins & MCPs, choose Add Marketplace, then Import from Repo, and enter https://github.com/DwaarAI/dwaar-agent. Members then install dwaar from Customize.
Without a team marketplace, anyone can add the skills:
npx skills add DwaarAI/dwaar-agent
MCP only: open this install link, or add the server to ~/.cursor/mcp.json:
{
"mcpServers": {
"dwaar": { "url": "https://console.stg.dwaar.ai/mcp" }
}
}
Cursor then shows Needs login next to dwaar in its MCP settings. Choose Connect to sign in.
VS Code with Copilot
Turn on the chat.plugins.enabled setting. Then open the Command Palette, run Chat: Install Plugin from Source, and enter https://github.com/DwaarAI/dwaar-agent.
MCP only:
code --add-mcp '{"name":"dwaar","type":"http","url":"https://console.stg.dwaar.ai/mcp"}'
VS Code asks you to sign in the first time a Dwaar tool runs.
Antigravity
agy plugin install https://github.com/DwaarAI/dwaar-agent
MCP only: add the server to ~/.gemini/config/mcp_config.json. Antigravity calls the address serverUrl:
{
"mcpServers": {
"dwaar": { "serverUrl": "https://console.stg.dwaar.ai/mcp" }
}
}
Antigravity opens your browser to sign in the first time it connects to the server.
Gemini CLI
Gemini CLI is for paid and enterprise Google accounts. Individual accounts moved to Antigravity on 18 June 2026, so use Antigravity unless your organization still runs Gemini CLI.
gemini extensions install https://github.com/DwaarAI/dwaar-agent
MCP only:
gemini mcp add --transport http -s user dwaar https://console.stg.dwaar.ai/mcp
To sign in, run /mcp auth dwaar in Gemini CLI.
OpenCode
OpenCode takes the skills and the server separately:
npx skills add DwaarAI/dwaar-agent
opencode mcp add dwaar --url https://console.stg.dwaar.ai/mcp
To sign in, run opencode mcp auth dwaar.
Windsurf, Devin, Zed, Cline and others
Add the skills with npx skills add DwaarAI/dwaar-agent. Then add https://console.stg.dwaar.ai/mcp as a remote MCP server in the agent's own MCP settings, and sign in the way it signs in to any MCP server. The agent has to support remote MCP servers that sign in with OAuth. If it does not, use an API key instead.
Dedicated tenants
The addresses on this page are the shared ones, and the default. A customer with a dedicated Dwaar tenant has its own console and MCP addresses, and the MCP address may be on a different domain from the console. Use a dedicated address only if Dwaar gave you one. Agents follow the same rule: they use a dedicated address only when you say you have a dedicated tenant and give the address, and never guess one.
The plugin carries the shared address, so on a dedicated tenant install the skills and add the server with your address using your agent's MCP only command. In Claude Code:
npx skills add DwaarAI/dwaar-agent
claude mcp add --transport http dwaar <your MCP address>
Then sign in with claude mcp login dwaar. Once the agent is signed in, the server's instructions name your console, and the agent links to that console.
Sign in
Signing in happens in your browser, never in chat. The agent opens a page, and you pass through up to three stops:
- Sign in or sign up, with your work email. Personal addresses, such as Gmail, are refused. This is the same sign-in as the console's, so the Quickstart applies.
- Create your organization, the first time only, when nobody from your email domain uses Dwaar yet. If your company already does, the screen says Ask for an invite instead: get an invite from one of its administrators, then start again from the agent.
- The consent page. It asks whether to allow the agent to use Dwaar Console, names the person and tenant you are signed in as, and says where the access goes. For an agent on your laptop that is "an app on this computer". Choose Read only or Read and write, then Approve.
After you approve, the agent is connected and works with what you approved.
Why source control is connected in the browser
A new tenant is empty, so the agent's first useful job is getting your code into Dwaar. It does not do that itself. It sends you to the console's integrations page for your provider, for GitHub https://console.stg.dwaar.ai/integrations/github, and asks you to choose Add connection. With the GitHub App method that is a few clicks on GitHub.
That is deliberate. Connecting a provider stores a credential, and a credential should never pass through chat, where it ends up in the agent's context and its logs. Connecting integrations is a permission no API key can hold, so the agent could not do it even if it tried. Connect source control has the details for each provider.
Once the connection is saved, the agent follows the first discovery and the scans it queues, and tells you when findings arrive.
What the agent can and cannot do
The key the agent receives is named after the agent, followed by "(MCP)". It lasts 30 days. It holds the permissions of the access you chose, but only those your role grants too, and on every call it is cut down to what you hold at that moment. Lose a permission in the console and the agent loses it too.
The agent sees only the tools its key can use. With Read only:
| Tools | What they do | Permission |
|---|---|---|
session, search, overview, orgs, providers | Who you are and whether your tenant is still empty, find anything by name, the Home page in numbers, your organizations, and the integrations Dwaar can connect | None beyond signing in |
products, product | Products with their grades and top risks | products:read |
resources, resource, components, component | Repositories, images, cloud workloads and more, and the packages they use | resources:read |
findings, finding, findingCounts | Findings, one finding in full, and counts | findings:read |
findingRecommendation | How to fix one finding: a summary, the steps, and a prompt for a coding agent | findings:read |
findingHistory, findingDecisions | A finding's events newest first, with who, through what and why; and every decision on your findings, as in Findings › Decisions | findings:read |
scans, scan | Scan runs, queued and finished | scans:read |
discoveryJobs | Discovery runs, to follow a new connection | discovery:read |
connections | Your connected integrations and their health | integrations:read |
Read and write adds:
| Tools | What they do | Permission |
|---|---|---|
rescanResource, retryScan | Rescan a repository or image, or retry a failed scan | scans:run |
cancelScan | Cancel a queued or running scan | scans:cancel |
syncConnection | Run a connection's discovery now | integrations:sync |
waiveFinding, unwaiveFinding | Mark a finding accepted risk or a false positive, with an optional expiry, or revoke a waiver | findings:update |
overrideFindingSeverity | Set a finding's severity, or clear it | findings:update |
commentFinding | Add a comment to a finding, shown in its Decisions section, up to 4,000 characters. Comments are never edited or deleted. | findings:update |
The consent page offers Read and write only when the agent asks for it and your role holds some of it.
Every change to a finding needs a reason of 20 to 2,000 characters. Dwaar tells the agent to read finding and findingHistory first and to name the evidence it checked, such as the code path, the call sites, the deployment, a VEX statement or a compensating control. It tells the agent never to change a finding to lower a count or without evidence, and to comment instead when unsure. People see each decision in Findings › Decisions, marked AI agent (MCP) with the key's name.
Some things are never tools. The agent cannot delete anything, connect an integration, or touch administration: members, roles, groups and API keys. Those happen in the console, by a person. Permissions lists which permissions a key may hold at all.
Revoke access
Open Settings › API and find the key in the API keys list, named after the agent with "(MCP)" after it. Choose Revoke. The agent loses access on its next call, and has to sign in again to get it back.
There is no refresh. When the 30 days are up, the agent asks you to sign in again, and you approve a new key.
Headless use with an API key
CI jobs and claude -p cannot open a browser. Give them an API key instead, sent as Authorization: Bearer dwc_….
-
Create the key in Settings › API with the Read only (MCP) preset. For anything that outlives you, such as a pipeline, create it for a service account. Service accounts and API keys walks through both.
-
Add the server with the key. In Claude Code:
claude mcp add --transport http dwaar https://console.stg.dwaar.ai/mcp \--header "Authorization: Bearer dwc_…"In Codex, keep the key in an environment variable and name it:
codex mcp add dwaar --url https://console.stg.dwaar.ai/mcp --bearer-token-env-var DWAAR_API_KEY
The key decides the tools, exactly as above. To let a headless agent triage findings, give its key findings:read and findings:update. A key keeps the permissions chosen when it was made, so an older key doesn't gain findings:update. Treat it like a password: anyone holding it can do what it allows.
Troubleshooting
| You see | Why | What to do |
|---|---|---|
| The agent says Dwaar needs authentication | It has not signed in yet, or its key expired or was revoked. | Run the sign-in step for your agent again. |
| The browser does not open, because the agent runs over SSH | Sign-in needs a browser on some machine. | In Claude Code, run claude mcp login plugin:dwaar:dwaar --no-browser after the plugin install, or claude mcp login dwaar --no-browser after the MCP only install. Open the URL it prints, and paste the address you land on back into the terminal. |
| This sign-in cannot continue | The request was wrong, or it expired. | Start again from the agent. |
| Ask for an invite | Your company already uses Dwaar. | Ask one of its administrators to invite you, then sign in again from the agent. |
| Use your organization email | You signed in with a personal address. | Choose Sign out, then sign in with your work email. |
| The consent page offers only Read only | Your role holds none of the write permissions, or the agent asked only to read. | Ask for a role with scans:run if you want the agent to rescan, or findings:update if you want it to triage findings. |
| A tool you expect is missing | The key holds only what your role grants. | Check your role in Roles and permissions. |
A tool answers not_found | It does not exist for this key, which includes anything on organizations where you hold no permission. | Ask the agent to run orgs to see the organizations it can read. |
Any other error carries a request id after "Ref". Troubleshooting explains each error code.
Try this next
Open a repository you work on and ask the agent: "What does Dwaar know about this repository, and what should I fix first?" It matches the repository to its resource, reads the worst open findings, and proposes a fix in the code in front of you. With Read and write, ask it to rescan when you have pushed, and watch the finding close.
Then let it triage what it can't fix:
Review the open critical findings in acme/payments-api. For each one you can prove is not exploitable here, lower its severity or waive it with the evidence; otherwise add a comment.
Read what it decided in Findings › Decisions, with Through set to AI agent (MCP).