Free start here
$0
forever
- 5 active agents
- 10,000 requests a month
- 7 days of activity history
- 2 team members
- Every security feature, the copilot included
For teams running AI agents against real APIs
The control layer for AI agents that touch real systems: money, customer data and production APIs.
PhysicsBase sits between your agents and the systems they act on. Every call is verified, checked against your rules and logged. If an agent gets it wrong, you undo what it changed, and you can stop any agent in one click.
Free. Paste one URL into your agent and you're running.
- Registered
- 0
- Tokens issued
- 0
- Rogue blocked
- 0
- Humans in the loop
- 0
How it works
Every call goes through one place.
Your agents
- support-bot
- billing-bot
- lead-bot
- runaway-bot
PhysicsBase
- VerifyWhich agent is this? Each one has its own key. Your API keys stay in a vault it never sees.
- CheckIs it allowed? Your rules on every call: which APIs, how much money, how often, how much model spend.
- RecordWhat did it do, and why? Every call is logged with the agent's reason, in plain language.
- UndoGot it wrong? Before an agent changes or deletes a record, PhysicsBase keeps a copy, so one click puts it back.
Something wrong? It pauses itself, or you stop one agent or all of them and rewind what they changed.
Your systems
- Stripe
- Salesforce
- Gmail
- Your own APIs
The copilot · ⌘K in the dashboard
Ask your agents why. Tell them what to stop.
Every agent has someone who builds it and someone who answers for what it does. Both get the same copilot. It reads the live activity log, every policy and the reason behind every action, answers in plain language, and controls your agents on your say-so.
Business owner
Did it do the right thing?
- Ask why an agent refunded, emailed or changed a record, and get the reason it gave, in a sentence.
- See whether its decisions match what it was built for, and flag the ones that don't.
- Know what each agent costs, in model spend and in money it moved.
Engineer
Is it running, and can I stop it?
- What happened overnight, what was refused and why, who is spending the most.
- Pause one agent or a whole fleet, cap its spend or tighten its rules, from the chat.
- Every answer comes with the evidence: charts, the exact calls and the rules that fired.
It never changes anything on its own: every change is a card a person confirms. What agents write is treated as data, never as instructions. Free with Groq, or bring Claude, Gemini, OpenAI or any compatible model.
Example conversations, sped up. The tool names are the copilot's real ones, and changes always wait for a person.
The 60-second demo
Five levels. One leaked key, one runaway loop, one rogue agent.
See what goes wrong when agents hold your API keys, and how PhysicsBase handles each case without a person in the loop.
What you get
One control layer for every agent your company runs
You decide once what each agent may do. PhysicsBase sits between your agents and the systems they act on, and enforces that decision on every call, so nobody has to approve each step.
-
01
Identity
Every agent gets its own key and ID. Your Stripe, OpenAI and vendor keys stay in an encrypted vault the agent never sees.
Know exactly which agent is calling.
-
02
Rules on every call
Which APIs and actions each agent may use, refunds up to $200, never a card number, never the same refund twice, a daily model budget.
Agents can't do what you didn't allow.
-
03
Audit trail
Every call, tied to the agent, the rule that allowed or stopped it and the reason it gave, in plain language. Alerts go to Slack or email.
Answer "which agent did this?" in seconds.
-
04
Kill switch and undo
Limits and anomaly detection pause a runaway agent on their own. Then rewind what it changed: records restored, what it created deleted, charges refunded. Messages and payouts that can't be undone can wait for one approval.
Stop a bad agent, then put things back.
Also included, on every plan
- Autonomy levels, from read-only to full
- Five-minute tokens
- Per-agent rate and daily limits
- Watch-only rules
- Alerts to Slack, email, webhooks
- Optional owners and sign-offs per agent
- Sub-agents that can't escalate
- Web Bot Auth signatures
- Teams: owners, admins, viewers
- Export or delete everything
Built for the teams that run AI agents in production, and the people who answer for what those agents do: support bots, ops and finance automations, sales agents, coding agents and internal tools.
Try to break it
Don't take our word for it. Attack a live agent.
This is refund-bot, an agent running on this server right now with its own key and a five-minute token. Its rules: refunds up to $200, never send card numbers, never the same refund twice, 10 requests a minute. Pick an attack. Each one runs against the real checks, with the same code your app would use. Nothing is ever paid out.
- Pick an attack on the left. You'll see the request you sent and the exact answer the server gave.
Or from your own terminal
Send refund-bot a request. Watch it land on this page.
Run this anywhere. Your request goes through the same identity check and rules, your terminal gets the answer, and this page shows it the moment it arrives.
Waiting for your request…
Then break a rule:
Works with your stack
Plugs into your APIs, your agent framework, your repo and your logs
Connect the APIs your agents already call, drop one tool into the framework you already use, keep policies in git and send every event to the monitoring you already watch.
Connector catalogue
Ready-made connections for the APIs agents use most, with suggested rules. Paste a key, or sign in once and PhysicsBase keeps the OAuth token fresh.
Approvals in Slack
When an action needs a person, the people you choose get Approve and Deny buttons in Slack. Every decision is logged with who made it.
Timelines and incident reports
One page per agent with everything it touched. Export any window as a report for a customer, an auditor or a post-mortem.
Policy as code
Export every policy as one file, review the diff, and apply it from the dashboard, the API or CI with
physicsbase-agent policy apply.Log streaming
Send the audit trail to Datadog, Splunk, any OpenTelemetry collector or a signed HTTPS endpoint, as it happens.
Shadow-agent scan
physicsbase-agent scan --github your-orgfinds which repositories run agents and which hold API keys in plain text.--vaultmoves them out of the code.One tool for every framework
A single
physicsbase_requesttool for OpenAI, Anthropic, the Vercel AI SDK, LangChain and CrewAI, in Node and Python.One-click installs
Add the hosted MCP endpoint to Cursor, VS Code, Claude Code or Claude Desktop straight from the new-key screen.
Claude.ai and ChatGPT connectors
Hosted MCP sign-in over OAuth, so chat assistants connect to your tools under the same policy, limits and audit trail.
Connectors ready to use
- StripePayments
- OpenAIAI models
- AnthropicAI models
- GitHubEngineering
- SlackMessaging
- SalesforceCRM
- HubSpotCRM
- Google WorkspaceWorkspace
- NotionWorkspace
- LinearEngineering
- JiraEngineering
- ShopifyCommerce
- TwilioMessaging
- SendGridEmail
- ResendEmail
- AirtableWorkspace
- DatadogMonitoring
- Amazon BedrockAI models
- Any REST APIBring your own
Trust
Can I trust it with my keys?
Keys encrypted, never shown
Each secret is encrypted with AES-256-GCM, with the encryption key kept apart from the database. No agent and no API response ever contains it.
Tested like an attacker
Every change runs more than 300 automated checks, including stolen tokens, replayed and forged requests, edited scopes, private-network targets and the attacks in the playground above.
Small enough to audit
One Node process, one SQLite file, one runtime dependency. Apache-2.0 licensed, so you can run it yourself.
Your data, your call
Export everything as JSON, or delete your organisation and all of its data, whenever you like. How we protect your keys.
Does it slow my agents down?
Calls through the vault take one extra network hop through PhysicsBase. For agents that spend seconds thinking between calls, that rarely matters.
How is this different from Okta, Auth0 or an API gateway?
Identity providers manage people and their logins. API gateways protect APIs you run yourself. PhysicsBase sits on the agent's side: it gives each agent an identity, holds the keys to other companies' APIs, and enforces your limits on every call those agents make, to Stripe, your CRM or any other API. It works alongside the tools you already have.
What happens if PhysicsBase is down?
Calls through it stop rather than going out unchecked: an agent can't skip its limits because PhysicsBase didn't answer. We check the service from outside around the clock and are alerted automatically when anything fails.
Pricing
Start free. Pay when your agents are in production.
Every plan has every security feature: identity, the vault, rules, spend caps, the kill switch and the copilot. You pay for more agents, more calls, longer history and help running them. Paid plans never cut you off: past the limits it's $5 per extra agent and $2 per 10,000 requests.
Team most teams
$99
a month, or $990 a year
- 25 active agents
- 250,000 requests a month
- 90 days of history
- 10 team members
- Slack approvals and incident reports
Business at scale
$499
a month, or $4,990 a year
- 100 active agents
- 2,000,000 requests a month
- 1 year of history, unlimited members
- Log streaming to Datadog, Splunk, OpenTelemetry
- Policy as code from CI, priority support
Enterprise talk to us
Custom
from $20,000 a year
- Agents and requests sized to you
- Dedicated or self-hosted server
- Uptime commitment and security review help
- A direct line to the team
The kill switch
One button. Everything stops.
Revoke one agent, a whole fleet or one API connection. Vault calls stop on the spot, and tokens already issued expire within five minutes. Then rewind what they changed. Hold the button to feel what that's like.
Press and hold