VxCloud
Agent runtime · prod-ready

Ship agents
that actually do the work.

A managed runtime for autonomous AI agents — typed tools, vector memory, multi-model routing, policy guardrails, and a full audit trail. Define once in Python or YAML, run from CLI, REST, cron, or your app, with the same observability you expect from any production service.

0+
built-in tools
0
model providers
0.99%
replayable audit
agent:refund-bot · liveTHINKING
planner·decomposed goal412ms
"refund order #4821" → 6 sub-tasks

Routes between the model providers and tool ecosystems you already use

Anthropic
OpenAI
AWS Bedrock
Azure OpenAI
Vertex AI
vLLM
MCP
LangChain
LlamaIndex
OpenTelemetry

One agent. Three surfaces.

Define in Python. Deploy via GitOps. Call from anywhere.

Author agents in idiomatic Python with typed tools and memory, or define them declaratively in YAML for GitOps. Either way, every workspace gets a signed REST endpoint, OpenTelemetry traces, and a one-click replay surface for every session.

  • Typed tool schemas — no JSON-string juggling, no silent failures
  • Per-tenant memory and Vault namespace — no cross-tenant leakage
  • Budget + tool-call ceilings enforced server-side
  • Sessions replayable in the dashboard or piped to OpenTelemetry
1# Pip install once: pip install vxcloud-agents
2from vxcloud.agents import Agent, Tool, MemoryStore
3
4stripe = Tool.openapi("https://api.stripe.com/openapi.json", auth="vault://stripe")
5gmail = Tool.mcp("gmail", server="mcp://workspace/gmail")
6policy = Tool.python(refund_policy) # local python function -> typed tool
7
8memory = MemoryStore.pgvector("orders-refunds", scope="tenant")
9
10agent = Agent(
11 model="claude-opus-4-7",
12 tools=[stripe, gmail, policy],
13 memory=memory,
14 guardrails={
15 "max_amount_usd": 500, # above -> human approval
16 "deny_paths": ["/users/delete"],
17 "max_tool_calls": 12,
18 "budget_usd": 0.05,
19 },
20)
21
22result = agent.run(
23 goal="Refund order #4821 if eligible and notify the customer.",
24 context={"order_id": "4821", "actor": "[email protected]"},
25)
26
27print(result.summary) # plain-English summary
28print(result.audit.url) # signed replay link

Want to draft an agent from a goal?

vxcli agent new "ship a refund bot for Stripe" — generates the Python and YAML, ready to commit.

agents · v1 · prod
Python · YAML · REST

The runtime, not just a wrapper

Everything an agent needs to ship to production

Memory, tools, planning, guardrails, audit, and routing — built in, not bolted on.

Plan, act, observe, repeat

Agents decompose a goal into steps, pick the right tool for each step, observe the result, and loop until success or a configured budget is hit.

Multi-agent orchestration

Compose specialist agents (planner, retriever, coder, reviewer) into supervised graphs. Hand off context cleanly; nothing runs unsupervised.

Long-term + episodic memory

Vector memory for semantic recall, episodic memory per task, and short-term scratchpads — all per-tenant, encrypted, and auditable.

Typed tool catalog

Wire HTTP, GraphQL, SQL, gRPC, and MCP servers as typed tools. Schema validation in, schema validation out, with retry and backoff baked in.

Human-in-the-loop guardrails

Policy engine gates risky actions: amount thresholds, PII writes, production resource changes. Approval via Slack, email, or dashboard.

Full audit trail

Every prompt, tool call, observation, and decision is recorded with cryptographic chain-of-custody. Replay any session in the dashboard.

Typed tool catalog

Every API your agent touches is a typed contract

Wire any HTTP, GraphQL, SQL, gRPC, or MCP endpoint as a typed tool. Schema validation runs on the way in and on the way out — no more silent JSON deserialization failures taking down your agent at 2am.

HTTP / REST
core
GraphQL
core
SQL (Postgres, MySQL)
data
Vector store (pgvector)
data
Stripe
apis
GitHub
apis
Slack
apis
Gmail / SES
apis
MCP servers
core
Shell / Bash
core
Browser (Playwright)
core
Custom Python
core

Multi-model routing

Pick the right brain for each step — automatically

Route reasoning to Opus, fast loops to Sonnet, classification to Haiku, and air-gapped tenants to your own self-hosted model. One config, one bill, one observability surface.

Claude Opus 4.7

reasoning
1M ctxbest for: planning, refactors

Claude Sonnet 4.6

fast
200kbest for: tool-calling loops

Claude Haiku 4.5

cheap
200kbest for: classification, routing

GPT-class reasoning

reasoning
128kbest for: multi-step coding

Self-hosted Llama

sovereign
BYObest for: air-gapped tenants

BYO endpoint

BYOM
anybest for: Bedrock, Azure OpenAI, vLLM

Real workloads, real numbers

Production-grade — measured on production traffic

Benchmarks aggregated from agent sessions running on the platform. Not synthetic; not hand-tuned demos.

Cold start (first run)1.8s
Warm tool-call latency p5078ms
Goal-completion rate94%
Audit log coverage100%

Real-world agents teams have shipped

From customer ops to incident response

Refund and dispute bot

Reads order history, applies refund policy, calls Stripe, emails the customer, and writes the audit entry — under 2s, $0.005 per resolution.

Tier-1 support copilot

Answers product questions from your docs, files tickets when escalation is needed, and writes the postmortem when an incident closes.

PR triage agent

Watches GitHub PRs, runs targeted reviews, requests changes with line-level comments, and merges low-risk dependency bumps after CI is green.

Incident commander

Joins the on-call channel, pulls traces and logs, suggests a fix, and drafts the timeline as the incident unfolds.

Data-quality watcher

Profiles new ETL outputs, flags drift, opens Linear tickets with a reproducible query, and assigns the right data engineer.

Sales-ops research agent

Enriches inbound leads with firmographic data, scores them against your ICP, and drops them into the right Slack channel with context.

Ship your first agent before lunch.

Open a workspace, point it at a tool, and run a real goal. No credit card, no cluster to provision — and no synthetic demo data.