Brainstead gives each of your agency's clients its own AI agent in Slack. Mention the client's brain in a channel, for example @brain-harbour, and it drafts in that client's voice from their website, brand files and skills.
Every draft lands in a review queue with a SHA-256 fingerprint of its exact bytes. The agent holds no credential to post to Slack, LinkedIn or anywhere else, so nothing goes out until a person sends it. It runs on Claude by default, or another frontier model your client chooses.
Design-partner stage · priced per client brain, not announced
How to
Create a brain and give it the client's name, for example harbour. Brainstead creates its own storage bucket and encryption key in AWS Frankfurt (eu-central-1) at that moment.
Add the site's address. The brain reads up to 25 pages into knowledge/website/, which becomes its first source of facts and voice.
Drop tone-of-voice guides, approved copy and product sheets into brand/. Skills such as /draft define how each kind of request is handled.
Each brain has its own Slack app, named after the client. Invite it to the channel where the account team works, for example #harbour-marketing.
Mention it with a skill: @brain-harbour /draft a LinkedIn post about our Q3 launch. A thread is one continuing conversation, so follow-ups keep the context.
The draft is saved as a file, for example drafts/q3-launch-linkedin.md, with its fingerprint and the reviewer's notes. A person edits it, sends it or drops it.
Good to know
Each brain has one owner, and only that person can mention it and get an answer. Anyone else gets nothing back, so a client's material isn't exposed to the whole workspace by accident. Roles for the rest of your team, and view-only access for clients, are in development.
What you get back
The reply in the thread is short: where the draft is saved, the angle it took and why, and the gaps it found. In a real run on a coffee roaster's brain, it explained that the site had no founding story, so the post led with how the beans are sourced instead of inventing a history.
The draft itself is a markdown file with the reviewer's notes underneath. Each voice choice names the file it came from, and each claim it couldn't back up carries [NEEDS-FACT], so the account manager knows exactly what to check before it goes to the client.
What it adds
[NEEDS-FACT] markers wherever a claim has no source, instead of a confident guess.Questions
No. The agent holds no channel credential. It writes drafts into a queue, and a person sends them. A one-click signed approval is in development.
Not yet. Today a brain has one owner. Roles for the rest of the team are in development.
No. Slack and tool tokens stay encrypted under the brain's own key. The agent only holds a 15-minute pass to our tool proxy, which adds the real token on the way out.
Claude by default, through Anthropic's API, or another frontier model your client chooses, on their own subscription or API key.
Brainstead is priced per client brain, not per seat. Pricing isn't announced yet; design partners agree it with us directly.
It isn't a scheduling or publishing tool, and it isn't immune to prompt injection. It limits what the agent can reach and puts a person before anything goes out.
Related
Each client brain also works with Brand voice, ClickUp, Client isolation and ChatGPT projects vs Brainstead. See what runs today and what's in development, the security page for questionnaires, or start from the Brainstead home page.