One layer routes, governs and records every agent. Here's what that does for each person on your team. Where something is still on its way, it says preview.
In line · on every request
Four jobs from the first request. Cost, budgets, keys and policy.
Developer on Claude Code or Codex
Same AI coding, smaller bill.
Keep your tool and your sign-in. Every turn shows what it cost and what it saved beyond your provider's own caching. Reading, searching and summaries can go to free or local models, within the privacy level you set.
Receipts per turn
Free until it saves you $500
Recoverable compression
Coaching tips in the status line
Engineering lead
One hub for the team, a budget for each person.
One machine or server serves the team. Developers join with an invite, without running anything extra. You see spend and savings per person, project and agent, and set the caps.
Team hub with invites
Budgets per person
FinOps per project and agent
Any OIDC provider preview
Building non-coding agents
Give each agent a key, not a credit card.
Every agent gets its own key on the team hub, with its own policy and budget cap. The providers behind it are abstracted away: the key reaches frontier, free and local models, and the agent never holds your real credentials.
Per-agent keys
Budget cap and privacy floor
Model and IP allow-lists
Rotate or revoke any time
Security and compliancepreview
Know what agents may do, and keep the proof.
Every tool call gets allow, ask or deny before it runs, judged on what the call actually does. Start in observe mode to see what would have been stopped. The audit log is tamper-evident and you can verify it offline.
Policy packs
Evidence mapped to SOC 2, ISO 42001, EU AI Act for your assessor
Option: local sandbox per agent, no Docker
OpenTelemetry export
Over time · it compounds
Three jobs that grow with your record. Autonomy, warm models, owned models.
Letting agents do morepreview
Autonomy that agents earn, task by task.
Agents work on isolated branches, every change climbs a ladder of checks, and nothing merges without your approval. As the record shows a kind of task going well, that kind of task earns more freedom. One bad outcome moves it back down.
Isolated branches
Verify ladder
Your approval to merge
Agent profiles that fail closed
Platform team, private cluster or neo-cloud
Your own models, kept warm.
PromptForce knows which context belongs together and how long it will be needed. KV Zones passes that to your vLLM, SGLang or llama.cpp servers as cache hints, and reads back what they kept.
KV Zones preview
Hub in your VPC
Kubernetes sidecar preview
Air-gapped policy pack preview
CTO and finance
From rented models to owned, at fixed cost.
Your record of what worked, and at what cost, is the evidence for moving a job off rented frontier models. We help you move work to models you own, one kind of task at a time, when the record says it's ready. You pay a share of what PromptForce saves, never per seat.