OpenAI Agents SDK MCP: give your agents live TikTok and Instagram data
Updated
Virlo is social listening for TikTok and Instagram Reels. Agents written with the OpenAI Agents SDK can plan, hand off and call tools, yet no built-in OpenAI tool reports how TikTok and Instagram videos are performing this week. Give the agent Virlo as an MCP tool and it can pull trends to copy, hooks, sounds and outlier videos from a database of 23.1M viral videos. The OpenAI Agents SDK is OpenAI's open-source library for building agents with tools, handoffs, guardrails and tracing, in Python or TypeScript.
- `HostedMCPTool`: OpenAI calls Virlo for you
- `MCPServerStreamableHttp`: your process holds the connection
- Same setup in Python and in TypeScript with `@openai/agents`
Part of Virlo's AI assistant guides. Want an agent inside ChatGPT instead of your own code? See ChatGPT agent. Coding with OpenAI's terminal agent? See Codex.
How do I add the Virlo MCP server to the OpenAI Agents SDK?
Six steps. Python and TypeScript each get a full example; use whichever language your agent runs in.
Step 1: Create your Virlo account for free
Step 2: Install the OpenAI Agents SDK
In Python, the
openai-agentspackage installs theagentsmodule with the hosted MCP tool and the MCP server classes. In TypeScript, the same pieces ship in@openai/agents.Terminalpip install openai-agents # or npm install @openai/agentsStep 3: Set both keys in your environment
The SDK reads
OPENAI_API_KEYfor the model. Add a Virlo key from the developer dashboard beside it; Virlo keys start withvirlo_tkn_.Terminalexport OPENAI_API_KEY=sk-... export VIRLO_API_KEY=virlo_tkn_YOUR_KEYStep 4: Attach Virlo as a hosted MCP tool
HostedMCPToolhands Virlo's address to the Responses API, which lists Virlo's research tools and calls them on OpenAI's side. Setrequire_approvaltoneverfor unattended runs, or leave approvals on while you test.agent.pyimport asyncio, os from agents import Agent, HostedMCPTool, Runner virlo = HostedMCPTool(tool_config={ "type": "mcp", "server_label": "virlo", "server_url": "https://dev.virlo.ai/api/mcp/mcp", "headers": {"Authorization": f"Bearer {os.environ['VIRLO_API_KEY']}"}, "require_approval": "never", }) agent = Agent( name="Trend researcher", instructions="Research TikTok and Instagram with Virlo. Cite a source video for every idea.", tools=[virlo], ) async def main(): result = await Runner.run(agent, "What are the top 3 trends to copy in [your niche] this week?") print(result.final_output) asyncio.run(main())Step 5: Or do it in TypeScript
The TypeScript library exposes the hosted route as
hostedMcpTool, with camelCase options and an optionalheadersrecord for the Virlo key.run()returns the final output once the agent is done.agent.tsimport { Agent, run, hostedMcpTool } from '@openai/agents' const agent = new Agent({ name: 'Trend researcher', instructions: 'Research TikTok and Instagram with Virlo. Link a source video for every idea.', tools: [ hostedMcpTool({ serverLabel: 'virlo', serverUrl: 'https://dev.virlo.ai/api/mcp/mcp', headers: { Authorization: `Bearer ${process.env.VIRLO_API_KEY}` }, requireApproval: 'never', }), ], }) const result = await run(agent, 'Which Instagram Reels sounds are trending in [your niche] this week?') console.log(result.finalOutput)Step 6: Run the agent with a research question
Run the script. The OpenAI traces dashboard shows each Virlo call the agent made, its arguments and the data that came back.
Your first promptFind the 10 TikTok videos about [topic] with the highest Virality Score this month and tell me which hook type they share.
Every agent you define can carry the Virlo tool, and handoffs let a research agent pass its findings to a writer agent that has no Virlo access of its own.
HostedMCPTool vs MCPServerStreamableHttp: which Agents SDK route should I use?
| Route | Where the Virlo connection lives | Code | Best for |
|---|---|---|---|
HostedMCPTool | On OpenAI's servers, through the Responses API | tools=[HostedMCPTool(tool_config={...})] | Short scripts and serverless jobs with no long-lived connection |
MCPServerStreamableHttp | In your own process | mcp_servers=[MCPServerStreamableHttp(...)] inside async with | Apps that log, filter or cache every Virlo call |
hostedMcpTool() in TypeScript | On OpenAI's servers | hostedMcpTool({ serverLabel: 'virlo', serverUrl, headers }) from @openai/agents | Node backends |
MCPServerStreamableHttp in TypeScript | In your own Node process | new MCPServerStreamableHttp({ url, name, requestInit }) with connect() and close() | Node services that need per-request control |
All four reach the same Virlo server and the same research tools. The difference is only who holds the connection.
OpenAI agent on its own vs with Virlo
| What you ask the agent | Without Virlo | With Virlo |
|---|---|---|
| Which TikTok formats are taking off? | Web search results about trends in general | Trends to copy with creators this week and an example video each |
| Write hooks for our launch | Hooks invented by the model | Hooks adapted from real openers on videos that beat their usual views |
| How is @[competitor] doing on Instagram? | No direct way to check | Their recent Reels with the ones that broke out flagged |
| Pick a sound for this Reel | Popular songs it remembers | Sounds trending on Instagram Reels over the last 7 days |
What would an OpenAI agent see when it asks Virlo for trends?
The first call most OpenAI agents make is the trends lookup. These cards are the live answer: each trend with how many creators posted it this week.
In Virlo's data, 659 creators posted "My moon my man couple montage" in the last 7 days, the most of any trend to copy as of November 19, 2026.
As of November 19, 2026 at 13:07 UTC

1. My moon my man couple montage
Fresh · 659 creators this week

2. While they're out partying btw
Rising · 346 creators this week

3. He wanna pay full price SNL trend
Fresh · 175 creators this week

4. But you're not a Buddhist trend
Fresh · 158 creators this week

5. Shoulder carry mirror selfie trend
Fresh · 158 creators this week
| Trend | Stage | Creators this week | Creators last week |
|---|---|---|---|
| My moon my man couple montage | Fresh | 659 | new this week |
| While they're out partying btw | Rising | 346 | 115 |
| He wanna pay full price SNL trend | Fresh | 175 | new this week |
| But you're not a Buddhist trend | Fresh | 158 | 33 |
| Shoulder carry mirror selfie trend | Fresh | 158 | new this week |
See all 39 trends to copy
Recipes, 30-day charts, client fit and every example video are in the Virlo app.
Open Trends to copyHow this sample is made
- Source
- Virlo's Trends to copy: formats found by Virlo's trend detector and published only after the team checks each one.
- Window
- Creators who posted each trend in the last 7 days, counted from TikTok and Instagram videos only. Trends made with an AI tool are left out.
- Method
- Ranked by creators posting this week. Stages: Fresh, Rising, Peaking, Fading.
- As of
- November 19, 2026 at 13:07 UTC. Refreshes every hour.
- Made by
- Numbers are pulled automatically from Virlo's data every hour; the text around them is written by the Virlo team.
Suggested citation: Source: Virlo, “OpenAI Agents SDK MCP: give your agents live TikTok and Instagram data”, virlo.ai/openai-agents-sdk
How do I move an Agent Builder workflow with Virlo to the Agents SDK?
OpenAI has scheduled its Agent Builder canvas to shut down on 30 November 2026 and recommends the Agents SDK for workflows that should keep running as code. Moving a Virlo-powered workflow takes a few minutes:
- Open the workflow in Agent Builder and click Code to copy the generated Agents SDK version.
- Find the MCP node in the exported code; it becomes a hosted MCP tool or an MCP server entry.
- Replace any inline token with a read from
VIRLO_API_KEYso the key lives in your environment. - Run the exported agent once with a known question and compare its answer with the canvas preview.
- Deploy it where your other services run, or keep serving it in a chat window through ChatKit.
Nothing changes on the Virlo side during the move: same server URL, same key, same research tools.
When should I keep the Virlo connection in my own process?
MCPServerStreamableHttp opens the connection from your code instead of OpenAI's. Choose it when you want to log every Virlo request, filter which tools the agent can see, or cache the tool list between runs. The connection is managed by the async with block, so it closes cleanly when the run ends.
Python example: from agents import Agent, Runner from agents.mcp import MCPServerStreamableHttp async with MCPServerStreamableHttp( name="virlo", params={ "url": "https://dev.virlo.ai/api/mcp/mcp", "headers": {"Authorization": f"Bearer {os.environ['VIRLO_API_KEY']}"}, }, ) as virlo: agent = Agent(name="Trend researcher", mcp_servers=[virlo]) result = await Runner.run(agent, "...")
What can an OpenAI agent research with Virlo?
Pass these to Runner.run in Python or run() in TypeScript. Replace the brackets.
Weekly trend brief
List the 5 trends to copy with the most creators in [your niche] this week, with what to keep, what to swap and one example video each.
Handoff to a writer
Research the top 3 outlier TikTok videos about [topic], then hand off to the copywriter agent to draft one script per video for [your brand].
Guardrailed creator check
Look up @[creator] on TikTok and Instagram and report how often their videos beat their usual views. Refuse if the handle is not found.
Sound picker
Which sounds are breaking out on TikTok this week? Return three that suit a [tone] product video, with the number of videos using each.
Hook swipe file
Collect 15 spoken hooks from videos in [your niche] scoring 18 or more on Virality Score and group them by hook type.
Transcript rebuild
Get the transcript of [TikTok link] and rebuild it as a 30-second script for [your product] with the same beats.
Structured output
Return the trending Instagram Reels sounds as JSON with title, artist and Reels in the last 7 days, matching my output schema.
ChatKit assistant
You answer questions from our content team. For every question about TikTok or Instagram, use Virlo first and link the videos you used.
OpenAI Agents SDK vs Agent Builder vs ChatGPT agent vs Codex: which OpenAI product is this?
OpenAI uses the word agent for several products. This page covers the developer library you build your own agent with.
OpenAI Agents SDK (this page)
The Python and TypeScript library for building, running and tracing your own agents. Virlo attaches as an MCP tool.
Agent Builder
The visual AgentKit canvas whose Code button exports workflows as Agents SDK code. The move section above shows how to carry a Virlo workflow over.
ChatGPT agent
The agent mode inside ChatGPT that browses and completes tasks for you. It uses Virlo through ChatGPT's apps, not through code.
OpenAI Codex
OpenAI's coding agent for the terminal and IDE. Add Virlo with codex mcp add.
ChatGPT
The chat assistant, where Virlo is added as an app in developer mode.
What are the limits?
- Virlo only reads public TikTok and Instagram data. Your OpenAI agent can analyse and write with it, but it cannot post or sign in to accounts.
- With
HostedMCPTool, OpenAI holds the connection, so the Virlo key travels in the tool config; keep it in an environment variable, never in source. - With
MCPServerStreamableHttp, restart or reconnect the server object to pick up new Virlo tools after a release; the hosted route discovers them on each run. - Trends to copy refresh every hour after a team check; sounds and outliers update through the day.
- For what each Virlo plan includes, see pricing.
Frequently asked questions
How do I add an MCP server to the OpenAI Agents SDK?
Add a HostedMCPTool to the agent's tools with type mcp, a server_label, the server_url and any headers, or open an MCPServerStreamableHttp and pass it in mcp_servers. For Virlo, use the dev.virlo.ai MCP URL and an Authorization bearer header carrying your key, then call Runner.run.
Does the OpenAI Agents SDK support MCP?
Yes. The Agents SDK offers HostedMCPTool, which lets the Responses API call a remote server for you, and MCPServerStreamableHttp, which connects from your own process. Both accept an Authorization header, which is all Virlo needs. The TypeScript library, @openai/agents, offers the same two routes as hostedMcpTool and MCPServerStreamableHttp, so a team can mix languages without changing the Virlo side.
What is HostedMCPTool in the OpenAI Agents SDK?
A tool you add to an Agent with a tool_config naming the server: type mcp, a server_label, the server_url, optional headers and an approval policy. OpenAI then lists and calls that server's tools during the run, so your code holds no open connection.
How do I move an OpenAI Agent Builder workflow with Virlo to the Agents SDK?
Open the workflow, click Code and copy the generated Agents SDK version. Its MCP node becomes a hosted MCP tool; swap any inline token for a read from VIRLO_API_KEY, run it once against a known question and deploy it beside your other services. The Virlo URL, key and tools stay the same.
What is AgentKit from OpenAI?
AgentKit is OpenAI's set of tools for building and shipping agents: the Agent Builder canvas, ChatKit for embedding chat interfaces, and connections to the Agents SDK. Virlo connects to the code side of it as a remote MCP server, and an SDK agent built with Virlo can sit behind a ChatKit chat window for your team.
Can an OpenAI agent get TikTok data?
With Virlo attached, yes. The agent can request TikTok trends to copy, outlier videos ranked by Virality Score, trending and breakout sounds, hashtags, creator lookups and full transcripts, then use them in later steps or handoffs. A research agent can pass its findings to a copywriting agent, which never needs its own Virlo access.
How do I use Virlo with the OpenAI Agents SDK in TypeScript?
Install @openai/agents and add hostedMcpTool with serverLabel virlo, serverUrl set to the Virlo MCP address, and an Authorization header. For a connection from your own process, use MCPServerStreamableHttp with the same URL, pass the header through requestInit, and call connect before the run and close in a finally block afterwards.
Can an OpenAI agent post to Instagram through Virlo?
No. Virlo is read-only research on public TikTok and Instagram data, so the agent can rank, compare and write but never publish or comment. Pair it with whatever publishing service your team already uses, and keep a human approval step between the agent's draft and anything that goes live.
Is this the same as using Virlo in ChatGPT agent mode?
No. ChatGPT agent runs inside ChatGPT and reaches Virlo through apps you add there. The Agents SDK is for agents you build and host yourself, with your own code deciding when Virlo is called. The ChatGPT agent guide covers the in-app route for people who never touch code.
Give your OpenAI agent a social media brain
One HostedMCPTool entry connects Virlo to every agent you build with the Agents SDK.
More AI assistant guides
- AI assistant guides
- Claude
- Claude Code
- Cursor
- MCP server
- TikTok MCP server
- Instagram MCP server
- ChatGPT
- ChatGPT dots
- OpenAI Codex
- Gemini CLI
- Grok Build
- Claude Cowork
- n8n
- Zapier
- Grok Bot
- Muse
- Claude Duet
- Hermes Agent
- Kimi K3
- Cue
- Manus
- Perplexity Computer
- Paperclip
- OpenClaw
- NanoClaw
- DeepSeek
- AI agents CLI
- ChatGPT agent
- Perplexity
- Microsoft Copilot
- Le Chat
- Notion
- Slack
- Discord
- Telegram
- Google Sheets
- Windsurf
- VS Code
- GitHub Copilot
- Cline
- JetBrains
- Kiro
- Google Antigravity
- goose
- Warp
- OpenCode
- Zed
- Amp
- Factory Droid
- Raycast
- Lovable
- Replit
- Bolt.new
- v0
- Ollama
- LM Studio
- Open WebUI
- LibreChat
- Make
- Pipedream
- Gumloop
- Lindy
- Relevance AI
- Airtable
- ClickUp
- Dify
- Flowise
- LangChain
- CrewAI
- Mastra
- Vercel AI SDK
- Claude Agent SDK