FuturixAI01
Uneven latency
Response times swung by region and time of day, and slow answers broke the flow of a coding session.
How OpenCode runs its models
on FuturixAI hosting.
FuturixAI runs multiple models as reliable hosted services for OpenCode, the open-source AI coding agent.
- Customer
- Open-source AI coding agent
- Scope
- Model hosting and inference
- Deployment
- Managed, multi-region
- Rollout
- 8 weeks
- Industry
- AI Infrastructure
About OpenCode.
OpenCode is an open-source AI coding agent used by developers in the terminal and the editor. Its users switch between models for different tasks, and expect each one to answer quickly at any hour.
OpenCode’s challenge.
As usage grew, serving several models became a second job for a small maintainer team.
02
On-call load
Maintainers were paged for capacity issues instead of working on the agent itself.
03
Cost without control
Capacity was sized for peaks, so idle hours still cost the same as busy ones.
What we built.
FuturixAI took over model hosting end to end, so the OpenCode team calls one stable endpoint per model.
01
Managed endpoints
Each model runs behind its own endpoint with versioning, so upgrades never break clients.
02
Autoscaling
Capacity follows demand per region and scales down in quiet hours.
03
Monitoring and on-call
Latency and error alerts reach FuturixAI first, with a runbook shared with maintainers.
How we rolled it out.
- Weeks 1–2
Load profiling
Measured request patterns, token volumes and peak hours per model and region.
- Weeks 3–4
First model live
Moved the most-used model behind a managed endpoint with a fallback to the old setup.
- Weeks 5–6
All models moved
Migrated the remaining models one by one, each with a rollback plan.
- Weeks 7–8
Handover
Agreed alert thresholds, on-call rules and monthly capacity reviews.
Before and after.
| Before | With FuturixAI | |
|---|---|---|
| Latency | Varied by region and hour | Stable median across regions |
| Incidents | Maintainers paged directly | FuturixAI on-call handles first response |
| Capacity | Sized for peak all day | Scales with demand |
| Upgrades | Manual, risky swaps | Versioned endpoints with rollback |
OpenCode results.
- 99.9%
- Endpoint uptimeMeasured across all hosted models over the first quarter.
- 40%
- Lower median latencyCompared with the month before migration.
- 6
- Models hostedEach behind its own versioned endpoint.
- 0
- Maintainer pagesFor capacity issues since handover.
“We stopped thinking about servers and went back to building the agent.”
Connected systems.
- Model endpoints
- Autoscaling
- Usage metering
- Alerting
- Status page
What other teams can take from this.
- Hand off hosting early, before on-call load slows the product.
- Move models one at a time, each with a rollback path.
- Share the runbook, so both teams know who acts first.
FAQ
OpenCode case study FAQs.
What does FuturixAI do for OpenCode?
FuturixAI hosts the AI models OpenCode uses as managed endpoints, scales capacity with demand and handles monitoring and first-response on-call.
How long did the migration take?
Eight weeks: two weeks of load profiling, then models moved one at a time with a rollback plan, then handover of alerts and on-call rules.
Can FuturixAI host models for other AI products?
Yes. FuturixAI runs models as hosted services for AI products that need reliable inference without running their own infrastructure.