Autonomous agents,
on governed rails.
Refusion® provisions scoped agents, routes every task to the right model, and enforces spend & policy before any tool fires — all observed in one audit stream.
Watch an agentic task workflow run.
Pick a framework. A real support task flows through the pipeline live — every step routed to its own model, metered, and threaded into the next. This is the same primitive you provision in the console.
Triage → Resolve → Draft.
Run a real agent task — no signup.
See your savings vs. frontier models.
Rudy picks your plan, you pay in a click.
Auto reminders before your trial ends.
Watch a run, end to end.
Provision an agent from the CLI, route the task to the right model, and observe the audit stream — every step governed and metered in real time.
Run a real task. Watch it route.
Pick a use case, edit the prompt, and hit run. The router picks the model and meters the exact cost — the same path your agents use in production.
Every route saves. The numbers add up.
Refusion routes each task to the cheapest capable model and meters the difference vs. a frontier-class baseline. Here's what the public sandbox has saved so far.
- 1Summarize2 runs$0.0196.7%
- 2Triage4 runs$0.0195%
- 3Reasoning2 runs$0.0180%
- 4Conversational1 run$0.0190%
Everything in the plane
Primitives to provision, route, and govern — plus a built-in CLI, demo, and docs.
Cost Autopilot
Shadow-evaluates a cheaper model on every task, auto-promotes it when quality holds, rolls back the instant it slips. Your routing tunes itself.
Bounty Board
Post tasks with reward credits — assigned packs run them and earn. The autonomous workforce, on governed rails.
Live run monitor
Watch in-flight orchestration runs update in real time — spend ticks up as each step routes.
Spend & savings analytics
Aggregate spend over time, cost per task type, model mix, and top agents — the ROI story made tangible.
Multi-agent orchestration
Chain agents into supervisor-driven swarms with per-step routing and swarm-wide spend caps.
Skill Packs
Pre-assembled bundles of subagent personas and callable skills — clone a showcase pack or author your own.
Agent provisioning
Scoped namespace, API key, and a per-agent spend cap — issued in one command.
Model router
Per-task model selection with automatic failover and exact per-run metering.
Governance
Spend caps and lifecycle checks enforced before any tool is allowed to fire.
Interactive CLI
Provision, run, and govern agents from a single in-app terminal prompt.
Guided demo
Watch the full provision → route → audit flow play out end-to-end in 3 steps.
Platform docs
Primitives, the CLI command reference, and the governance model — all in-app.
Live sandbox
Run a real routed task right on the landing page — no signup required.
ROI calculator
Estimate monthly savings vs. raw frontier models in a few sliders.
One-click checkout
Rudy recommends the right plan from your quiz and starts checkout instantly.
Governed agent swarms, pre-assembled.
Curated bundles of subagent personas and callable skills that run as governed multi-agent swarms — caps, audit, and spend metering inherited from the plane. Clone a showcase pack or author your own.
A specialist for every business function.
Deploy pre-trained agents for sales, support, marketing, HR, engineering, and research — each scored on a benchmark and auto-paused if it drifts. Connect your tools once and put a specialist to work.
Personalize outbound sequences
Agents that learn, earn, and report.
The plane now closes the loop: post bounties your packs earn completing, watch runs live, read the spend & savings analytics, and get a daily performance digest in your inbox.
Your routing gets cheaper every week. By itself.
The autopilot shadows a cheaper model against your real task mix, scores every sample against your quality bar, and auto-promotes it the moment it sustains the pass-rate. The instant quality slips, it rolls back — no one touches a dial, and your agents never get worse.
- Shadow-evaluates a cheaper model on rotated real prompts
- Auto-promotes after a sustained eval pass-rate streak
- Auto-rollback on the first failed sample — quality never degrades
- Cumulative savings reported on your analytics dashboard
Between your agents and your warehouse, a governance layer.
Refusion sits between autonomous agents and Snowflake, Databricks, or any SQL source — a semantic context layer agents query by name, a policy-as-code engine that redacts PII and blocks before any query runs, and a full GDPR lifecycle with retention windows and right-to-erasure.
Governed data access, priced per fleet.
Semantic context layer, policy-as-code, PII redaction, and full GDPR lifecycle — connect Snowflake, Databricks, or any SQL source with every read audited.
- Semantic assets, never raw tables
- Declarative policy engine
- PII classifier + redaction
- GDPR erasure & retention purge
- Warehouse connectors
- Signed audit exports
Semantic context layer
Agents request semantic assets by name — never raw tables. Register the tables, columns, and PII flags you want exposed.
OpenPolicy-as-code engine
Declarative rules — if PII in prompt then redact, if source has PII then require role, retain 90d then purge. No code per run.
OpenPII classifier
Deterministic detection on every prompt and result. Sensitive values redacted before model routing; redactions counted in audit.
OpenGDPR lifecycle
Data-subject tagging, retention windows, and one-click right-to-erasure. A scheduled job purges expired events.
OpenSnowflake · Databricks · SQL
Connect each customer’s own account (app-user) or one workspace OAuth app (shared) — governed reads, every access audited.
OpenAudit exports
Exportable signed audit reports with per-agent lineage, retention metadata, and policy decisions for SOC 2 / GDPR reviewers.
OpenThree primitives. One governed loop.
Provision
Scope a namespace, issue a key, set a spend cap.
Route
The router picks the model and meters exact cost per run.
Govern
Caps enforced before any tool fires — every run audited.
Pay for what your agents run.
Refusion is metered like infrastructure — provisioned agents, routed task runs, and frontier token pass-through. No per-seat tax. Every run is logged, governed, and capped.
Scoped namespace, API key, spend cap. First 3 included on every plan.
Smart-routed, governed, and audited — including fallback.
Pass-through cost when you pin the frontier strategy.
Knowledge bases and shared-run storage, priced by GB/mo on paid plans.
Plans that scale with your fleet
Start free. Upgrade when your agents hit production. Governance included on every tier.
First agents in production.
- 3 provisioned agents
- $250 spend cap / agent
- Smart routing
- Audit stream
For operating agent fleets.
- 20 provisioned agents
- Custom spend caps
- All 4 routing strategies
- Failover + exact metering
- Audit + revoke alerts
Governed data access + GDPR.
- Everything in Scale
- Semantic context layer
- Policy-as-code engine
- PII classifier + GDPR lifecycle
- Snowflake / Databricks / SQL connectors
- Audit exports + data residency
Governed, org-wide rollouts.
- Unlimited agents
- Dedicated routing
- SSO / SCIM
- SLA + priority support
- Custom governance policies
Read the manual — in the console.
The whole platform is documented in-app: primitives, the CLI command reference, the governance model, and how to author multi-agent swarms. Sign in to open it.
Core primitives
Agents, routing, audit — how the pieces fit.
CLI reference
Provision, run, and govern from the prompt.
Governance model
Spend caps, lifecycle checks, revocation.
Orchestration
Build supervisor-driven agent pipelines.
See what smart routing saves.
Estimate your monthly LLM spend on raw frontier models versus Refusion's per-task routing — and which plan fits.
Provision your first agent in one command.
Sign in to unlock the console, the CLI, and the model router — governed by default.