AI agents and production pipelines built for marketing agencies
Knowlytica designs and builds AI systems for marketing and creative agencies: production pipelines that teams drive from Claude through MCP, and per-client knowledge agents that work inside Slack, Notion, Linear, Figma and Canva. Three of our builds show the range. Flow is a shared generation canvas with an Adobe Premiere Pro plugin, which one agency’s studio teams used to produce more than 2,500 videos. Brainstead gives each client account its own AI agent. AgencyOS brings an agency’s Slack, meetings, e-mail and tasks into one knowledge base, with about 30 agents and an assistant that cites its sources. Work is quoted per scope after a free 30-minute call.
Status: available now. Brainstead is live at brainstead.si; production pipelines are built per agency.
How to bring Claude and AI agents into agency production
Six steps, in the order we ran them for Flow and Brainstead.
- Step 01
Map one client’s week
Briefs, assets, review rounds and the tool each step happens in. The first build targets the step that takes the most hours and comes back every week.
- Step 02
Give each client its own space
Brand guidelines, past work, tone of voice and contacts live in a separate space per client, with its own keys, so nothing crosses accounts.
- Step 03
Connect the tools where work happens
Agents read from and post to Slack, Notion, Linear, Figma and Canva. Each connection is read-only until the agency switches writing on for that tool.
- Step 04
Put production on a shared canvas
Image, video and voice steps become nodes on one canvas the whole team sees. Through its MCP server, a producer runs the canvas from Claude Desktop by describing what they need.
- Step 05
Keep sign-off with people
Anything client-facing waits for an account lead to press Approve. The agent drafts; a person sends.
- Step 06
Hand approved work to the edit
A custom Adobe Premiere Pro plugin brings approved generations into the editor’s project, so the cut starts from the canvas instead of a downloads folder.
Client knowledge must never cross accounts
One shared assistant for every client looks efficient until it drafts a pitch in the wrong brand voice, or quotes another client’s unreleased campaign. The model has no sense of account boundaries; only the system around it does.
We separate client knowledge into spaces with their own keys and permissions, and an agent working for one client cannot search another’s. Brainstead is built this way.
What Knowlytica adds to Claude for agency work
Claude writes and reasons well on its own. Agency production needs the pieces around it.
An MCP server for production
Teams drive the whole canvas from Claude Desktop in plain language.
A Premiere Pro plugin
Approved generations land in the editor’s project.
2,500+ videos produced
By one agency’s studio teams on Flow.
One agent per client
Separate knowledge, keys and permissions for every account.
Human sign-off
Client-facing output waits for an account lead’s approval.
Agent operations
Pigeonhole coordinates agents across a team, self-hosted or in a managed cloud.
Flow, Brainstead, AgencyOS and Pigeonhole are on the work page, with the other builds.
Frequently asked questions
How much does it cost?
The first 30-minute call is free. Each build is quoted in a written proposal for the agreed scope; generation costs from model providers are billed to your own accounts.
What isn’t Knowlytica?
We are not a creative agency and do not make your campaigns. We build the systems your teams make them with, and hand them over.
Does it work with Adobe tools?
Flow includes a custom Adobe Premiere Pro plugin. Connections to other Adobe apps are scoped per project.
Can we use Flow or Brainstead directly?
Brainstead is available at brainstead.si. Flow was designed and built for one agency; we build pipelines on the same pattern for yours.
Which AI models do you use?
The ones that fit each step. We are vendor-neutral and take no commission, so image, video and text models can be swapped when better ones appear.
A draft for one client used another client’s campaign line. What happened?
The agent could search more than one client’s material, usually because all client documents were loaded into one shared space. Splitting them into separate spaces with their own keys stops it.
Also works with
The production canvas runs on a custom MCP server, and agency-wide use of Claude starts with a Claude rollout. Agencies with Swiss or EU clients should also read revFADP compliance for AI assistants and the EU AI Act for Swiss companies; the Act’s labels for generated content apply from 2 August 2026. Every service is on the Knowlytica home page.