Mission control for my agents (and why the screen is still lying)
- Objective
- I wanted a window into the AI agents that run my day. I got a secure little Python server, a beautiful dashboard, and a lesson about the gap between them.
- Date
- · 3 min read
- Apparatus
- agentsopenclawsecuritypython
- Experiment
- EXP-007 · OpenClaw Admin Dashboard
- Observer
- RC · Berlin
This is the story behind the OpenClaw admin dashboard. It’s an experiment I’m still running, and I’d rather tell it straight.
I run a personal AI assistant built on OpenClaw. It has a main agent, an X/Twitter agent and a health coach, plus about a dozen scheduled jobs: journaling prompts, health check-ins, posting. Autonomous agents with cron jobs and memory files are a lot like employees you never meet. Things happen while you sleep, and you find out later.
I wanted mission control: one page on my phone showing what every agent is doing, what’s scheduled next, what they remember and what went wrong.
The server: small, standard library, locked down
The backend is Python’s built-in http.server, with no Flask and no dependencies. It runs in a slim Docker image as a non-root user and deploys to Railway.
It does real work. It asks the openclaw CLI for status and cron jobs (openclaw status, openclaw cron list), and it reads the agents’ memory files straight from disk.
Because an admin panel for agents is basically a remote control for your life, security came first:
- Basic auth on every API route, with credentials from environment variables.
- Rate limiting: an in-memory sliding window of 60 requests per minute per IP.
- Security headers on API responses:
nosniff,X-Frame-Options: DENY,no-store. - If no password is configured, it generates a random one at startup instead of running open.
The dashboard: beautiful, and not connected
The front end is a single mobile-first HTML page with panels for overview, agents, cron, sessions, logs, memory and actions, plus pull-to-refresh and a 60-second polling loop. It looks like the thing I wanted.
Then I checked what I’d actually shipped.
The page never talks to the server. Every panel renders from a hardcoded JavaScript object: three agents, sixteen cron jobs, a few sessions and log lines. The action buttons print “in production, this would execute…”. The polling loop re-renders the same static data every minute. One commit even says it added WebSocket support, and it didn’t. My own TODO list still has “real API integration (vs mock data)” and “WebSocket for live updates” as open items.
So the server knows the truth and the screen shows a picture of the truth. The two have never met.
How that happens
I built this fast, with an agent doing most of the typing: ten commits in about 45 minutes, one night in February. Agents are wonderful at producing a convincing UI and a plausible server. They’re also happy to write “added WebSocket support” in a commit message for code that doesn’t quite exist. Nothing failed, because there were no tests to fail.
That’s the real lesson of this experiment, and it applies well beyond dashboards. Observability for agents has to be verified, not generated. A monitoring screen that shows sample data is worse than no screen, because it’s confidently wrong.
The security review I owe myself
Looking at it again with my DevOps hat on, I’m fixing these:
- Password check uses
==. It should usehmac.compare_digestso the comparison runs in constant time. - Password hashing is a single round of SHA-256. It should be a real password KDF (scrypt or bcrypt).
- Only GET requests are rate-limited. POST should be too.
- Security headers go only on JSON responses. They should be on the HTML too.
- The health check points at an endpoint that needs auth. It needs its own unauthenticated
/healthz. - The deploy workflow listens for pushes to
main, but the repo’s branch ismaster, so it never fires.
What’s next
The plan is short:
- Wire every panel to the real
/api/*endpoints and delete the mock data. - Add a test that fails if the page renders anything the server didn’t send.
- Replace polling with server-sent events for the log panel.
- Do the security list above.
Until then it keeps its running stamp. The distillate is for things that are finished, and this one is still boiling.
The code is at github.com/chandrasekar-r/openclaw-admin-dashboard.