Stop rebuilding guardrails for every AI agent.
SteerPlane sits between your agent and its tools. It stops loops, caps spend and blocks unsafe calls while the agent runs. Host it yourself, or let us run it for you.
No signup. Install, set a dollar limit, and watch your agent stop at it.
$ pip install steerplane$ npm install steerplaneWhat goes wrong when nothing stops an agent
In their own words, from our conversations with teams running agents.
“People don't actually trust any agent to actually go and fix files. So we are focusing on the approval part right now, because that's the most important thing.”
“It just shows up as a tool call. So until and unless you go into that tool call and see what input it's sending, you won't notice it.”
“...notorious for going off on tangents and working forever with no end in sight.”
SteerPlane puts a dollar limit on every run, stops repeated calls, and logs every step, so these end in a stop, not a bill.
A runtime safety layer for AI agents
SteerPlane sits between an agent and its tools and checks every step while the agent runs. It stops repeated loops, caps spend per run, and blocks calls you have denied, such as delete_* or drop_*. Every step is logged.
- Who it is for
- Teams running AI agents in production, where an agent can spend money, send messages or change data.
- How it decides
- Fixed rules, not a model, so the same step always gets the same answer. Checks run before the step does.
- Works with
- LangChain, CrewAI, AutoGen, the OpenAI Agents SDK, plain Python and TypeScript. Any OpenAI-compatible client works through the gateway.
- How to run it
- Self-host it, or let us run it for you. The core guardrails are free. Install with pip or npm.
Agents need limits. Most teams write their own.
We spoke with teams shipping agents, from startups to regulated enterprises. The same three problems came up.
Every project gets its own guardrails
Step caps, timeouts, regex checks, cost limits. Teams write them again for each agent. They drift apart and nobody owns them.
Built for teams that cannot take risks
Large and regulated teams ask one question first: where does the data go? With SteerPlane it stays in your environment. Every check runs before the step does.
Rules decide, not a model
Every check is deterministic. The same step always gets the same answer.
Checks before the step runs
A blocked call never reaches your tools or database. It is stopped, not reported later.
Fails closed
If the control plane goes down, the SDK keeps enforcing limits locally.
Your data stays yours
Self-hosted keeps everything in your environment. Cloud-hosted runs on a dedicated instance. We never sell or share your data.
Keys stay on the server
Your provider key never sits in the agent process.
Open code you can audit
The core engine is on GitHub. Your security team can read every line.
Running in three steps
No rewrite. Wrap your agent, or point your client at the gateway.
- 1
Install
pip install steerplane or npm install steerplane.
- 2
Wrap or proxy
Add @guard to your agent, or change one base_url.
- 3
Set your limits
Cost, steps and denied actions. In code or a YAML file.
from steerplane import guard
@guard(
agent_name="support_bot",
max_cost_usd=5.00,
max_steps=50,
denied_actions=["delete_*", "drop_*"],
)
def run_support_agent():
agent.run() # your code, unchangedFree to start. Hosted when you need it.
The core guardrails are free and open source, for good. Paid plans add advanced protections, hosting by us, and support for large teams.
Open source
FreeSelf-hosted. Yours to run.
- Loop detection
- Cost ceilings
- Step limits
- Action policies: deny, allow, rate-limit, approve
- Zero-code gateway
- LangChain, CrewAI, AutoGen, OpenAI Agents SDK
- Run dashboard
Cloud-hosted
PaidWe run it for you, set up for your team.
- Everything in Open source
- Advanced protections, shared on a call
- Dedicated instance in your chosen region
- Setup help and direct support
Enterprise
CustomFor large and regulated teams.
- Everything in Cloud-hosted
- Runs in your VPC or on-premise
- No data leaves your environment
- Help with your security review
- A named contact and agreed response times
Run SteerPlane on one real agent. Free, with us setting it up.
Our first 10 design partners get every paid feature free.
- We wire it into one of your agents with you, usually in under an hour.
- A direct line to the founders, and your requests go to the top of what we build.
- We ask you to use it on one real agent for 4 weeks and tell us what breaks.