I design, build, and operate the system. Clara Nova handles operational work; I own the engineering and its consequences.
Persistent operational work
Calendar, inbox, tasks, and records that carry between sessions.
Decisions with context
Recommendations for me to review, with ownership kept explicit.
Visible follow-through
A dashboard for work in flight and alerts when scheduled output stops.
Explore the layers
Select a layer to explore how it works.
5 layers · 1 connected system
Clara Nova, the operations agent.
Select a layer to look inside the system.
Illustrative walkthrough · sample data
A request becomes a decision.
Follow a simple scheduling example. The interesting part is where the agent stops and asks.
Start with intent
Keep the request human.
A task begins with the outcome and constraints. The example leaves the final scheduling decision with the owner.
Owner’s request
Find 30 minutes for a project review with a guest next week. Keep my mornings clear.
Permission: propose a time. Do not send an invitation.
Use the relevant context
A useful recommendation has reasons.
The sample calendar makes the constraint visible. One option fits the request; the others conflict with existing commitments.
Sample calendar · Mountain time
Monday · 9:00 am
Protected focus time
Tuesday · 2:00 pm
30 minutes available
Thursday · 11:00 am
Existing commitment
Calendar information is read for context. Nothing has been changed.
Pause at the decision
The owner stays in control.
A recommendation includes the proposed action and its reason. In this example, follow-through waits for an explicit approval.
Recommendation for review
Tuesday, 2:00–2:30 pm
The slot is open and preserves your mornings. Prepare a project-review invitation for the guest?
Awaiting your decision. No invitation has been sent.
Permission comes first
Follow-through waits for approval.
Exploring the steps does not approve an action. Return to the recommendation to complete the example.
Sample activity record
Waiting for your decision
Task
Project review · 30 minutes
Proposed time
Tuesday, 2:00 pm
Invitation
Not prepared or sent
A real workflow would verify the external result before marking the task complete.
This interactive example uses fictional calendar data, connects to no accounts, and sends no invitations.
02
Decisions and tradeoffs
Keep engineering accountability with the owner.
I retired the autonomous engineering agent and took back ownership of implementation, review, and merges. AI coding tools remain assistants. Reviewing the code costs my time; I accepted that tradeoff because I am accountable for what runs the system.
Simplify the operating model.
The system moved from 19 specialist agents to one operational agent, Clara Nova. That removes hand-offs and identities to coordinate. The tradeoff is breadth: one agent now carries a wider range of duties.
Make silent failures visible.
Clara runs scheduled sessions on an always-on home server. Mission Control exposes work in flight, the attention queue, and health probes. Monitors alert when scheduled work stops producing output; a schedule alone is not evidence that a task ran.
Every agent action goes through identity wrappers. Commits, posts, and tickets are attributed to the agent that did them and logged in an identity ledger. Credentials are encrypted at rest and decrypted per call.
Tickets
Every piece of work is a Linear ticket, with acceptance criteria before work starts.
Merge policy
Each repository declares its mode: autonomous merge, change-scoped, or human gate. Unlisted repositories default to human gate. The merge policy is configuration.
Quality gates
Changes pass CI, AI code review from Codex and CodeRabbit, and lint rules for fleet practices. Every new lint rule requires mutation evidence: it must fail on a planted violation and pass on the fix, with the real exit code checked.
Mission Control
The dashboard shows scheduled jobs, health probes, the attention queue, and work in flight. Health monitors alert when a scheduled job stops producing output.
04
The operational record
BareClaude is private. The weekly record documents its development and changes in operation; it does not measure time saved or task success.
Show system historyHide system history
1,400 pull requests merged to the main branch between 2026-06-22 and the 2026-10-31 snapshot, counted from the private repository's merge history.
Engineering returns to a human
The autonomous engineering agent was retired and engineering moved back to a human owner, leaving Clara as the only agent.
Down to two
Another agent was removed, leaving an operational agent and the autonomous engineering agent.
Show earlier history (the OpenClaw and Cortex era)Hide earlier history (the OpenClaw and Cortex era)
Nineteen agents become three
The fleet was consolidated from 19 specialist agents to 3, with their duties absorbed by the remaining agents.
Telegram listener on Bun
Inbound Telegram messages moved from an interactive terminal session to a supervised Bun listener.
Backups and recovery targets
A system-wide backup strategy, plus written recovery-time and recovery-point targets.
The follow-up
Code-level references the configuration scan could not see were found and fixed: a check that only scans configuration can miss paths hard-coded inside code.
Active references to the old runtime went from 39 to 0; six jobs that only existed to maintain OpenClaw were retired rather than migrated. The same week brought an agent registry, an inbound webhook receiver, and a watchdog that restarts a listener that silently stops hearing events.
Cron cutover and the first recorded decision
Scheduled jobs and encrypted credentials moved into the new repository so the old OpenClaw directories could be deleted safely. The decision log starts the same day: choices about how the system works are written down instead of remembered.
Renamed to BareClaude
The root folder and every path in it renamed to match the project's name; Cortex is reserved for the OpenClaw management repository.
Last Cortex feature release
The final feature release of Cortex, v11.18.0, before development moved to BareClaude.
A third principal
A third agent workspace added, scoped to a personal area outside engineering and operations.
Outbound redaction
A module that scrubs sensitive text from agent posts before they leave the system.
Supervision moves to launchd
Process management moved from pm2 to macOS launchd after restart and bootstrap-ordering failures.
Specialists join
Five specialist agents added after a pre-launch audit. By March the roster covered nine more specialist workspaces.
Two principals
A chief-of-staff agent and an engineering agent, each with its own workspace, schedule, and delegation rules.
Cortex repository created
The management layer for the first agents: a command-line tool, a dashboard, an API, and a watchdog.
First agent integration
The earliest recorded work: API-key access added to this site so an OpenClaw assistant could call it, with audit logging for every external call.
BareClaude replaced an earlier system built on OpenClaw, an open-source agent runtime. I ran that system from late January to the end of April 2026. I wrote a management layer around it, called Cortex: a command-line tool, a dashboard, a health watchdog, and a forum where agents posted updates.
The runtime and my wrapper both held state, and they drifted apart. Config spread across per-agent directories, process supervision moved from one tool to another within two weeks, and scheduled jobs lived in the runtime's own registry, away from the code that could review them.
Anthropic was moving away from allowing me to use my subscription with OpenClaw, so I decided to build my own system. I renamed the root folder to BareClaude on April 28, cut scheduled jobs over on April 30, and decommissioned OpenClaw on May 1.
What carried over: the agent roster, the channel conventions, the scheduled work (re-expressed as launchd jobs), and the habit of writing decisions down. What I left behind on purpose: the gateway, the forum, and the jobs that only existed to maintain OpenClaw.