Research memo · 14 May 2026

Agent tooling inspiration for Hermes + Cloop

Pi’s package ecosystem is becoming a fast-moving bazaar of small, installable agent affordances. The useful lesson is not “switch to Pi”; it is to make Hermes feel more like a personal operating layer: delegated workers, thin CLIs, auditable memory, browser-auth utilities, and artifacts that turn agent work into something inspectable.

Pi packagesOpenClaw-compatible skillsSteipete reposNico Bailon reposHermes skills

Bottom line: Hermes already has the right primitives — tools, skills, memory, cron, messaging, delegation, browser, GBrain. The next leap is productizing them as small named capabilities with visible state: a delegation cockpit, context/truth adapters, browser-auth capture tools, and a personal package layer.

What I looked at

Pi ecosystem

Pi package catalog, awesome-pi-agent, npm package metadata for memory/delegation/workflow extensions.

Notable: installable extensions, skills, prompt templates, themes via pi install npm:<package>.

Nico Bailon

pi-subagents, pi-mcp-adapter, pi-messenger, surf-cli, visual-explainer.

Steipete

summarize, agent-scripts, oracle, CodexBar, claude-code-mcp.

Patterns worth stealing

Delegation is a first-class UX, not just a backend tool

pi-subagents makes child agents plain-language addressable: reviewer, oracle, scout, parallel reviewers. pi-messenger adds presence, file reservations, status, feed, and stuck detection.

Hermes implication: keep delegate_task, but add named crews, persistent job handles, task artifacts, and a status surface Connor can inspect from Telegram.

MCP needs a lazy, searchable adapter

pi-mcp-adapter attacks the core MCP problem: hundreds of verbose tool schemas burn context. It exposes one proxy tool, discovers tools on demand, caches metadata, and lazy-connects servers.

Hermes implication: build/use a “thin MCP gateway” mode where Hermes sees search/describe/call, not the entire universe of MCP schemas.

Memory is becoming governed, local, and evidence-based

Pi packages like pi-persistent-intelligence, pi-semantic-memory, titan-pi-memory, and pi-continuous-learning emphasize typed records, evidence, confidence, review cadence, supersession, local search, and earned instincts.

Hermes implication: GBrain should be surfaced as a governed “truth interface”, with fact candidates, supersession, provenance, and review rather than a magic memory blob.

Browser auth is a capability multiplier

surf-cli and Steipete’s oracle/summarize lean into browser sessions: use existing logins, capture network requests, automate ChatGPT/AI Studio/YouTube-like surfaces, and avoid API-key-only thinking.

Hermes implication: browser-auth tools are not a nice-to-have; they are the answer for YouTube blocks, authenticated research, dashboards, and private web apps.

Agents need observable interactive processes

pi-interactive-shell has modes for interactive, hands-free, dispatch, and monitor. User can watch, take over, and get wakeups only when meaningful events happen.

Hermes implication: Hermes already has background processes, but should expose “observable task sessions” with logs, notifications, and structured completion summaries.

Artifacts beat terminal output

visual-explainer turns diffs, plans, diagrams, and recaps into styled HTML. Steipete’s summarize is media-aware and cache-aware.

Hermes implication: keep defaulting substantial outputs to docs.cforsyth.com. Add specialized artifact generators for audits, delegation reports, memory reviews, and context maps.

Comparison matrix

Delegation

Named subagents, background jobs, parallel reviewers, inter-agent chat, file reservations.

Build a Hermes delegation cockpit: named agent presets, job IDs, artifacts, review gates, Telegram status.

Skills

Pi packages make installable skill bundles feel lightweight and social. Steipete keeps cross-agent skills in one canonical repo.

Create a personal skill registry: install/update/validate skills; support repo-owned symlinks; make skills emit verification checklists.

CLIs

Small CLIs are preferred over huge MCP surfaces when possible: summarize, surf, CodexBar, oracle.

Favor thin, inspectable CLIs for Cloop context: email, calendar, YouTube, browser auth, GBrain, market, CRM.

Context

Dynamic pruning, boomerang summaries, lazy MCP metadata, session search, semantic memory.

Add context budgets and source accounting: what was loaded, why, cost, provenance, and what was deliberately ignored.

Truth/usefulness

Memory packages increasingly store evidence, confidence, supersession, review cadence.

Expose GBrain as “truth ledger”: candidate → reviewed → canonical, with source and confidence visible to Hermes.

Human interface

Status bars, dashboards, package catalogs, visual explainers, observable PTYs.

Telegram + docs + dashboards: every long task should have a live status and a final inspectable artifact.

Recommended Hermes upgrades

1. Delegation cockpit

Create named presets — builder, reviewer, scout, oracle, ops, memory-curator. Each run returns: objective, files touched, evidence, tests, blockers, confidence, next action.

Use existing delegate_task for short leaf work.
Use spawned Hermes/Codex/Claude for long-running jobs.
Persist summaries to GBrain and optionally docs artifacts.

2. Thin MCP / CLI adapter layer

Do not let every MCP server dump schemas into context. Build a single adapter shape: search_tools(query), describe_tool(id), call_tool(id,args). Prefer CLIs when a compact CLI is enough.

3. GBrain truth console

Add a review layer over candidate facts: provenance, speaker/source, confidence, supersession, review cadence. Make “what is true/useful?” a queryable surface, not an implicit memory prompt.

4. Browser-auth utilities

Install or adapt a Surf/Browser Harness style local browser controller. This directly helps YouTube transcription, logged-in SaaS dashboards, Cloop QA, authenticated web research, and network/API discovery.

5. Artifact generators

Formalize dark HTML outputs for recurring agent work: cron audits, PR reviews, delegation run reports, GBrain fact reviews, context maps, market digests.

6. Continuous learning with curation

Borrow the “earned instincts” idea: observe repeated corrections and successful workflows, create candidate skills/rules, but require evidence and review before hot-prompt injection.

Practical roadmap

Now: 1–2 days

Finish cron audit: retire old cloop-ears assumptions, unify summaries around audio-memory + GBrain.
Add a “delegation run report” HTML template.
Install/evaluate Surf or Browser Harness on a machine with useful browser auth.
Add YouTube blocked-video fallback docs to youtube-memory.

Next: 1–2 weeks

Build Hermes named-agent presets and job ledger.
Implement thin MCP adapter proof-of-concept.
Create GBrain candidate review dashboard.
Make skills validation/update a cron or command.

Later: 1 month

Personal package registry for Connor’s tools/skills.
Multi-agent workspace with file reservations and stuck detection.
Context budget dashboard: token spend, loaded sources, ignored sources.
Continuous-learning candidate pipeline.

YouTube transcription tangent

Finding

Pure CLI transcript extraction works when captions or media download are accessible. For the blocked test video, multiple CLI/API approaches failed from this VPS because YouTube demanded sign-in / anti-bot verification. The durable solution is not another freemium transcript API; it is a tiered fallback:

1. captions first via youtube-memory.
2. yt-dlp + faster-whisper when media download is available.
3. authenticated browser/cookies or residential/local machine when YouTube blocks the cloud IP.
4. Browser Harness/Surf-style browser auth to make that repeatable and observable.

Important policy already implemented: external video speech is candidate evidence only, not Connor’s belief.

One architectural sentence

Make Hermes the operator, GBrain the truth ledger, browser-auth CLIs the reality interface, and skills/delegation the compounding execution layer.