For developers / Technical overview
Technical overview
On this page
The technical side of TakePilot, for whoever sets up the agent or handles development on your team. If you only want to know how to use it, read What is TakePilot.
No API keys
- The API keys for the AI that makes video, images and voice all live on TakePilot's servers. You sign no contracts with AI companies.
- You pay for what you make with TakePilot credits. Every AI call goes through TakePilot's servers, which meter and charge each one (Credits and plans).
- Your agent plans the script and cuts. That runs on your agent's own subscription; you pay TakePilot nothing for it.
How it connects
TakePilot serves its tools as a remote MCP server.
| Item | Value |
|---|---|
| MCP server URL | https://api.takepilot.ai/v1/mcp |
| Authentication | OAuth 2.1 with dynamic client registration. The first tool call opens a browser where you sign in to TakePilot and allow access |
| Plugin | Skills (the steps for making ad videos) plus the MCP tools. Distributed from the public repository TakePilot/takepilot-plugin |
| Agent | What you install |
|---|---|
| Claude Code, Codex CLI, Grok Build | The plugin (skills, commands and tools) |
| Claude Desktop | The plugin zip (chat mode) |
| ChatGPT desktop app | A custom MCP server (tools and the get_playbook guide only; Chat mode) |
| Other MCP clients | The remote MCP server and the skill zip (steps). If the client can't take skills, the guide comes from get_playbook |
Setup steps are in Make videos with the plugin.
Scopes
The consent screen grants three scopes.
| Scope | Allows |
|---|---|
read |
Read offerings, avatars, templates, My files and ad results |
write |
Register offerings, upload files, send ad results (free) |
spend |
Call tools that use credits (make voice, images and clips; render) |
Tools
The tools are building blocks for a video. Your agent plans the video; the tools only make and save things.
| Kind | Cost | Main tools |
|---|---|---|
| Read | Free | list_products, list_avatars, list_templates, list_files, get_ad_results, get_playbook |
| Estimate | Free | estimate |
| Make | Provider cost + 20% | generate_speech, generate_image, generate_clip, build_video, render_video |
| Wait | Free | wait_jobs (returns within 50 seconds) |
| Register and send | Free | create_product, request_file_upload, sync_ad_results, set_spend_caps |
Long jobs (clips, building, rendering) return a job_id at once; wait for them with wait_jobs.
Spending safeguards
- Paid tools require
max_usd(what this call may spend) andidempotency_key(8–128 characters). If the estimate is overmax_usd, nothing is made. - Calling again with the same key and arguments returns the first result and makes nothing new. The same key with different arguments returns
idempotency_mismatch. - The server counts caps per call ($10 by default), per agent per day ($20) and per person per day ($50). Lowering a cap works at once from a tool or the web. Raising one needs you to confirm on a confirmation page; your agent cannot raise caps itself (Plugin costs and limits).
Building and rendering
- Building always runs on TakePilot's servers, in a throwaway box with no network for each build. Nothing is built on your computer.
- Rendering runs on your computer when it has at least 16 GB of memory and 10 GB free, with the HyperFrames CLI build TakePilot ships, at no cost. Otherwise your agent asks before rendering on TakePilot's servers (paid). Map videos, videos from another author's template that have a voice, and versions that use a hidden or removed file always render on the servers.
What only people do on the web
Your agent cannot do these; a person does them on the web. The tools only return the page's URL.
- Publish to Community
- Raise a cap
Let AI read these docs
https://docs.takepilot.ai/llms.txt: a list of every page- Add
.mdto any page URL to read it as markdown (for examplehttps://docs.takepilot.ai/en/developers.md)
Next steps
Updated