Pricing
The runtime is free. Always.
- Apache 2.0
- no strings, no rug pull
Pilot is Apache 2.0 and runs entirely on your own infrastructure. You pay only if you would rather not operate the trace store, the eval runner, and the approval queue yourself.
Plans
Free to run. Pay for what you keep.
Every plan runs the same runtime — the open-source one. What Cloud and Enterprise add is somewhere to put the traces, the evals, and the approvals.
Open source
The full framework, self-hosted.
$0 forever
Read the guides- The whole runtime, Apache 2.0
- Durable loop and checkpoints
- Typed tools and schema derivation
- Local traces and replay
- Bring your own models and storage
- Community support
Cloud
Hosted traces, evals, and approvals.
$99 / month
Get started- Everything in Open source
- Hosted trace store, 90-day retention
- Eval suites that run in CI
- Approval routing to Slack and email
- Spend metering and budget ceilings
- Five seats included
Enterprise
Your VPC, your compliance team.
Talk to us
Contact sales- Everything in Cloud
- Self-hosted control plane in your VPC
- SSO, SCIM, and audit logs
- Unlimited retention
- Private model endpoints
- A named engineer, not a queue
Every plan, line by line
| Feature | Open source | Cloud | Enterprise |
|---|---|---|---|
| The runtime and durable loop | Included | Included | Included |
| Typed tools and validation | Included | Included | Included |
| Trace retention | Local only | 90 days | Unlimited |
| Eval suites in CI | — | Included | Included |
| Approval routing | — | Included | Included |
| Spend metering and budgets | — | Included | Included |
| Seats | Unlimited | 5 included | Unlimited |
| Self-hosted control plane | — | — | Included |
| SSO, SCIM, audit logs | — | — | Included |
| Support | Community | Priority | Named engineer |
Questions
The ones people actually ask
Still unsure? Write to us and a person answers.
hello@pilot.devIs this another orchestration DSL?
No. Pilot is a library you call from ordinary code — a function, a queue worker, a controller. There is no graph to draw and no YAML to learn. If you can write a class and a few typed methods, you have written an agent.
What happens when a tool call fails?
The step retries with backoff. If it keeps failing, the run parks at its last checkpoint with the error attached rather than unwinding. You fix the tool, resume the run, and it picks up from that step instead of starting over.
Which models does it work with?
Any of them. Pilot speaks the OpenAI and Anthropic wire formats natively and anything OpenAI-compatible through configuration. Model choice is per-step, so a cheap model can plan and an expensive one can write.
Do I have to send you my data?
Only if you want hosted traces. The open-source runtime writes traces to your own Postgres and never calls home. Cloud is a place to put them, not a dependency of the framework.
How is this different from just calling the API in a loop?
For a demo it is not. The difference shows up the first time a run dies at minute nine of twelve, or a tool gets an argument that does not type-check, or someone asks why the agent refunded that customer in March.
Can a human approve things mid-run?
Yes, and this is the part teams underestimate. Put a gate anywhere in the loop; the run parks, routes to whoever should decide, and resumes exactly where it stopped once they do. Runs can wait for days.
Start with the open-source runtime
Nothing to sign up for. Add the dependency, write one tool, and decide about the hosted parts later.
Apache 2.0 · Bring your own models · Self-host everything