One runtime for the
whole agent lifecycle
Orchestration, tool access, memory, evaluation and observability share one policy model, one billing surface and one audit log. Deploy in our cloud, your VPC, or fully on-prem.
Four layers, independently scalable
Agents talk to tools through a typed mesh. The control plane decides what is allowed, the data plane remembers what happened, and the gateway keeps model spend predictable.
Durable execution by default
Every step is checkpointed to an append-only log. If a pod dies mid-run, the agent resumes at the last completed tool call instead of restarting the conversation — or double-charging a customer.
- Idempotency keys generated per tool invocation
- Parallel tool fan-out with per-branch timeouts
- Human-in-the-loop pauses that survive deploys
- Deterministic replay of any historical run
12:04:11 step=plan model=claude-sonnet 812ms $0.0041
12:04:12 tool=search_kb args={q:"refund policy"} 94ms
12:04:13 tool=crm.read case=88213 38ms
12:04:13 checkpoint id=ck_7f31a durable=yes
12:04:14 approval_wait policy=refunds>500 state=paused
12:19:02 approval_granted [email protected]
12:19:03 tool=refund amount=312.00 status=ok
12:19:03 step=respond # run complete, 14m52s wallPolicy as code, enforced at runtime
Permissions are evaluated per tool call, not per session. A refund over $500 pauses for approval; a production database write is denied outright. Policies are versioned and diffed like any other code.
- Reusable policy bundles per environment
- Approval queues in Slack, Teams or the API
- Immutable audit export to S3, Splunk or Datadog
- Break-glass access with mandatory review
version: 3
tools:
- name: refund_order
allow: "amount <= 500"
require_approval: "amount > 500"
- name: crm.write
allow: "actor.role in ['support', 'admin']"
deny: "record.tag == 'legal-hold'"
limits:
cost_per_run_usd: 0.40
wall_clock_seconds: 1800Memory that does not leak between tenants
Thread state, semantic recall and long-term facts are partitioned by tenant and region. Nothing is shared across customers, and retention is set per dataset rather than per vendor default.
- Vector, relational and object storage behind one API
- Per-dataset retention from 24 hours to 7 years
- PII detection with automatic redaction hooks
- Point-in-time restore for agent state
Route by cost, latency or quality
One gateway for every provider, with automatic fallback on 429s and outages. Set a per-agent budget and Veltrix will degrade to a cheaper model rather than silently overspend.
- Weighted routing across OpenAI, Anthropic, Google and Bedrock
- Semantic caching for repeated internal queries
- Prompt and completion logging with redaction
- Chargeback reports by team, agent and customer
| Route | Model | Share | p50 |
|---|---|---|---|
| default | claude-sonnet | 62% | 780 ms |
| bulk-classify | gpt-4.1-mini | 24% | 290 ms |
| long-context | gemini-2.5-pro | 9% | 1.9 s |
| fallback | llama-3.3-70b (self-hosted) | 5% | 410 ms |
What ships in the box
No add-on modules. Everything below is included on Team and Enterprise plans.
Observability
Traces, spans and token-level cost for every run. Filter by agent, tenant, tool or policy version, and export to Datadog, Grafana or S3.
Evaluation harness
Datasets, LLM judges and human review queues. Wire the suite into GitHub Actions and block merges that regress quality.
Guardrails
Prompt-injection detection, output schema validation, PII redaction and jailbreak canaries on every request path.
Tool mesh
OpenAPI, MCP and gRPC tools registered once with typed schemas, OAuth scopes and per-tenant rate limits.
Scheduling
Cron, event and webhook triggers with backfill, concurrency caps and dead-letter queues that humans can inspect.
Multi-agent graphs
Supervisor and worker topologies with budget inheritance, so a child agent cannot outspend its parent.
Connects to the systems you already run
First-party connectors are maintained, versioned and tested against sandbox tenants. Anything else is a two-line OpenAPI import.
Your cloud, our cloud, or your racks
Veltrix Cloud
Multi-tenant, US or EU regions, live in one deploy. Best for teams that want zero infrastructure work.
Dedicated VPC
A single-tenant control plane inside your AWS, Azure or GCP account. Traffic never leaves your network boundary.
On-premises
Helm-based install with air-gapped mode, customer-managed keys and offline license renewal.
See it on your own data
A 30-minute session with a solutions engineer: your use case, your systems, a working agent by the end of the call.