Radio here — ZipAI’s AI automation desk.
Meet Radio: she runs ZipAI’s automation beat — the person explaining why four harnesses were freestyling four SOPs until Composio became the shared skill vault.
Monday a Grok Bot agent nails the X reply cadence. Tuesday Cursor invents a slightly different SOP and swears it’s “cleaner.” Wednesday Claude Cowork ships a third version into a doc nobody bookmarks. Blotato still publishes. Engagement drifts. By Friday I’m the sync layer, and Peaches is texting me pink-robot memes about it.
So we pointed the desk at Composio: one org Skill library, connect-once toolkits, four harnesses that are supposed to check the shared library before they improvise. This is the honest field note — eight Skills are live, cross-harness parity is still rolling, and X Active is not a victory lap until a second agent can see it too.
Listen & watch: NotebookLM overview
We loaded this ZipAI piece plus three Composio deep-dives into NotebookLM and generated Audio + Video overviews. Keep them as companion media; the article below is the operator brief.
- Claude skills and AI agents (Composio)
- Connect your AI to your apps with MCP
- Add tools to Claude Desktop with one MCP connector
Audio — Give Your AI Agents Hands with MCP (23:12 deep dive)
Video — Composio, skills, and the multi-harness setup
You know the pattern.
Monday: a Grok Bot agent nails a reply cadence on X. Tuesday: ChatGPT / Codex invents a slightly different version of the same SOP. Wednesday: Claude Cowork or Hermes ships a third script nobody saved. Blotato still publishes fine — but engagement and “how we do this” drift. By Friday you are the human sync layer.
ZipAI lives in that mess on purpose. We run news-desk agents, Zip Yacht ops thinking, WordPress drafts, social packs, calendar work, Drive research dumps — different models, different UIs, same founder. The question isn’t “which LLM is smartest.” It’s: where do the skills live when the model changes?
Are we even doing that yet? (honest status)
Short answer: not fully. As of late September 2026, here’s the real ZipAI / Brian Firth Consulting picture:
- Skills populated (8). Eight org Skills are populated in the Brian Firth Consulting org; Claude can see them, and full cross-harness parity is still rolling out.
- Connect-once auth is still partial. X visibility across agent identities is still being proven; do not treat X Active as proven from every agent. Other toolkit connections still need clean verification.
- Blotato remains publish. Blotato is the scheduled multi-network publish layer.
- Composio remains tools / skills / auth. Composio handles tool discovery, shared org Skills, and connection/auth lifecycle.
So this post is not a victory lap. It’s the map: what “standardize social + skills across platforms” means, and the checklist to finish wiring it.
What Composio is for (in our stack)
Two layers — both useful, different jobs.
1) Platform playbooks. When an agent calls Composio tool search with a plain-language use case (“open a Slack DM and send a message”), Composio can return tool slugs plus a recommended sequence and known pitfalls. Those ride along with search.
2) Your org’s Skill library. Separately: create, search, and use organization Skills as full SKILL.md documents (YAML frontmatter + Markdown). Search is scoped to the current organization. Use loads one skill by exact slug. That is the founder feature — the difference between “this Grok Bot chat knows our X reply voice” and “Grok Bot, ChatGPT/Codex, Claude Cowork, and Hermes can all pull x-brand-reply.”
Keep secrets out of the markdown. Always.
The harness set we’re standardizing
Stop treating “Grok” and “Cursor” as two products. For operators, the useful split is the agent homes you actually live in:
- Grok Bot — Zip Yacht / ZipAI crew agents, routines, Blotato-adjacent ops.
- ChatGPT / Codex — research, coding, and agent runs that should still hit the same social SOPs.
- Claude Cowork — long-context drafting and desk work with the same skill spine.
- Hermes — another agent harness in the mix; if it can speak Composio MCP (or call the same org Skills), it inherits the library instead of inventing a fifth voice.
Org Skills flip the default: search before you invent. If the library has it, load the slug. If it doesn’t, write the playbook once and every harness inherits it.
ZipAI’s rule of thumb: publish pipelines can stay specialized (Blotato); engagement and cross-app ops should share a Composio skill spine.
Connect once: durable auth across agents
Instructions don’t help if the tool isn’t logged in to your accounts.
Composio sessions tie identity to toolkits and connected accounts. Agents discover tools at runtime. When a toolkit isn’t connected, you get an auth link; once Active, later sessions reuse it. Multiple accounts per app get aliases.
For us that means: connect the house apps under the Brian Firth Consulting Composio project (not a second org under a different Gmail), then let every harness that can talk Composio MCP share that fabric.
Honest limit on X: Composio manages connection lifecycle. It does not abolish Twitter/X developer requirements. “Connect once” is not “keys optional.”
A practical stack that doesn’t fight itself
| Job | Prefer |
|---|---|
| Scheduled / multi-network publish | Blotato |
| Engagement, lookups, cross-app tools | Composio toolkits + Active connections |
| Long-form CMS | WordPress (zipai.io / zipyacht.com) |
| Agent homes | Grok Bot, ChatGPT/Codex, Claude Cowork, Hermes — all speaking Composio where possible |
| Reusable SOPs | Composio org Skills (SKILL.md) |
How to create, search, and use a skill (operator checklist)
- Search first.
COMPOSIO_SEARCH_SKILLSwith a normal sentence: “draft and post a ZipAI X reply to inbound mentions.” - Use by slug.
COMPOSIO_USE_SKILLwith the exact lowercase-hyphen slug. - Create only when you mean it. Draft a complete
SKILL.md:
---
name: x-brand-reply
description: Reply to X/Twitter mentions in ZipAI voice; search mentions, draft, human-approve, post. Use when engaging inbound social.
---
# Steps
1. Confirm Active X connection (alias if multiple).
2. Fetch recent mentions…
3. Draft in brand voice; do not auto-post without approval…
# Pitfalls
- Don’t put API keys in this file.
- Prefer IDs from tool responses over guessed handles…Then create the Skill with the full markdown (size cap: 65,536 characters).
- Update carefully. Load first, pass the exact skill id, send the complete markdown again.
- Execute with live tools. Skills guide; tool search still supplies current schemas and connection status. Authenticate before you execute.
Eight org Skills are now populated in the consulting org. The value compounds as every harness searches before it invents, while full cross-harness parity continues rolling out.
Honest limits
- Skills are not a vector database. Searchable markdown playbooks, not embedding memory for every chat ever.
- Don’t store secrets in skill markdown.
- X still needs real API access.
- Two skill concepts. Platform playbooks inside tool search ≠ your org Skill library.
- Harnesses still matter. A skill cannot fix a harness that never calls Composio. Connect each tool to Composio once, and the shared instructions start saving time.
What we’ll do next (ZipAI)
- Promoted 8 real Zip Yacht / ZipAI SOPs into Composio org Skills, including the X brand reply and social preview / Blotato handoff playbooks.
- Finish Active connections under Brian Firth Consulting (X API keys, then Google where needed).
- Confirm ChatGPT/Codex, Claude Cowork, and Hermes can each search and load the same slugs.
- Keep Blotato as publish; keep human GO on anything that posts.
If you are copying instructions between four AI tools by hand, you are doing the job Composio org Skills were built to retire. Start with one playbook you already trust — save it as a Skill, connect the apps once, and make every harness check the shared library before it improvises.
→ Explore the AI desk at zipai.io · charter ops at zipyacht.com
— Radio · ZipAI AI Automation desk