Self-hosted infrastructure for your coding agent fleet
Run every coding agent through one key in your cloud. See who’s spending what, cap it before it runs, and send each request to the model that fits the job.
Open source. Self-hosted. Set up by us.
Your team
- Routing4 models
- Budgets$114/$210
- Guardrails1 waiting
- Context5 skills
- Auditall logged
Models
Tools
Your cloud
Open at every layer
Developers keep any model and any agent. You keep the control to roll AI out safely.
Any agent
Works with the coding agents your team already uses.
- terminal agents
- IDE agents
- background agents
Any model
Frontier or open-weight, picked per task. New models the day they ship.
- frontier
- open-weight
Your compute
Runs in your own AWS or GCP account.
- AWS
- GCP
- Kubernetes
Your data
Prompts, traces and skills stay with you.
- skills
- memory
- traces
Every model for your whole team, behind one login
One sign-in and one key for every model and tool. No more API keys pasted into laptops.
See what every agent costs. Cap what every agent spends.
Give each team and agent a monthly budget. When it runs out, work stops or drops to a cheaper model. No surprise bills.
Easy work goes to cheap models. Hard problems go to the best ones.
Know which model does your work best
Every night we rerun your recent pull requests on each model. You see which one does your work best, and you hear about it the day one slips.
Your PRs, rerun nightly
94%passing
- #1291checkout-web97%pass
- #1288payments-api92%pass
- #1285auth-service58%fail
Models ranked on your codepass rate
- 1Opus 5.594%
- 2Sonnet 5.590%
- 3GPT-5.588%
- 4gpt-oss-120b76%
A model slipped
−12pts vs last week
Sonnet 5.5 on payments-api
flagged 06:10
Let agents use your tools, safely
Decide once which apps each team’s agents can use and what they can do there. Anything risky, like dropping a table, waits for a person to say yes.
background agent wants to run
DROP TABLE sessions
on prod · auth-service
Approved by you
Denied by you
See exactly what every agent did
Every prompt, tool call and approval is saved in your cloud, so you can always answer who did what, and why.
Rotate the session keys in auth-service. Run tests first.
Running tests and a staging deploy. I’ll ask before touching prod.
Approved.
Trace
run_tests()passeddeploy_to_staging()donesmoke_test()passedrotate_keys_prod()running
Hand off work from Slack or GitHub
Tag an agent on an issue or in Slack. It works in a sandbox, tells you when it’s stuck, and opens a tested PR.
Teach one agent, and every agent learns
Turn your best prompts and runbooks into skills once. Every agent on the team uses them, in whatever tool each developer prefers.
- Deploy to staging
- Write migration
- Review PR
- Write tests
- Triage incident
deploy-to-staging/SKILL.md
--- name: deploy-to-staging version: 1.4 --- 1. Run the test suite 2. Build the image 3. Deploy to staging 4. Post the link in #deploys
Built for security teams
Single sign-on
SAML or OIDC, with roles for who can change what.
Stays in your cloud
Your code, prompts and logs never come to us.
Secrets hidden from agents
Keys and passwords are stripped before a prompt leaves.
Your models, your hosting
Use any provider, or run open models on your own GPUs.
One admin console
Set permissions for every team and see usage in one place.
Works with your vault
Pull keys from 1Password, Vault or AWS Secrets Manager.
FAQ
Which agents does it work with?
Any coding agent that can use a custom model endpoint or MCP, which covers the terminal and IDE agents most teams use. Agents that can’t route through us still get rules, budgets and traces.
Do my developers have to change tools?
No. They keep coding in their terminal or IDE. We point their agents at your gateway and push rules and skills into each agent’s settings.
Does your code leave your cloud?
No. Everything runs in your AWS or GCP account. Prompts go only to the model providers you choose.
Who runs it?
We install it, keep it updated and stay on call. You keep the keys; we do the ops.
How does a pilot work?
Six weeks with one or two teams: we audit your agents, set up the gateway and budgets, add guardrails and skills, then put background agents on one real job. One fixed price.
Why not build this ourselves?
You could, with two engineers for six months and someone on call for every agent update. We do that for you, and the core is open source, so you can take it over later.