MCP is the connection, not the creative assistant. The assistant runs the conversation, Kabo returns public creator evidence and our production tools handle the finished work. We’ll start with a beginner camera lesson so we can test the whole chain before building a bigger workflow.
Kabo, an AI agent for creator research, exposes research tools over MCP. The client chooses calls, reads results and formats the handoff; the protocol itself does not write our content.
Choose the client and confirm the connection
Connect https://kabo.sh/mcp-pub in a compatible client and complete Kabo sign-in. The MCP setup guide covers the connection; installation options are available for Codex and Claude Code.
For a first file-based test, choose Codex or Claude Code using the installation options. If our existing compatible chat client supports remote HTTP MCP and Kabo sign-in, we can keep working there and request copyable text instead.
Prepare the materials for this MCP task
[ ] Client choice — Remote HTTP MCP support and Kabo sign-in support. [ ] Audience brief — Who the content serves and one question to investigate. [ ] Reference scope — Platform, exact dates and optional public URLs. [ ] Owned proof — What we can demonstrate ourselves. [ ] Handoff choice — Authorized local file access or copyable text.
Pick the client around the handoff you need
A creator-operations lead wants to use the same Kabo research service from different assistants. Some collaborators work in local folders, while others need a copyable report in a chat.
A local Codex or Claude Code workflow can be useful when you want files beside existing brand material. A conversation-oriented client can be enough for a short research memo. Check its current MCP and OAuth support. Kabo does not provide your client’s file permissions, document editor or publishing connection.
Inspect the connected Kabo tools and call get_research_capabilities with {}. Tell me which capabilities are ready for [question]. Show a small plan with tool purpose, required inputs, expected output and whether any step needs further confirmation. Don’t invent results for unavailable tools.Connect once, then discover what is ready
Add the remote server through the client’s documented setup, complete the Kabo sign-in and inspect the tool list. Call get_research_capabilities with {} before planning a large collection. A missing capability should reduce the scope or change the evidence source, not prompt the assistant to invent a replacement result.
Test one complete handoff before expanding. A ten-minute source lookup that produces a usable brief tells us more than a long list of connected tools.
The video-analysis recipe shows the saved-analysis path; it uses a different state record from a transcript data job.
Run a small research pass on [topic] for [audience], within [dates]. Show the source candidates before synthesis. Keep URLs, dates and missing fields. Return one original brief with a viewer promise, our own demonstration and next actions. Save locally only if this client has authorized file access; otherwise give me copyable Markdown. Keep any returned job ID until its result is actually read.
A clear division of responsibilities
A successful first run should leave the roles and deliverable this clear:
- Kabo: Return public source evidence and available analysis.
- Assistant: Select relevant tools, inspect results and write the brief.
- Creator: Supply positioning and approve the concept.
- Production tool: Edit the owned footage.
- Output: One brief with source links and missing-data notes.
Run a small end-to-end research task
Give the assistant a bounded question and ask it to preserve source URLs and dates. Let Kabo supply public candidates and context. Have the assistant show the source table before drafting the creative brief. The reasoning runs in your assistant; this connection does not delegate the whole job to Kabo’s separate web agent.
Carry state forward instead of starting again
If a tool returns a data job, retain its job identifier and read the correct status and artifacts. A saved video analysis has a different id and reader. Do not turn a pending response into a completed report. Keep any failure visible and decide whether the remaining evidence can still answer the question.
Leave with the output your client can actually deliver
Ask a file-enabled client to save the brief in an authorized folder. Otherwise request copyable Markdown or a table. Review it before moving it into your script or planning tools. A successful MCP connection is a research capability, not automatic permission to publish.
Tell me which step is unavailable and which parts of the question the collected evidence can still answer. Do not silently replace the missing tool result with general model knowledge.
Move the deliverable out of the tool conversation
Move the reviewed brief into Google Docs for script collaboration, or keep it as a file in your content project. Choose the production editor based on your own footage and workflow rather than treating MCP as an editor.
Choose the next task from the missing deliverable: a source dossier when evidence is thin, or an original idea brief when the accepted sources are ready.
Worked examples are illustrative. Product screenshots show existing Kabo web results, not a run of the example in this guide. Use your own materials and the evidence returned in your task.