1
One Mac/PC
Run the MVP from one machine first. Claude Code is installed there. Multica CLI and the daemon run there. Keep the blast radius tiny.
Multica MVP guide
One Mac/PC. Four agents. One visible loop.
Keep the first run simple: Claude Code and Multica on one machine, one Chief of Staff coordinating the work, and three specialist agents returning readable HTML reports instead of hard-to-review Markdown.
The shape
The minimum viable pod.
No multi-machine orchestration. No enterprise theatre. Prove the loop on one Mac/PC first, then add complexity after the thing works.
2
Claude Code runtime
Claude Code is the execution layer. Multica agents loop back to this same local runtime through the daemon.
3
Multica Cloud workspace
Multica Cloud is the board: agents, issues, comments, assignments, statuses, and the audit trail.
4
Chief of Staff
The Chief of Staff owns the parent request, writes the brief, creates the relevant child tickets, and returns the final report.
5
Four strong agents
Chief of Staff, Researcher, SEO Consultant, and Blog Writer. Small enough for the MVP; useful enough to survive in a larger team later.
Before you start
Install the two things this MVP depends on.
The setup file assumes Claude Code and Multica are already available. If either one is missing, start with the official docs below before asking Claude to build the pod. Otherwise you get tutorial soup and a terminal quietly judging everyone.
Prerequisite
Install Claude Code
Claude Code is the local execution runtime in this MVP. Install it and confirm the `claude` command works before trying to create Multica agents.
Prerequisite
Set up Multica
Multica is the workspace and issue board. You need the Multica CLI installed, authenticated, and the local daemon/runtime connected.
Setup
6 steps: inspect the instructions, or hand Claude the action file.
Each step now has two file types. The Instruction MD explains what the step means. The Give-to-Claude MD is the executable prompt/file you can hand to Claude Code to perform that specific setup step.
- 1
Give Claude the setup
First, you need Claude Code running on the Mac/PC where Multica CLI is installed. Give Claude the setup MD file, or copy/paste its contents into the Claude Code session.
- Run Claude Code on the machine that will host the Multica runtime.
- Give Claude the setup MD and ask it to follow the instructions.
- Claude should discover workspace/runtime IDs instead of guessing.
- Claude will set up steps 1–5 for you, or you can follow them manually below.
- 2
Create the four MVP agents
Create a small pod: one coordinator and three useful specialists. The demo uses SEO/blog work, but the agents should be capable beyond that one example.
- Chief of Staff coordinates requests, briefs, tickets, reviews, and final reports.
- Researcher gathers evidence, sources, competitors, and customer language.
- SEO Consultant audits sites and reviews content for search fit.
- Blog Writer writes blog content for many purposes, matching the organisation's voice and site style.
- 3
Paste agent role files
Each agent gets role-specific instructions. These should make the agent best-of-breed at the role, not a deterministic worker for one SEO demo.
- Paste Chief of Staff instructions.
- Paste Researcher instructions.
- Paste SEO Consultant instructions.
- Paste Blog Writer instructions.
- 4
Attach skills
Skills teach routing, evidence labels, final review, and readable HTML reports/previews.
- Attach routing and final-review skills to Chief of Staff.
- Attach evidence labels to all agents.
- Attach HTML report/preview skills to agents producing approval artifacts.
- Keep approval gates: nothing public/published/customer-facing without a human yes.
- 5
Run the first request
In the Claude Code session, tell the Chief of Staff to run an SEO audit and take ranking-improvement actions. If setup is right, you should not need to explain ticket creation.
- Give the Chief of Staff the website, audience, and goal.
- The Chief of Staff creates the parent ticket and relevant child tickets.
- Specialists return readable HTML reports or previews where useful.
- 6
Review the full cycle
The demo works when one human request becomes visible Multica work, specialist handbacks, readable HTML reports, and one clear final recommendation.
- Parent ticket exists and is owned by Chief of Staff.
- Relevant child tickets were created and assigned.
- Reports are readable HTML, styled like the analysed website where relevant.
- Chief of Staff returns final SEO actions and blog opportunities if supported by findings.
Sample Claude Code request
Tell the Chief of Staff the outcome. Let Multica split it.
If setup worked, the assignment roles already know how to create the relevant Multica tickets, return readable HTML reports, and suggest blogs when the SEO findings point that way. You should not have to explain the plumbing.
Do an SEO audit on https://example.com and have the team take actions to help us rank better.
Give me back a clear report on what you found, what the team did, and what needs approval before anything goes live.Expected ticket flow
What should happen next.
Healthy loop
- Chief of Staff creates/owns the parent coordination ticket.
- Chief of Staff writes the brief and creates relevant child tickets.
- SEO Consultant audits the site and recommends ranking actions.
- Researcher gathers supporting evidence, competitor examples, or buyer language when needed.
- Blog Writer creates content only if the findings support it, then returns an HTML approval preview.
- All substantial reports are readable HTML, styled like the analysed website where relevant.
- Chief of Staff returns one final SEO/action report and approval list.
Bad loop
- One agent tries to do everything.
- No child tickets are created.
- Specialists are hardcoded to one demo instead of being useful agents.
- Reports come back as raw Markdown walls nobody wants to read.
- Anything gets published without human approval. Absolutely not, cowboy.
Important limitation
This MVP proves the workflow. It is not the full operating layer.
This version is deliberately deterministic: Multica holds the issues, and Claude Code is the local runtime doing the work. That is useful because it makes the loop easy to see — but it is still mostly a routed code session, not a persistent operator with memory, channels, schedules, tools, and long-running context.
MVP layer
- Great for proving issue routing and handbacks.
- Mostly deterministic: setup files, role files, and Claude Code runtime.
- Limited by whichever tools the runtime has available.
- Useful framework. Not yet the full AI operating team.
Next layer
- Hermes acts as the persistent operator / Chief of Staff.
- Specialists can keep their narrow roles, but get memory and better orchestration.
- Agents become much more useful when connected to real tools: files, browser, GitHub, Slack, email, cron, MCP, and internal systems.
- The Multica MVP shows the shape. Hermes is how you start wiring the nervous system.
The point is the loop, then the operating layer.
If one request becomes Multica tickets, specialist work, readable HTML artifacts, and one clear Chief of Staff recommendation, the MVP works. The next step is putting Hermes above it as the cheaper, persistent operator that can remember, schedule, use tools, and keep the team moving.
References: Multica, Multica agent docs, and Hermes Agent docs.