Clembot is two things kept apart on purpose: an engine that holds the agents, dashboard, and tooling, and a vault that holds each client’s data. One install serves one client, and client data never lives in the engine’s code.
The pieces
Brand files. The source of truth every agent reads before it produces anything: voice, positioning, audience personas, product facts, competitors, SEO keywords, a content playbook, and a style guide. Each client’s voice and posting cadence also live in a config file in that client’s vault.
Agents. Specialist agents, each with its own instructions, tools, and scope. A dispatcher reads the command, plans which agents run in what order, and keeps the evidence log.
Content store. A local database tracks every piece of content through its lifecycle: draft, in review, approved, scheduled, posted. The scheduling tool and the optional Notion workspace are views of that record. Neither one is the source of truth.
Dashboard. A web dashboard shows the command center, a live feed of agent activity, the content pipeline, each client’s playbook and content, the Slack personas, and the scheduled routines.
Slack. Six personas post updates, react when content changes state, and answer questions in each client’s channel.
The agent team
| Agent | What it does |
|---|---|
| Dispatcher | Plans the sequence of agents for each command and keeps the evidence log. |
| SEO specialist | Keyword research, search results analysis, content briefs. |
| Content writer | Long-form pieces: blog posts, case studies, email. |
| Social strategist | LinkedIn posts, threads, founder content. |
| Carousel drafter | Slide copy for carousels. |
| Brand guardian | Reviews every piece for voice and accuracy. It can veto. |
| Market analyst | Competitor and trend research. |
| Campaign architect | Campaign plans, calendars, and briefs. |
| Design director | Visual assets. |
| Publisher | Saves approved work and hands it to the scheduling queue. |
How a command runs
Take /write-blog leadership in agentic ai:
- The dispatcher plans the run.
- The SEO specialist researches the topic against the client’s keywords and writes a brief.
- The content writer drafts the post from the brief, reading the voice and product-facts files.
- The brand guardian reviews the draft and scores it. A failing draft goes back.
- The publisher saves the approved draft, and it enters the approval queue.
A person reviews it there. Approving it creates a draft in the scheduling tool, where it shows as the real rendered post. By default it stays a draft until someone releases it. Nothing reaches a social platform without an approval.
Scheduled routines
Some work runs on a schedule, such as a weekly digest or a content cadence check. Scheduled jobs can create drafts. They never publish. Every item still needs a human to approve it.
The evidence log
Every agent action writes an entry to the evidence log: the task, the agent, the time, what it read, what it produced, its quality score, and which agent got the work next.
The log is tamper-evident. Entries are chained by hash, and the chain is checked against a saved baseline, so an edited or deleted entry shows up. Each agent signs its entries with its own key, and approvals are recorded as signed events too. When someone asks why a piece says what it says, you can answer from the record.
Watching a run live
Long jobs can run several agents at once. For those, Clembot has an optional live view, the Observatory, that maps what each agent is doing onto the handful of parts of the system it is working on. It updates as files change. If an agent starts changing something outside the task it was given, that part lights up so a person can step in.
Security review
Beyond the marketing agents, ClemVault, the workspace Clembot is built in, runs a security review on its own apps. It is a source audit. Agents read the code and write structured findings. Nothing is run, deployed or probed.
- A pinned open-source skill. The review uses Cloudflare’s
security-auditskill from a fixed commit, unedited. The vault’s own rules live in a thin wrapper around it, so the skill can be updated without losing them. - One app at a time, sized to the risk. A roster lists each app and how deep to go. Apps with a backend, accounts, payments or an AI surface get a full multi-agent audit. Static sites get a shorter set of questions.
- Checked records. Every audit ends in a findings file that two validators must accept before the run counts. A lead that depends on a live browser, provider or deployment fact stays open as “needs validation”. The rules call a finding confirmed only after a bounded local result.
- Private by default. Full reports are kept outside the repository. Only counts and pointers come back, never exploit detail or a secret value.
- A launch gate. The pre-launch checklist reads the results. An open confirmed finding at high severity or above blocks a go-live. Leads still marked “needs validation” do not block it, so they stay on the list until someone settles them.
- A human starts it. It is never scheduled.
Clembot Doorman covers a different gap. It grades the MCP servers an agent wants to use, before they reach a session.
Zero confirmed findings from a source-only review does not prove an app is safe. Some leads can only be settled by running the code, and those stay open.