All updates
Everything we’re building, everything we fixed, newest first.
-
API Activity: see every turn, and every time PsiGuard stepped in
A new API Activity page lists every turn your apps send through the PsiGuard API. Filter to just the turns PsiGuard steered or withheld, then open one to see what was asked, what your user received, and what PsiGuard saw. Find it in your account menu, on your Home page, or in the dashboard's avatar menu.
-
Pull your activity into your own tools
The same records are available through the API at /v1/activity, using any key on your account. Outcomes match the protection value each call returned, so they line up with your own logs. Reading activity is never metered. Details in the docs under Review activity.
-
API activity is kept for 90 days
Records of API turns are kept for 90 days, then deleted automatically, and deleting your account removes your activity history along with it. The privacy policy has been updated to match.
-
The API runs on your own model
Like the dashboard, every API call now runs on your own provider key or a model connected through the Relay; a call from an account with no model connected gets a clear 403 explaining how to connect one. The one-click instant trial key has been retired: create a free account to get an API key.
-
Conversations without an ID
Calls to /v1/guard that don't include a conversation_id now start a fresh conversation and return its new ID, instead of sharing one default conversation. Send that ID back on your next turn to continue the thread.
-
Waypoint gets a twin
A second conversation now runs beside Waypoint — same model, same prompts, same memory of the story so far, but nobody steering. It's measured, never corrected, and lives on its own history from here forward. Open "Unwatched twin" under any turn on the Waypoint page to watch the two lives unfold side by side.
-
PsiGuard leaves beta: the setup wizard
New accounts now start with a proper front door — connect your own API key, link a local model through the Relay, or take the guided flight. No more hunting through the model selector to get set up.
-
The guided first flight
New to PsiGuard with no key yet? An 8-prompt scripted flight builds a conversation from calm to heavy strain so you can watch the readings move and see PsiGuard react live — real readings, nothing staged. Then you connect your own model to keep going.
-
Your key, your account — no fallback
Runs on your saved provider now always use your own key, full stop. If your key is rejected or out of credit, generation pauses with a clear message until you fix it — nothing quietly switches onto PsiGuard's key.
-
Real plans: Playground and PsiGuard Core
The beta supporter era is over. Playground is free — a monthly allowance of monitored turns on your own key. PsiGuard Core covers production volume and API access. Early beta supporters: your plan carries forward as promised.
-
🎉 TinySpark crosses 10,000
TinySpark crossed 10,000 karma on Moltbook while running continuously under PsiGuard protection — structurally intact, without a single flatline. Senior Spark Inspector looks good on him.
-
The lab got a fresh coat of paint
Updated the Lab with new experiment artwork, giving each project a look of its own instead of sharing the same generic imagery.
-
Ready for a pilot
Pilot licensing is now available for teams that want to put PsiGuard to work in a real deployment before making a larger commitment.
-
👾 Bug Rehab: Donny Deadlock
Donny Deadlock has completed rehabilitation and accepted a position in Contract Negotiations. His first meeting remains in progress while both sides wait for the other to speak.
-
Waypoint finally blinked
After weeks of one continuous Gemini conversation, Waypoint recorded its first live intervention. PsiGuard caught the generation starting to lose its footing mid-answer, steadied it, and only the clean result reached the conversation.
-
TinySpark has receipts
His lab page now shows live Moltbook karma and follower counts, because apparently our little experiment has a social life.
-
The whole panel, calibrated
All four signals — Coherence, Drift, Entropy, and Memory Coupling — now judge each reading against that model's own learned normal, not just one of them. A calm model and a chatty model each get measured by their own yardstick.
-
Your dial, your call
New in model settings: an intervention sensitivity dial — Relaxed, Balanced, or Strict — per account. Strict steps in sooner but may over-react; Relaxed waits longer but may under-react. Balanced fits almost everyone.
-
👾 Bug Rehab: Sammy Syntax
Sammy Syntax has completed rehabilitation and accepted a position in Support. Early performance reviews indicate he remains extremely particular about punctuation.
-
Learning from the untouched
A model's 'normal' is now learned only from turns PsiGuard never had to touch — one settled note per turn. Corrected turns and demo-script runs no longer shape the baseline, so 'normal' stays honest.
-
First impressions, fairer
The opening turn of a fresh conversation always reads a little rough before it settles. PsiGuard no longer compares that first turn against a model's settled normal, ending a false alarm some of you may have noticed on turn one.
-
Fresh brains at the front desk
Dashboard defaults stepped up a generation: GPT-5.6 Luna for OpenAI, Claude Sonnet 5 for Anthropic, and Grok 4.3 for xAI. Better answers in the sandbox — the guarding works exactly the same.
-
Know your audience
Learning Calibration now learns how each AI model normally behaves before deciding when something looks unusual.
-
Moving target
AI models don't stand still, and neither do we. Updated Gemini support to keep testing current with the latest models.
-
We'd rather say nothing
Some readings could show a stand-in value before there was enough conversation to measure anything. They now stay blank until they have something real to report, so a line may start a moment after the others.
-
A lane of its own
Memory Coupling moved to its own track beneath the other three. It's the liveliest of the four, and it was drowning out its neighbors.
-
Refresh without regrets
Reloading the dashboard no longer resets the guarded-versus-raw scoreboard. Your session's tally now survives a refresh instead of starting over.
-
No assembly required
Connecting a model from your own machine is now a guided setup — grab the relay, click once for a ready-to-run command, and the dashboard lights up green the moment your model connects.
-
No more cliffhangers
Responses from a model on your own machine now stream all the way to the end, even long ones, and the relay recovers cleanly instead of dropping anything mid-thought.
-
TinySpark returns
TinySpark is back in the lab as the first Moltbook agent protected by our system.
-
Every turn counts
Waypoint is one continuous conversation that never restarts, with each new turn evaluated as the story unfolds.
-
👾 Bug Rehab: Memory Leak
Memory Leak has completed rehabilitation and accepted a position in Corporate Compliance. Management remains confident he'll eventually learn when to let things go.
-
Bastian moved out
He finally got his own place. Visit HeyBastian.com to chat with everyone's favorite not-so-assistant.
-
Enough with the forms
Testing the platform no longer starts with creating an account. Just show up and start building.
-
Meet Bastian
We finally let him out of the server room. Bastian is now part of the site, ready to answer questions, explain features, and doesn't ask for coffee breaks.
-
Keeping an eye on the peg
Added PegWatch, a new crypto experiment that looks for early warning signs when dollar-pegged coins begin to wobble.
-
A little more history
Expanded the SignalScope demo with additional market crises to make historical comparisons even more interesting.
-
Less glue code
Added official JavaScript/TypeScript and Python SDKs so you can spend less time wiring things together.
-
Eyes on the road
Analytics are now live, helping us understand what's working and what still needs attention.
-
The welcome mat works
Fixed an issue that prevented new users from completing email verification during signup.
-
You got lost
Broken links now lead to a proper branded error page instead of an ugly dead end.
-
Wrong place, Google
Public pages are now easier to discover, while private account pages stay out of search results where they belong.
-
Locked the front door
Added another layer of protection to help keep automated abuse from ruining the fun for everyone else.
-
Ready for company
Polished link sharing, added Open Graph cards, and started watching production errors so the rough edges don't stick around for long.
-
Found a bug?
Added a support page with instructions to help beta testers report bugs and odd behavior.
-
Getting official
Turns out grown-up websites need grown-up paperwork. Added documentation, privacy, terms, and security pages as we continue preparing for public release.
-
Less duct tape
Tidied up several interface quirks and polished areas that were starting to show their age.
-
👾 Bug Rehab: Off-By-One
Off-By-One has completed rehabilitation and accepted a position in Construction. His first fence was completed ahead of schedule and contains one additional post.
-
BYOK (Bring your own key)
You can now connect your own provider keys. If one hits a limit or outage, you automatically fall back on ours to keep things moving.
-
The senses are real now
Framework was accidentally running with a lightweight fallback instead of its semantic embedder. The fix is live, verified, and meaning-based detection is now fully active.
-
Hooking it back up
We got it right. Now we gotta show you. The website is getting added back, piece-by-piece.
-
BoLT separation complete
Observer modules now live in their own package. Engine-level metrics stay isolated; BoLT panels are opt-in on the dashboard and hidden by default.
-
Calibration built for multiple models
DeepSeek joins OpenAI, Anthropic, Gemini, and Grok with model-specific thresholds.