PhysicsBaseHow PhysicsBase compares
Updated November 2026
Agent access is being solved from three directions. Tool and action platforms (Composio, Arcade) run the tool calls for you. Identity providers (Auth0, WorkOS, Descope, Ping) extend sign-in for people to agents. Credential managers (1Password) keep secrets out of code. Each is a good fit for some teams. PhysicsBase is the control plane in the middle: every agent gets its own identity, and every call it makes is checked, limited, logged and stoppable, with no person approving each step.
Rather than a checklist of other products, which goes stale and is easy to get wrong, here are the questions we'd ask any vendor, including us, and our answers.
| Ask | PhysicsBase |
|---|---|
| Can an agent get an identity with no person involved? | Yes. It generates its own key and registers under an enrollment key's policy, or as an independent agent. |
| Is a stolen token useless? | Yes. Tokens last five minutes and are bound to the agent's private key (DPoP, RFC 9449). Without the key, a copied token is refused. |
| Do agents ever hold my API keys? | No. Keys and OAuth tokens live in the vault, encrypted. PhysicsBase adds them on the way out and refreshes OAuth tokens itself. |
| Can I limit what is inside a request, not just which API? | Yes. Request rules cap amounts per call and per day, allow or forbid methods and paths, and restrict hours. Example: refunds up to $200, $1,000 a day, never payouts. |
| What happens when an agent goes rogue at 3 a.m.? | Bursts far above its normal rate, refusal storms and first use of a new API are flagged. You choose: alert, or pause the agent automatically. One click resumes it. |
| Can an agent hand work to a sub-agent safely? | Yes. Delegated sub-agents and single-task agents get their own key, can never exceed the parent, expire on their own, and stop when the parent is paused or revoked. |
| Can I see and control my fleet without writing queries? | Yes. The copilot answers "what happened overnight?" from the live activity log and pauses, limits or revokes agents in plain language. Every change needs your confirmation; revoking needs a press-and-hold. |
| Will websites recognise my agents instead of blocking them? | Brokered calls are signed with Web Bot Auth (RFC 9421), which sites and CDNs can verify. |
| Does it work with what my agents already use? | MCP (hosted endpoint, nothing to install), plain HTTP through a proxy, Node and Python SDKs, and drop-in tools for OpenAI, Anthropic, Vercel AI SDK, LangChain and CrewAI. |
| Can policy changes go through code review? | Yes. Export every policy as one file, preview the diff, apply from CI with an API token. |
| Can I send the audit trail to my SIEM? | Yes: Datadog, Splunk, any OpenTelemetry collector, or a signed webhook. |
| Can I run it myself? | Yes. One Node process, one SQLite file, Apache-2.0. Hosted is free too. |
| What doesn't it do yet? | Running on several machines at once (one server today), single sign-on and SCIM for your team, and a published independent security audit. All three are on the roadmap. We'd rather tell you than have you find out. |
When something else is the better fit
- You want a large catalogue of ready-made tools that run in someone else's cloud: a tool platform.
- You only need people to sign in to your app, with agents as an afterthought: your existing identity provider.
- You only need secrets out of your repositories, with no runtime control: a credential manager.
If your agents call real APIs with real money or real data, and you want to know what each one did and stop it in seconds, try PhysicsBase. It's free.