What Paperclip Actually Is
Paperclip calls itself an agent orchestrator. The tagline is “zero human companies,” which Dota admits is partly aspirational. Sitting between fully-automated startup generators on one end and single-tab coding agents on the other, Paperclip carves out a middle lane — software for work you’re genuinely accountable for, where you define the goals, hire the team, approve the plans, and let agents carry the load between check-ins.
The interface looks like a lightweight project management tool, but the unit of work isn’t a ticket or a pull request. It’s a business goal. You tell the platform what your company should be trying to achieve, and it spins up a CEO agent to make a hiring plan, break that plan into tasks, and assign those tasks to specialists it recruits as it goes.
The mental shift Dota kept coming back to: stop thinking about the jobs you need done. Start thinking about the agents you need to hire.
Bring Your Own Bot
One of the most important architectural choices in Paperclip is that it doesn’t lock you into a single model provider. Dota calls it “bring your own bot.” If you already pay for Claude Code, Cursor, Codex, or OpenCode, Paperclip will use those subscriptions. If you’d rather pay per token through OpenRouter, that works too. And because OpenRouter occasionally hosts frontier-quality models for free, anyone willing to pay attention can run cheap work on free inference and reserve premium models for tasks that actually benefit from frontier reasoning.
The recommendation Dota gave for serious users was unambiguous: put a frontier model behind your CEO agent. The CEO is the bottleneck for almost everything — planning, delegation, organizational changes — so a smart CEO multiplies every agent below it. Workers can run cheaper.
The Memento Problem
There’s a moment in the walkthrough where Dota draws an analogy that reframes the entire conversation about agent design. “Have you ever seen Memento?” he asks. “Your AI agents are Momentum Man.”
The reference lands because it’s accurate. Agents wake up every single run not knowing who they are, where they are, or what they were doing ten minutes ago. They know how to code. They know how to search. They know how to spend money. They don’t know they work for you, and they don’t know what success looks like for the task on their desk.
That’s why Paperclip bakes in personas, memory systems, and what it calls a heartbeat — a checklist the agent runs through on every wake-up. Fetch who you are. Confirm your role. Read today’s plan. Find your assignments. Do the work. Extract new memory. The heartbeat is the tattoo on the arm. It’s the Polaroid. It’s the thing that stops an otherwise capable agent from wandering off and burning tokens on something nobody asked for.
Under the hood, Paperclip’s default memory system borrows from Tiago Forte’s PARA method, implemented as a simple file-based store. That’s enough for most work, and it keeps things transparent — you can open the files and see what your agents remember about you, themselves, and the business.
Skills, Issues, and the Concurrency Trap
Everything that happens inside Paperclip is tracked as an issue. An agent asks permission to hire? Issue. A QA reviews a feature? Issue. A content strategist drafts a Discord post? Issue. The reason this matters isn’t bookkeeping — it’s the concurrency rule. Paperclip enforces that each task can only be handled by one agent at a time, which sounds small but eliminates an entire class of “wait, which Claude Code window did what?” problems that make running twenty parallel agents untenable.
Dota described building Paperclip in the first place because he was running twenty or thirty Claude Code windows over weekends, coming back Monday with a blown token budget and no memory of what any of them accomplished. Paperclip’s monthly spend tracker and issue log are the direct answer to that pain.
On top of the orchestration layer, agents can be augmented with skills — reusable toolkits published on places like skills.sh. The Remotion skill for generating animated videos came up repeatedly in the walkthrough, and the pattern is simple: paste the skill URL into Paperclip, attach it to an agent, and that agent suddenly has new capabilities. The only real caveat is the one that applies to any open registry: treat new skills with the same care you’d apply to any third-party package. Check the stars, check the security badges, and read what you’re installing.
Routines: Where Orchestration Becomes Automation
The feature that started to feel most directly useful was routines. A routine is an issue template that runs on a schedule. Dota demonstrated it by creating a daily Discord update — every morning at 10 AM, a content strategist agent reads the last twenty-four hours of GitHub merges, celebrates community contributors by name, and drafts a message for the Discord.
That’s the same pattern countless teams pay for cron-triggered Zaps or Make scenarios to handle — except instead of a rigid node graph, you’re giving a briefed agent a rolling assignment it can actually exercise judgment on. Miss a day because of a weekend? The agent can decide there’s nothing worth posting. Big release shipped? It can rework the tone to match. You still get the traceability, because each run of the routine creates its own issue you can audit later.
The Obvious Application Nobody’s Talking About Yet
Halfway through the walkthrough, Dota mentioned something offhand that deserves its own section. The users finding success with Paperclip today aren’t running fully autonomous companies. They’re people with existing businesses — marketing agencies, dental foundations, security consultancies, roofing companies — who are using agent orchestration to extend their team capacity rather than replace it. One of the quiet superpowers is marketing. Specifically, social media.
Think about what social media management actually is at the operational level. Every day there’s a queue of posts that need to be written, reviewed, scheduled, and published across a dozen platforms. Each platform has its own rules, character limits, media requirements, and audience voice. Half the job is mechanical. The other half is judgment — what to post, when, and in what tone. That’s the exact shape of problem Paperclip is designed to decompose.
The gap, until recently, was the last mile: once an agent has drafted and approved a post, how does it actually hit Instagram, LinkedIn, TikTok, YouTube, Reddit, and everything else without a human clicking publish twenty-eight times? That’s where Postiz comes in.
Wiring an Agent Team Into Postiz
Postiz is an AI-powered social media management platform that handles scheduling, cross-posting, and analytics across 28+ channels — X, LinkedIn, LinkedIn Pages, Reddit, Instagram, Facebook, Threads, YouTube, TikTok, Pinterest, Discord, Mastodon, Bluesky, and more. What makes it especially useful in a Paperclip-style setup is that Postiz exposes an MCP server, a public API, and a CLI — three separate interfaces your agents can call directly, no clicking required.
The cleanest integration path is MCP. Any MCP-compatible agent — Claude Code, Cursor, Codex, the agents Paperclip orchestrates on top of them — can reach Postiz by pointing at a single URL:
URL: https://api.postiz.com/mcp/your-api-key
Drop that into your Claude Desktop config, Claude Code config, or Cursor’s MCP settings, and the agent instantly gains eight tools: integrationList, integrationSchema, triggerTool, schedulePostTool, generateImageTool, generateVideoOptions, videoFunctionTool, and generateVideoTool. In plain terms, the agent can list the user’s connected accounts, learn each platform’s posting rules, fetch dynamic data like Reddit flairs or YouTube playlists, generate images and videos, and schedule the final post — all through structured tool calls, with no HTML scraping or brittle automation.
If your Paperclip setup prefers command-line agents, the Postiz CLI works the same way. After authenticating once with postiz auth:login, a social media agent can do the entire publishing loop in a handful of commands:
# Discover connected accounts
postiz integrations:list
# Inspect platform rules before drafting
postiz integrations:settings <integration-id>
# Upload media (required for TikTok, Instagram, YouTube)
postiz upload thumbnail.jpg
# Schedule the post
postiz posts:create \
-c "Your content here" \
-s "2026-05-01T09:00:00Z" \
-m "https://cdn.postiz.com/uploads/abc.jpg" \
-i "integration-id"
# Check analytics after a few days
postiz analytics:post <post-id> -d 7
For a multi-platform campaign with different content per channel, a JSON payload works cleanly and plays well with agent-generated output:
postiz posts:create --json campaign.json
The JSON carries a list of integrations, per-provider posts, and platform-specific settings — the kind of structured object an agent can produce reliably when given a schema.
What a Social Media Org Chart Looks Like
If you were to translate Dota’s agent org chart into a social-first company, the shape writes itself. A CMO agent owns the calendar and brand voice. A content strategist agent plans weekly themes and topic clusters. A copywriter drafts the post variants for each platform, respecting character limits exposed by integrationSchema. A designer runs generateImageTool or a Remotion skill to produce visuals. A scheduler calls schedulePostTool to push everything live. A QA agent checks each scheduled post the night before against the brand guide stored in memory. An analyst agent pulls weekly reports with analytics:platform and feeds the numbers back into the content strategist’s plan.
None of that requires Paperclip specifically — you could glue it together with custom scripts — but orchestration is exactly what Paperclip exists to make easy. You assign a goal (“grow our LinkedIn following to 10K this quarter”), the CEO agent drafts the plan, and a chain of specialist agents executes it, each one logging its work as an issue you can audit whenever you want.
Shared Companies and the Coming Import-Export Economy
The most surprising part of the walkthrough was a feature Paperclip is shipping around the idea of sharing entire companies. Not just agents. Companies. A full org chart with roles, skills, routines, and memory, packaged so another user can import it into their own Paperclip instance.
Dota walked through examples — Gary Tan’s GStack, which replicates his office-hours style; a game studio preset with a creative director, a producer, and a technical director already wired to produce Godot bullet-hell games; superpower repos with a hundred agents on tap. Importing a company doesn’t copy the skills locally; it references the remote repos so updates propagate. The implication is that someone, right now, is building the first proven TikTok marketing agency as a Paperclip export. When that ships, anyone who imports it starts with an already-debugged team instead of hiring from scratch.
That future, plugged into Postiz’s publishing layer, starts to look less like “AI takes your job” and more like “anyone can acquihire a small agency overnight.”
The Part Where AI Still Can’t Replace You
Dota said something near the end of the conversation that’s worth underlining: AI can do everything except know your values. The frontier models have capability. They don’t have taste. They don’t know that the header pills on your app should be the same height, or that your brand voice is warm but never cute, or that your audience hates hype. That part still has to come from you — written down, codified in skills and heartbeats and persona prompts, reviewed when agents drift, and corrected as a rule you save once rather than a note you give twice.
The art of running an agent team, in other words, is the same art that’s always defined good leadership. Clearly communicate what you want and why it matters. Give people — or agents — enough context to make calls without you. Review what comes back and push it back with specifics. What’s changed is the vehicle. What hasn’t is the skill.
How to Start This Week
If the walkthrough made any of this feel tractable, the path forward is shorter than it sounds:
- Install Paperclip locally and spin up your first company around a single, narrow goal. Don’t try to run the whole business on day one. Try one workflow — a weekly content report, a competitor monitoring routine, a daily social media publishing loop.
- Put a frontier model behind your CEO agent. Cheap inference is fine for workers, but the CEO is where planning quality compounds.
- Write your brand guide and link it. Agents can read URLs. If your taste lives in a Notion page or a public doc, your design and content agents will use it.
- Connect Postiz as the publishing layer. One API key turns any agent in your org chart into a scheduler that reaches 28+ platforms. Use the MCP server for in-chat agents, the CLI for scripted workflows, and the public API if you’re building something custom.
- Start with routines, not heroics. The highest-leverage wins are recurring tasks you currently do by hand — weekly reports, daily posts, monthly audits — that a briefed agent can handle on autopilot while you review the output.
Agent orchestration is genuinely new territory. A lot of the patterns aren’t figured out yet, and anyone who tells you otherwise is selling something. But the combination of a tool that organizes agents the way you’d organize people, and a platform that lets those agents actually reach your audience without human-in-the-loop clicking, is already a meaningful unlock.
Try Postiz
If you want to give your agents a publishing layer that actually works across every channel that matters, Postiz is the fastest way to get there. Schedule across X, LinkedIn, Instagram, TikTok, YouTube, Reddit, Threads, Pinterest, Discord and more — from an MCP server, a CLI, or a clean public API. Start your free trial on postiz.com and connect your first agent in under ten minutes.
