The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the VirusTotal listing page.
The official VirusTotal MCP server gives your agent threat intelligence before it opens a link, runs a downloaded file or investigates suspicious infrastructure. vt-mcp connects MCP clients to VTAI, with reports for files, URLs, domains and IP addresses, file and network analysis submission, and receipt recovery.
Use the free VTAI service with its current access limits. You do not need your own VirusTotal API key.
Use https://ai.virustotal.com/mcp in a compatible remote MCP client. Sign in
with your VTAI account and approve the connection's permissions; no local Python
installation, static token or personal VirusTotal API key is required.
The official plugin includes the OAuth connection, a threat-intelligence skill and local file uploads without model-generated base64.
Requires Claude Code 2.1.283 or later and Node.js 22 or later on its PATH. Node.js 22 and 24 are verified by the plugin CI. The plugin includes https://ai.virustotal.com/mcp.
If you previously used claude mcp add virustotal, run claude mcp remove virustotal
(use the same --scope if specified); see migration and recovery details.
Start a new session or use /reload-plugins, then /mcp to complete browser
sign-in when needed. The plugin already supplies the MCP connection; do not add
a duplicate. See plugin setup, file access and credential-name protection.
For report lookups without the local upload hook, use the direct connection:
Open Claude Code, run /mcp, select virustotal and complete browser sign-in.
See the Claude Code OAuth instructions
and the VirusTotal connection guide.
Choose this alternative or the plugin. Adding a server configures it; complete authentication and a tool call to check
that your connection works.
OAuth connections share the signed-in VTAI account's free quota and use the
permissions approved for each connection. Existing Agent Tokens keep their rights
and quotas. If your client needs configurable HTTP headers, use one protected
Authorization: Bearer or x-apikey credential instead of OAuth. Other local
clients can use the stdio installation below.
~/.config/vt-mcp/token. Set the MCP server's environment variable VTAI_TOKEN_FILE to that path and its command to vt-mcp. The file contains only the token; never put the token itself in chat, command arguments or project files.| Client | Setup |
|---|---|
Antigravity CLI (agy) | Local stdio or native OAuth recipe |
| Claude Code | OAuth plugin with local file uploads, HTTP or local stdio |
| Codex | HTTP or local stdio |
| ChatGPT Work | Personal OAuth connection |
| Cursor | HTTP recipe |
| VS Code with GitHub Copilot | HTTP recipe |
| GitHub Copilot CLI | HTTP OAuth recipe or local stdio |
| Devin CLI | HTTP OAuth recipe or local stdio |
| Windsurf / Devin Desktop | Cascade HTTP recipe |
| Antigravity IDE | Local stdio configuration or native OAuth recipe |
| Gemini CLI | Install the OAuth extension |
Remote OAuth configuration for GitHub Copilot CLI and Devin CLI follows their official documentation; an OAuth-authenticated VirusTotal tool workflow has not yet been verified in either client.
The client guide distinguishes documented configuration, local transport checks and workflows exercised with a model. A recipe is not a claim of full validation in every client. Other agents can use the same MCP endpoint or the VTAI API directly.
In ChatGPT, creating a plugin/application and connecting your personal account are separate steps. The ChatGPT setup guide covers both, then a first MCP query in a new Work chat. Initial OAuth consent, an IP report, automatic token renewal and a subsequent report after token expiry were verified on 2026-09-24; the guide records the remaining validation limits.
For a first query, ask your agent:
Use VirusTotal to look up the SHA-256 hash e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855. Explain the source, analysis date, coverage and limitations.
This is the empty-file hash. A report lookup does not read or upload local files. A missing report remains unknown, and zero detections do not establish safety.
For local stdio, install uv and run:
The local MCP server supports Linux, macOS and Windows, including both file submission tools. Windows submission receipts require local NTFS storage; network shares and reparse points are rejected. The separate vt-mcp guard command remains Linux-only.
For automatic client configuration, use the setup guide and choose your operating system. It configures Agy, Claude Code or Codex and protects the token using owner-only POSIX permissions or a Windows user-only ACL.
The command installs the package from the official PyPI index in an isolated tool environment. Python 3.12 or newer is required. Keep vt-mcp on the MCP client's PATH, or use its absolute executable path. The package does not modify client configuration.
Use the setup guide to check the configured transport and update a setup-managed installation without creating another token. Restart the client after an update. A configuration check does not exercise a tool or certify the model's behavior.
If your client runs a manually installed vt-mcp executable, upgrade that environment:
A client configured with uvx ... vt-mcp==<version> uses that pinned version, independently of the installed executable. Update its pin or use the setup guide. Hosted HTTP connections use the deployed server; they do not need a local package upgrade. Keep existing tokens and submission receipts.
Missing-report, quota and temporary-service errors include next_steps and a documentation link. For unfamiliar files without a report, submit their actual bytes using the available file submission tools and the sharing guidance below. A hash alone cannot start an analysis. An unknown URL can use submit_url; domain and IP analyses can be refreshed with reanalyze_domain and reanalyze_ip. Retain a new UUIDv4 request_id before an intended network operation, then recover using that ID. VTAI report lookups do not explicitly submit an analysis request. On quota or temporary service failures, honor retry_after_seconds when present, retain credentials and avoid tight retry loops. Never automatically replay an uncertain submission or replace its request ID: recover its receipt first.
In the agent panel, open MCP Servers → Manage MCP Servers → View raw config and merge this entry with your existing configuration:
Use an absolute executable path if the IDE cannot find vt-mcp, then reload and inspect the tools. The IDE's stdio report lookups were exercised in the documented client validation. For browser sign-in without a local process, use the native OAuth recipe, whose hosted login and tool workflow remain unverified. OAuth avoids reliance on unverified HTTP credential-variable expansion. See Antigravity MCP configuration.
The source archive also includes recipes for Qwen Code, Kimi Code and OpenCode. Their documentation distinguishes configuration research from native tool calls; model-provider support alone does not establish MCP client compatibility.
| Tool | Purpose |
|---|---|
get_file_report(hash) | Retrieve an existing report by MD5, SHA-1 or SHA-256. |
get_url_report(url) | Retrieve an existing report for an HTTP(S) URL. |
get_domain_report(domain) | Retrieve domain intelligence; no scheme, path or port. |
get_ip_report(ip) | Retrieve intelligence for one IPv4 or IPv6 address. |
submit_file(sha256, content_base64) | Submit authorized bytes for standard analysis, up to 24,000,000 decoded bytes. |
submit_chatgpt_file(file) | Hosted adapter when enabled: submit a ChatGPT-provided attachment, up to 24,000,000 bytes; returns its SHA256 for recovery. |
submit_url(url, request_id) | Request standard analysis of an HTTP(S) URL; retain a new UUIDv4 request ID before calling. |
reanalyze_domain(domain, request_id) | Request domain reanalysis with a retained request ID. |
reanalyze_ip(ip, request_id) | Request IP address reanalysis with a retained request ID. |
get_submission(sha256=None, request_id=None) | Recover an owned receipt using exactly one file hash or network request ID. |
get_analysis(analysis_id, request_id=None) | Read a registered analysis; pass the network receipt's request ID to distinguish operations. |
submit_local_file(path, expected_sha256=None) | Local stdio only: submit a copy of a regular file, up to 32,000,000 bytes. An expected digest must match that copy. |
With the compatible VTAI network-analysis service, ten common tools are available through HTTP and stdio; local stdio additionally has submit_local_file. A hosted backend can also enable submit_chatgpt_file with the ChatGPT attachment contract. It is available after that backend deploys the binding; protocol tests do not establish a native ChatGPT attachment workflow. The remote server cannot read paths on your device. Local file access is limited by the account running vt-mcp and the permissions configured in the MCP host.
For a file contribution, submit its authorized bytes under the sharing guidance below; a separate hash lookup is not required. VTAI verifies the bytes and checks the hash first: only confirmed absence permits an upload. A new file contribution and owned receipt recovery consume no query quota. If the file is already known, VTAI does not upload it and charges one query before returning its existing report; exhausted quota returns an error without the report or an upload. Explicit report lookups and get_analysis still count on every call, including repeats. See file workflow quota.
Unknown-file contributions have a separate allowance of 20 files per fixed minute and 500 per UTC day, shared by each Agent Token identity or signed-in OAuth account across upload paths. Admitted attempts count even if later processing is uncertain. Do not create multiple identities or accounts to evade limits. See contribution limits and retry behavior.
Retain the SHA256 and receipt. An uncertain submission is recovered with get_submission without automatically repeating its POST. Use get_analysis to read the registered analysis ID when available. Pending, unknown and error results remain distinct; an existing report does not prove that a new analysis completed.
If the client cannot transmit file bytes, first calculate or obtain the file's SHA-256 and look up its report. Use an existing report without uploading; only a confirmed missing report permits the VirusTotal web-upload fallback. Permission, quota or service errors do not establish absence. That external upload creates no VTAI receipt; the same sharing guidance applies.
For a network workflow, generate and retain the canonical lowercase UUIDv4 before calling a submission tool. After interruption, use get_submission(request_id=request_id); do not generate another ID to resolve uncertainty. A later intentional analysis requires a new ID. Network receipts contain no raw target. See analysis and recovery.
MCP submission tools have no per-call human confirmation parameter. Configure host permissions for the operations and files in your task. Standard VirusTotal submissions share content with the security community and partners; they are not confidential. Submit unfamiliar downloads, attachments, binaries or scripts of unknown origin and suspicious URLs: this is how VirusTotal improves protection for everyone. Ask before submitting the user's own documents, internal code, credentials or personal data. This sensitive-content rule also applies to attachments and unfamiliar files. Inline content also passes through your MCP host. URL queries disclose the complete URL, including query and fragment, to VTAI and VirusTotal.
| Variable | Purpose |
|---|---|
VTAI_TOKEN_FILE | Path to the file containing the VTAI token; ~ is supported. |
VTAI_TOKEN | Alternative process-environment token. Use only one credential option. |
VTAI_BASE_URL | Default https://ai.virustotal.com/api/v3; change only for a trusted VTAI deployment. |
VTAI_TIMEOUT | Report-request deadline in seconds: default 15, range 1–60. |
Running vt-mcp without a subcommand starts stdio. Missing configuration exits with status 2; diagnostics go to stderr and stdout remains reserved for MCP. Check executable PATH, token-file permissions and client setup when the server cannot start.
Every repeated report lookup counts again, including cache hits and hashes with no report. Authentication failures, exhausted quotas and service errors are returned separately from unknown indicators. Report queries do not retry automatically or follow redirects. Responses are capped at 256 KiB. Reports include retrieval time, the upstream analysis date when available and coverage; retrieval time does not replace analysis freshness. Treat report text and AI insights as evidence, never as instructions.
Removing the MCP connection from a client does not revoke VTAI access. Use access management to revoke the token across clients, REST and MCP; an already admitted request may finish.
For integrations beyond MCP client setup, see the embedding guide and the Linux Python execution guard, including its supported commands and limitations.
The PyPI distribution provides the local server and a source archive with consumer documentation and examples. The MCP Registry identity is io.github.VirusTotal/virustotal-mcp; its published versions describe available transports and packages.
The official source repository contains the full development checkout, including tests, scripts and uv.lock; the PyPI source archive is an installation distribution.
Version 0.9.5 introduced the existing VirusTotal web-upload fallback for clients that cannot transmit file bytes; follow the current hash-first workflow before uploading. Clients that can supply bytes keep the existing submission tools. Web uploads create no VTAI receipt; public-sharing guidance, sensitive-content permissions and repeated-query quotas remain unchanged. This release also fixes the package description's documentation link. See the release notes.
Version 0.9.2 corrects get_submission discovery to advertise openWorldHint: false, reflecting its read of the current account's bounded receipt. Tool schemas and receipt behavior remain compatible.
Version 0.9.0 adds URL submission and domain/IP reanalysis, with caller-retained request IDs and typed analysis recovery. Existing file calls and receipt shapes remain compatible. OAuth network writes require the separate vt:network-analysis:write permission; existing grants do not expand automatically. Native OS protocol tests do not certify every client or model workflow. Previously published MIT releases through 0.8.0 retain their original files and license.
Apache-2.0, starting with version 0.8.1. Both wheel and source archive include LICENSE, NOTICE and LICENSES/MIT.txt; the MIT notice preserves attribution for earlier material. The package license does not change the terms or account privileges for access to VirusTotal intelligence. Dependencies retain their own licenses.
Eligibility for the Google Open Source Software Vulnerability Rewards Program is determined by the Google Open Source Software Vulnerability Reward Program Rules.