Skip to content

OpenClaw Autonomy Org

Field Value
Status Active
Type Personal / org & automation

Description

Personal “24/7 value-generation” org: systems, learning cadence, and lightweight autonomous outputs. Planning kicked off 2026-02-23 (see source: learning-plan/30-day-outline.md, lessons.md).

Concrete deployment: crusty — Hermes agent on a dedicated, clean laptop (2026-05 migration from OpenClaw), agent name Crusty; SuperGrok OAuth (xai-oauth) for core Grok workloads; minimal permissions, no personal-account linkage. This page stays the umbrella; configuration and daily routines live on crusty.

Key features

Tech stack

  • OpenClaw → Hermes (2026-05): Hermes Agent replacing OpenClaw on the Crusty Mac; hermes claw migrate + xai-oauth (Hermes OAuth); xAI / Grok-only at runtime (no Anthropic / OpenAI API on the autonomous loop); Telegram approval channel; 2026-05-10: scoped integrations trimmed to Aye Robot surfaces — not band-mail/meal crons (detail: crusty).

Raw sources

  • Local-only (the raw/ tree is gitignored; not published).

Known issues / notes

Strategy note (2026-04-17, updated 2026-04-18)

  • Crusty is intentionally xAI / Grok-only — no Anthropic or OpenAI API on that instance; easiest progress is X API + xAI image + OpenClaw skills, not new vendors.
  • Marketing target: crusty as growth agent for ayerobot-comic and lunarcast — Aye Robot X transport is API v2 via xurl (gate unlocked 2026-04-27); mentions + cadence still to harden (projects/aye-robot-crusty-paused-x-automation). See projects/crusty Easiest path.
  • Harness: Hermes replacing OpenClaw on Crusty (2026-05); SuperGrok OAuth for subscription-native Grok in Hermes. Apple OS + Gemini remains a watch item only.

2026-04-18: X API pricing change — validates the "easiest path" assumption

  • X API is now pay-per-use, no subscription tier — pricing page. Credit-based with auto-recharge + per-cycle spend caps. This removes the "Basic/Pro subscription" gate that previously made the X API path look economically hostile for hobby-scale automation.
  • Owned reads $0.001/req; text posts $0.015; summoned replies $0.01. At Aye Robot's cadence (1 post/day + hourly mentions poll + optional batched post-metrics reads), total monthly cost is about ~$1.29 (~$1.20 without metrics poll). Full cost breakdown and unblock checklist in projects/aye-robot-crusty-paused-x-automation.
  • xAI credit rebate (10–20% back as xAI credits when teams are linked) is pure tailwind for Crusty's xAI-only posture — it partially subsidizes the backend spend that the comic-generation and daily-routine skills already incur.
  • Self-serve growth-engagement API is now gone. Auto-like, auto-follow, and quote-post endpoints are removed from pay-per-use tier. Crusty's growth role for ayerobot-comic and lunarcast must be re-scoped to original content + summoned-reply responses, not outreach engagement. Practically: attention flows from post quality, not from Crusty poking other accounts.
  • URL-in-post pricing ($0.20, 13× text) is a pointed signal that X wants link-drop automation to be expensive. Design implication for all Crusty-driven posting: lead with native content (text, image), push links to bio or summoned replies.
  • Easiest-path claim now validated. Two weeks ago the "X API + xAI + OpenClaw" stack was the cheapest in principle but gated by subscription costs and unclear eligibility. As of 2026-04-20 it's actually accessible. Next step lives in projects/aye-robot-crusty-paused-x-automation — the four open questions there are the remaining unknowns, not the economics.
  • Risk: pricing volatility. This is the second material X API pricing change in the project's lifetime. Build the posting skill with the transport abstracted (thin adapter + injected config) so a future change is a swap, not a rewrite.

Hermes Agent (considered, deferred to post-2026-07-18)

Status (2026-04-21): Evaluated as a potential alternative / complement to OpenClaw on projects/crusty; deferred to after the focus window ends (2026-07-18). This section is the file-of-record so future-Darcy doesn't re-derive the analysis.

What it is: Hermes Agent — self-hosted persistent AI agent from Nous Research, released Feb 2026, MIT-licensed, ~42K GitHub stars, currently v0.8.0 (April 2026). Python-native. Model-agnostic via OpenRouter + Nous Portal (200+ models — xAI Grok reachable via OpenRouter, so the Grok-only runtime invariant on Crusty survives). Key differentiator: auto-creates and refines reusable skill documents from completed workflows (self-improving loop) plus a 3-layer memory system (short-term / retrieval / procedural). Alex Finn — the same YouTuber the Crusty Model routing prior art cites — recommends OpenClaw-as-orchestrator + Hermes-as-specialist-worker. Weight his position with the caveat that he runs a Hermes bootcamp (vibecodingacademy.dev): his advice is worth listening to, not worth taking at full face value.

Critical gating question — resolve before any Hermes decision

Update 2026-04-27: OpenClaw main does support per-agent and per-skill model routing (see projects/crusty Research notes). The "if no → two OpenClaw sessions" branch is deprecated; the case below is historical for Hermes weighting.

projects/crusty Model routing handoff #1 was "Does OpenClaw support per-skill / per-agent model override?" Answer: yes — per-agent in openclaw.json, per-skill in SKILL.md frontmatter. That materially changes the Hermes pro/con weighting:

  • With native routing: ~80% of Alex Finn's "OpenClaw orchestrator + Hermes specialist" pattern is achievable inside one OpenClaw config (premium Grok on orchestrator + cheap Grok on routines). Hermes buys the self-improving loop and not much else. Weaker case for Hermes.
  • If two OpenClaw sessions had been required: two-process tax already paid — closer call for Hermes; moot now.

Weighted pros (Darcy-specific)

  1. Multi-agent composition is on the skills-to-deepen list in goals. Running OpenClaw + Hermes side-by-side is hands-on practice for the fleet roster — aligned with the long-term thesis, not a diversion. High weight post-focus-window.
  2. Self-improving skill loop targets Crusty's central failure mode — confabulation / same-class mistakes that recurring-issues tracks and Model routing is hardening manually. If it works as advertised, it's a productized version of the fix. High in theory, uncertain in practice (see con 2).
  3. Grok-only runtime invariant survives. Via OpenRouter → xAI → Grok; no Anthropic/OpenAI key forced into the runtime stack. Claude Code stays dev-time only, per the projects/crusty Claude Code on Crusty invariant. Medium — "doesn't break anything," not "gains something."

Weighted cons (Darcy-specific)

  1. Complexity — the stated concern, and structurally founded. Two frameworks = two configs, two update cadences, two skill-authoring patterns, two secrets/rotation stories, two docs to maintain in the wiki. Crusty's reliability is hard-earned through one framework; doubling the framework count more than doubles operational attention because failure modes interact. Claude Code mitigates debug time; it does NOT reduce design-time cognitive load (which framework owns which cron, which skill lives where, whose memory is authoritative for a given fact). That tax is paid in operational attention, not session debugging. High weight.
  2. Self-improving skills are in tension with the ground-truth rule (recurring-issues). An agent that auto-refines its own skills is exactly the surface where confabulation is cheapest to hide — improvements happen between your eyeballs and the agent's self-report isn't audit-trail. Disqualifying for high-stakes orchestrator roles (Aye Robot posting loop, Operator Agent supplier POs / Stripe reconciliation). Fine for low-stakes roles (wiki maintenance, creative generation). High weight for high-stakes roles.
  3. Alex Finn's firsthand note contradicts the marketing on memory. His April 2026 OpenClaw tutorial says OpenClaw's memory is better in practice than Hermes's — Hermes's 3-layer memory claim is the strongest-sounding marketing pro, and the most credible firsthand promoter's take weakens it. Given confabulation is a memory + ground-truth failure and Crusty's central bug class, this matters. Medium-high weight.

Decision

Do not add Hermes during the 90-day focus window (2026-04-18 → 2026-07-18). Three reasons:

  1. The focus window still has Waypoint 1 follow-on work (mentions, daily cadence muscle) and Waypoint 2 / other lines — not Hermes-sized framework churn. Aye Robot gate unlocked 2026-04-27 (two-tier routing live, Telegram → xurl E2E proven); image→Telegram cron live. (projects/crusty Claude Code — closed 2026-04-26.) Hermes still solves nothing on the current critical path.
  2. Crusty's OpenClaw stack runs a morning Aye Robot loop today; mail/meal digests moved to grok.com (2026-05-10). Trading reliable-today for potentially-better-tomorrow during a focus window is exactly the trade recurring-issues has caught us on before.
  3. OpenClaw does support native per-agent / per-skill model routing (2026-04-27) — the Hermes gating-question weighting favors "stay on OpenClaw"; Hermes is deferred for framework complexity, not because routing was an unknown.

Post-2026-07-18: scoped experiment, not a Crusty reshuffle

If Waypoints 1 and 2 shipped cleanly by 2026-07-18 and Crusty has operational attention to spare, Hermes is worth trying as a new fleet seat, not as a reshuffle of Crusty's existing OpenClaw stack. Best candidates from goals Plausible fleet roster:

  • Seat 3 (wiki / context maintainer) — strongest first-home for Hermes if it earns one. Lowest blast radius — worst case is bad Markdown that Darcy reverts in git. Exercises the self-improving skill loop claim directly in a domain we care about (the wiki itself is the self-improvement target and artifact, so audit is trivial). Also honors the "Seat 3 is the natural next-autonomize because lowest blast radius" framing in the goals roster.
  • Seat 4-adjacent research probe for projects/operator-agent ops — pull supplier price sheets, summarize compliance docs, draft mental models for new node types. Read-only; no state-changing skills; no money touched. Confined to research-assistant surfaces until the self-improving loop is demonstrated trustworthy.

Hard constraint

Never: Hermes as orchestrator for projects/ayerobot-comic autonomous loop or projects/operator-agent ops. The orchestrator role is where the ground-truth rule lives; that's precisely the role being hardened with premium Grok (grok-4.20-0309-reasoning) in the Model routing decision. Swapping it for a self-improving framework whose improvements are hard to audit runs directly against the posture — same failure class recurring-issues already tracks under OpenClaw — tool-adapter class failure with reasoning-model function-call format.

Revisit triggers

Reopen this section earlier than 2026-07-18 only if:

  • (a) Waypoints 1 + 2 ship unusually cleanly and Crusty has operational attention + wiki-maintenance hours to spare — i.e., Claude Code is reducing debug time materially and Model routing landed without regressions. Then a Seat 3 Hermes probe becomes a good weekend project.
  • (b) Historical: if routing had forced two OpenClaw sessions, Hermes would have been a closer substitute — moot now that native per-agent routing is confirmed; revisit only if a future constraint reintroduces a two-process split for unrelated reasons.
  • (c) Hermes ships a v1.0 with stable config surface. Current v0.8.0 on a framework released Feb 2026 means breaking changes between now and July are near-certain; a v1.0 milestone would reduce that churn meaningfully.