A runtime, not a library
Durable execution, checkpointing, idempotent tools and automatic retries. Long-running agents survive restarts, timeouts and partial failures without bespoke glue code.
Explore the runtime
Series B · $48M raised · SOC 2 Type II
Agents that survive
Agents that survive
production.
Veltrix is the runtime teams use to deploy autonomous AI agents with real permissions, measurable quality and an audit trail your security team will accept. Bring any model. Ship in days, not quarters.
No credit card · 5,000 free agent runs every month · Deploy in your own VPC
1.2B agent steps / month
99.99% control-plane uptime
42 ms p50 tool latency
support-triage-01
1,204 runs · 99.2% pass
invoice-recon-03
318 runs · 97.8% pass
security-triage-02
eval drift · review
claims-intake-us-east
842 runs · 99.6% pass
7-day success rate+4.2% vs. last week
Trusted by operations teams at
LUMEN FREIGHT
PACIFICA HEALTH
CADENCE RETAIL
ORBIT FINANCIAL
NIMBUS LOGISTICS
TERRA MFG.
1.2B
Agent steps executed monthly
99.99%
Control-plane uptime, trailing 12 months
42 ms
Median tool-call latency, US regions
87%
Tickets resolved without a human handoff
Why teams move to Veltrix
Frameworks get you a demo. Runtime gets you a launch.
Most agent projects stall between prototype and production. The gap is not the model — it is permissions, retries, cost control, evaluation and the paper trail.
Governance by default
Scoped credentials, per-tool policy, human approval gates and an immutable action log. Every step is attributable to a user, an agent and a policy version.
Read the security modelEvals that gate deploys
Golden datasets, LLM judges and regression gates in CI. A prompt change that drops quality below your floor never reaches a customer.
See evaluation gates
The platform
Every layer an agent needs — already wired.
Stop stitching together a workflow engine, a vector database, a secrets manager and a tracing tool. Veltrix ships them as one runtime with one billing surface and one audit model.
- Model-agnostic: OpenAI, Anthropic, Google, Mistral or your own weights
- Tool mesh with typed schemas, OAuth scopes and rate limits
- Durable memory: thread state, vector recall and long-term facts
- Replay any run step-by-step, including the exact model inputs
Control planepolicy · evals · audit
Agent runtimeplanner · tools · memory
Data planevector · SQL · blobs
Model gatewayrouting · fallback · budget
Deploy targetVeltrix cloud · your VPC
From zero to production
One file to define. One command to deploy.
Define the agent, register its tools, then promote it through evals to production.
agent.py
from veltrix import Agent, tool
# Tools are typed, scoped and audited automatically
@tool(scopes=["crm:read", "crm:write"])
def update_case(case_id: str, note: str) -> dict:
return crm.cases.update(case_id, note)
agent = Agent(
name="support-triage",
model="claude-sonnet",
tools=[update_case, search_kb, refund_order],
policy="policies/support.yaml",
approval="refunds_above_500_usd",
)
# Same agent: local run, CI eval, or production
agent.run("Where is order 88213?")1. Define
Declare tools, policies and approval gates in code or YAML. Veltrix validates schemas and refuses ambiguous permissions at deploy time.
2. Evaluate
Run your dataset in CI. Quality, cost and latency are scored per case; the deploy is blocked if any metric falls below your floor.
3. Operate
Live traces, cost budgets, incident replay and one-click rollback. No pager duty required at 2 a.m. because a tool timed out.
Security first
Built for the review your CISO will run.
SOC 2 Type II, HIPAA and ISO 27001 controls. Customer data stays in the region you choose, and never trains a model.
SOC 2 Type II
Independently audited, report under NDA.
HIPAA
BAA available on Team and Enterprise plans.
ISO 27001
Certified ISMS across platform and support.
GDPR & CCPA
EU and US data residency, DPA on request.
Customer evidence
Measured outcomes, not pilots.
“We replaced a 14,000-line orchestration service with two Veltrix agents. Resolution time dropped 62% and we finally have an audit trail legal accepts.”
MRMaya ReyesVP Support Operations, Lumen Freight
“The eval gates are the reason this shipped. Every prompt change runs 420 historical tickets before it touches production.”
DKDaniel KimHead of AI Platform, Orbit Financial
“Deploying inside our VPC was non-negotiable. Veltrix passed our architecture review in three weeks instead of three quarters.”
SPSarah PatelCIO, Pacifica Health
Put an agent in production this week.
Start free with 5,000 runs a month. Talk to an engineer about your architecture, not a sales script.