Use Kimi K3 in Kimi Code to research TikTok and Instagram
Updated
Virlo is social listening for TikTok and Instagram Reels. Kimi K3 is strong at long agent sessions, but the model has never seen this week's TikTok and Instagram numbers. Add Virlo to Kimi Code CLI's MCP config and K3 can pull trends, hooks, sounds and outlier videos from a database of 23.1M viral videos, then write files and scripts from them. Kimi K3 is Moonshot AI's agent-focused frontier model, available in Kimi Code (Moonshot's terminal coding agent), on kimi.com and through the Kimi API.
- One entry in ~/.kimi-code/mcp.json
- Browser sign-in with /mcp-config login
- K3 turns Virlo data into files, scripts and plans
Part of Virlo's AI assistant guides. Comparing terminal agents? See Claude Code or Codex.
How do I connect Virlo to Kimi K3 in Kimi Code?
Kimi Code CLI is the MCP client, and K3 is the model doing the thinking inside it. You add Virlo to the CLI once and every K3 session can use it.
Step 1: Create your Virlo account for free
Step 2: Add Virlo to the Kimi Code MCP config
Open
~/.kimi-code/mcp.json(make the file if it is missing) and add the entry below. Rather not edit JSON? Run/mcp-configinside Kimi Code and add a server named virlo with this URL.~/.kimi-code/mcp.json{ "mcpServers": { "virlo": { "url": "https://dev.virlo.ai/api/mcp/mcp" } } }Step 3: Sign in to Virlo from Kimi Code
Start Kimi Code, then run the login command. Your browser opens the Virlo approval page; approve it and Kimi Code stores the token for you.
Kimi Code/mcp-config login virloStep 4: Confirm the server is connected
Type
/mcpto see the status of every server. Virlo should be listed as connected, with its research tools available to K3 next to the built-in ones.Kimi Code/mcpStep 5: Give K3 a research task
Describe the job in plain English. K3 plans the calls, reads the results and writes whatever files you ask for in the working folder.
Your first promptUse Virlo to find the 10 TikTok videos about [topic] with the highest Virality Score in the last 30 days. Save them to research/videos.md with the hook, format and link for each.
Want Virlo in one repository only? Put the same entry in .kimi-code/mcp.json inside that project. A project entry with the same name wins over the one in your home folder.
OAuth, API key or skill: which Kimi Code route should I pick?
| Route | Config | Sign-in | Best for |
|---|---|---|---|
| OAuth | url only in ~/.kimi-code/mcp.json | /mcp-config login virlo opens the Virlo page | Your own laptop |
| API key header | url plus a headers map with Authorization: Bearer virlo_tkn_… | None at run time | Remote boxes and scripted Kimi Code runs |
| Project scope | The same entry in .kimi-code/mcp.json in the repo | Either of the above | One research repo shared with a teammate |
| Virlo skills | npx skills add Virlo-AI/Virlo-Agent-Plugin -a kimi-code-cli | Uses the server you added | Better tool choice and Virality Score ranking |
K3 only sees Virlo inside Kimi Code. The model on its own, or through the raw Kimi API, has no tools until a client attaches them.
Which hook types win on Instagram, and could K3 prove it on TikTok too?
This is the kind of table K3 can pull from Virlo and drop into docs/hooks.md: hook types ranked by how often Instagram videos using them reach a Virality Score of 18 or more, with the TikTok rate beside each.
In Virlo's snapshot of 264,108 hook-tagged Instagram videos, comparison hooks reached a Virality Score of 18 or more 19.0% of the time, the best rate on Instagram.
As of November 19, 2026
| Hook type | Instagram: share at 18+ | TikTok: share at 18+ | Instagram videos |
|---|---|---|---|
| Comparison | 19.0% | 29.8% | 2,947 |
| POV setup | 18.1% | 32.9% | 8,409 |
| Shock statement | 16.8% | 31.8% | 12,033 |
| Trend reference | 15.3% | 30.3% | 1,106 |
| Relatable scenario | 15.3% | 31.2% | 19,312 |
| Before and after | 14.6% | 28.6% | 1,892 |
Hook types by niche, with the Reels behind them
Split by 20+ niches, with every hook and video behind each number, in the Virlo app.
Open hooks in VirloHow this sample is made
- Source
- Virlo's daily hook snapshot: 264,108 Instagram and 1,365,774 TikTok videos whose opening hook was tagged by type.
- Window
- All tagged videos in the snapshot. Hook types with fewer than 1,000 videos on either platform, and videos with no hook, are left out.
- Method
- Share at 18+ = the share of videos with that hook type whose Virality Score is 18 or more. Sorted by the Instagram rate.
- As of
- November 19, 2026. Rebuilt every day.
- Made by
- Numbers are pulled automatically from Virlo's data; the text around them is written by the Virlo team.
Suggested citation: Source: Virlo, “Use Kimi K3 in Kimi Code to research TikTok and Instagram”, virlo.ai/kimi
How does K3 use Virlo during a long Kimi Code session?
K3 was built for long, multi-step work with little supervision, which suits research that ends in real output. In one Kimi Code session it can query Virlo, check the numbers, follow up on the strongest videos and keep writing until the deliverable is done.
A content workflow that plays to that strength:
- Ask K3 to pull this week's trends to copy in your niche into
research/trends.json. - Have it fetch the hooks behind the top videos on each trend and group them by hook type.
- Ask it to read the transcripts of the three best TikTok videos and split each into hook, body and call to action.
- Let it write
scripts/with one script per trend for your product, keeping each video's beats. - Finish by asking for a Markdown summary you can paste into Notion or a client deck.
Virality Score is how far a video beat its creator's usual views. 18 or more counts as strong, so K3 can rank a small account's breakout above a large account's normal post.
What goes in mcp.json when Kimi Code runs without a browser?
Swap the OAuth entry for a header. Kimi Code reads a static headers map on any HTTP server, so a key from your Virlo developer dashboard is enough. Keys start with virlo_tkn_; keep the file out of version control or use the project file only on machines you trust.
Header version of the entry: { "mcpServers": { "virlo": { "url": "https://dev.virlo.ai/api/mcp/mcp", "headers": { "Authorization": "Bearer virlo_tkn_YOUR_KEY" } } } }
What can I ask Kimi K3 to build with Virlo?
Paste into a Kimi Code session with Virlo connected. Replace the brackets.
Trend dataset
Pull this week's top 15 trends to copy for [your niche] from Virlo into data/trends.json with name, stage, creators this week and an example video.
Hook study
Compare which hook types reach a Virality Score of 18 or more on TikTok versus Instagram in [your niche], and write the findings to docs/hooks.md as a table.
Script batch
Read the transcripts of the 3 strongest TikTok videos about [topic] and write a 30-second script for [your product] for each one in scripts/.
Competitor file
Look up @[competitor] on TikTok and Instagram with Virlo and write competitors/[competitor].md listing their outlier videos and the pattern behind them.
Sound shortlist
List the breakout TikTok sounds and the trending Instagram Reels sounds this week and save the five best fits for [your brand] to content/sounds.md.
Dashboard page
Build a single HTML page that reads data/trends.json and shows each trend as a card with its stage and example video link.
Weekly refresh script
Write a shell script that starts Kimi Code each Monday and refreshes research/trends.md from Virlo for [your niche].
Kimi K3 vs Kimi Code vs kimi.com: which one connects to Virlo?
Moonshot AI uses the Kimi name for a model, a terminal agent and a chat app. This guide wires Virlo into the agent, where K3 does the work.
Kimi K3 (the model)
Moonshot AI's agent-focused model. It powers the other two and reaches Virlo whenever the client running it has the Virlo server attached.
Kimi Code CLI (this setup)
The terminal agent. It reads MCP servers from ~/.kimi-code/mcp.json, signs in with /mcp-config login and lets K3 call Virlo.
kimi.com chat
The Kimi chat app in the browser and on mobile. Use Kimi Code when you want K3 to work with Virlo data and files.
Other terminal agents
Claude Code, Codex and Gemini CLI reach the same Virlo server with their own one-line setup.
What are the limits?
- K3 can analyse and write from Virlo data, but Virlo itself only reads public TikTok and Instagram videos and cannot post or log in anywhere.
- MCP servers live in
~/.kimi-code/mcp.json(or$KIMI_CODE_HOME/mcp.json) and in a project's.kimi-code/mcp.json; restart the session after editing either. - Only enable project-level MCP files in repositories you trust, since they load when a session starts.
- Transcripts are available for TikTok; Instagram Reels return hooks, formats and numbers. Trends to copy update every hour.
- Plan details are on the pricing page.
Frequently asked questions
How do I use Kimi K3 with MCP servers?
Run K3 inside Kimi Code CLI, which is an MCP client. Add the server to ~/.kimi-code/mcp.json or through /mcp-config, sign in with /mcp-config login and the name, then ask. For Virlo, the entry is a server named virlo with the URL https://dev.virlo.ai/api/mcp/mcp.
Where is the Kimi Code MCP config file?
User-level servers go in ~/.kimi-code/mcp.json, or $KIMI_CODE_HOME/mcp.json if you set that variable, and apply to every project. Project-level servers go in .kimi-code/mcp.json inside the working folder, and an entry there overrides a user entry with the same name, which is handy for a research repo with its own key.
Does Kimi Code CLI support remote MCP servers with OAuth?
Yes. An HTTP entry needs only a url, and /mcp-config login followed by the server name runs the browser sign-in. Kimi Code also accepts a static headers map or a bearer token variable when you would rather use an API key.
Can Kimi K3 see TikTok videos?
Not by itself. Through Virlo it receives what matters inside a video: the spoken hook, on-screen text, format, views and, for TikTok, the transcript. That gives K3 enough to explain why a video beat its usual views and draft a new script.
What is the difference between Kimi K3 and Kimi Code?
Kimi K3 is the model; Kimi Code is the terminal agent that runs it, reads and edits files, and connects to MCP servers. Virlo plugs into Kimi Code through mcp.json, and K3 decides during a task when to call Virlo, how to read the results and what to write next.
How do I add skills to Kimi Code?
Run npx skills add Virlo-AI/Virlo-Agent-Plugin -a kimi-code-cli from your terminal. It installs two skills: virlo, which picks the right research tool and ranks results by Virality Score, and short-form-trend-research, which turns trend data into a weekly content plan. Kimi Code loads them when a task matches.
Is Kimi K3 better than Kimi K2 for research tasks?
K3 is the newer model and is built for longer agent sessions with less supervision, which helps when a task chains many Virlo calls, follows up on the strongest videos and then writes files. Both models run inside Kimi Code, and the Virlo entry in mcp.json is identical for either one.
Can Kimi Code post to Instagram through Virlo?
No. Virlo is a read-only research source that reads public TikTok and Instagram videos and never signs in to your accounts. Kimi Code can turn that data into reports, datasets, captions and scripts in your project folder, and you post the finished content with the tools you already use.
How do I check that Virlo is connected in Kimi Code?
Type /mcp in a Kimi Code session. It lists every configured server with its status, and virlo should show as connected with its research tools available. If you edited mcp.json while a session was open, start a new session so Kimi Code reloads the file.
Give Kimi K3 the numbers it is missing
Add one entry to Kimi Code and let K3 research TikTok and Instagram before it writes.
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
- 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
- OpenAI Agents SDK
- Mastra
- Vercel AI SDK
- Claude Agent SDK