Using Reelwire from an AI assistant
Reelwire ships a Model Context Protocol (MCP) server: the same API, as tools an assistant can call, with this documentation bundled in beside them. It is a thin layer on purpose. Nothing underneath is different, and nothing is more permissive.
Everything you can do over the API you can do here. The server has a tool for every job a workspace key may do, one per job rather than one per route, so an assistant can set a workspace up from nothing, run it day to day, make a clip without publishing it, and answer questions about what went out, without anybody opening the dashboard.
That includes the agents that live in a terminal or an editor. Claude Code, Codex, Cursor and VS Code all speak MCP and all connect with the same key, which is worth knowing if the job is less "tell me what went out" and more "write the script that pushes our data in, then watch it arrive". Install it has the configuration for each.
| Page | What it is for |
|---|---|
| Install it | Connecting Claude, ChatGPT, Claude Code, Codex, Cursor, VS Code or any other MCP client, step by step |
| Every tool | All of them, grouped by permission, with the API route each one calls |
| Working with it | Worked examples and prompts: setting up, publishing, investigating, making a clip or several, making pictures, videos and music with AI |
The security model is the key
Every request the server makes acts as one ordinary API key: the one you gave your assistant, or the one Reelwire made for it when you connected by signing in. No session, no admin credential, no way to mint one. Whatever the key may do, the assistant may do; whatever it may not, neither may the assistant.
That makes the interesting decision a one-line one: give the assistant a key with the permissions you would give a colleague doing that job.
| What you want | The key to give it |
|---|---|
| Answer questions about what is going out, and why something failed | Read only |
| Draft and schedule posts for a person to approve | Manual post lists, Media, Templates and Brands read |
| Run a producer: push events in | Custom feeds - Send raw data |
| Make pictures, videos, music, fonts and texts with your AI services | Media - AI assistant, write |
| Set a whole workspace up and run it | Run the whole workspace |
Connecting from claude.ai or ChatGPT asks the same question in fewer words: reading only, which is the Read only set, or reading and acting, which is Run the whole workspace.
A refusal comes back to the assistant as a fact it can report, naming the permission that is missing and where to widen it, rather than as an error it will retry the same way.
What is worth telling your assistant
Start with the guided run. reelwire_getting_started is the shortest way from an empty
workspace to a published clip. It reads what your workspace already holds and answers with one thing: either a question
to put to you, or the next step to take. So the conversation goes question, answer, question,
rather than the assistant deciding on your behalf and presenting you with a workspace. If yours
launches into making things unasked, say "use the getting started tool and ask me one question at a
time".
Let it show you what it renders. A clip cannot come back as a tool result, so how you see it depends on the client. In Claude and ChatGPT a small player appears under the answer and plays the clip when the render is done. Elsewhere the assistant hands you a download link to click. When the server runs on your own machine, the file is written to your disk: ask the assistant to open it in your own browser, the one your machine opens links with, not its built-in browser or preview pane. That is not fussiness. A post's download is a page that plays the video sitting beside it on disk with nothing loaded from the network, so a sandboxed viewer draws an empty player, and an empty player looks exactly like a render that went wrong.
Validate before sending. reelwire_validate_event and reelwire_validate_feed_event cost
nothing and refuse for the same reasons the real send would. An assistant that validates first gets
the field name it got wrong instead of publishing something malformed to every channel you have.
Times are UTC, everywhere. "Thursday at nine" has to become an ISO 8601 UTC instant before it
reaches reelwire_schedule_post, and the workspace's own screens will show it in UTC too.
Pictures, videos, music, fonts and texts can be made with AI, through the AI services your workspace
has connected: reelwire_ai_models says which models are ready, reelwire_ai_generate starts one,
and reelwire_ai_keep puts a result you like into the media library. The service bills your own
account at its own prices, so a good assistant asks before it makes anything. The keys for those
services are added by an owner or an admin on the API keys screen; no tool can add one. See
Working with it.
What it deliberately cannot do
No tool creates or revokes API keys, adds or removes the keys for your AI services, rotates the webhook signing secret, changes the plan, buys allowance, touches invoices, the billing address or your commission, changes how storage is split, renames or leaves the workspace, signs anybody in, or reaches another workspace.
That is not restraint in the tool layer, which an assistant could be talked out of by something it reads. The API refuses those paths to every key, whatever it holds, so the limit holds even if somebody writes their own client. Money, credentials and the account itself are for a person who signed in.
One permission reaches people. A key with Team: write can invite colleagues and give them powers, including admin, which can make keys. It is there so a workspace can be set up entirely through an assistant; give it only when that is the job.
One thing the API allows and the tools do differently:
- Uploading a file. A tool call is text, so an assistant cannot pass a picture it can see
straight into one. Instead it asks for a one-time upload link (
reelwire_media_upload_link) and sends the file to it from its own code sandbox, brings a file in from a web address (reelwire_import_media), or sends you to the Media screen to upload it yourself. Then it points a brand at it withreelwire_set_brand_asset, or uses it in a clip.
Everything else a key may do has a tool. If you find a route with no tool for it, that is a gap worth telling us about.