“Who approved that?”
An agent changed something that matters. The approval lived in a chat, if it happened at all.
ConvOps is the operations layer for AI agents, on your own infrastructure. Teach your process once. Every AI and every person runs it the same way, with the gates, the record and the controls your enterprise needs.
self-hosted · your Kubernetes · your models
Cloud ConvOps Cloud runs the operation. Your AI tools, models and repositories stay yours and connect over HTTPS.
The operations layer for AI agents across the enterprise: one process for every AI tool and team, gates that hold a step until a condition is met, an audit trail, and the option to self-host on your Kubernetes.
ConvOps is the operations layer for AI agents. It holds your team's process as workflows with gates, a shared memory, scheduled and unattended runs, programs, a plan board, metrics and a full audit trail. It does not sell or host models.
key facts · october 2026
best for
not for
updated
Security asks that first. The business asks the next three, usually after something went wrong.
“Who approved that?”
An agent changed something that matters. The approval lived in a chat, if it happened at all.
“Where did that decision go?”
It was made in one AI tool, in one session, by one person. Nobody else can find it now.
“What happens when someone leaves?”
The process lived in their prompts. It leaves with them, and the AI never learned it.
Your AI tools on top. ConvOps in the middle, holding the process. Your executors and your cluster underneath, doing the work.
The cluster, the models, the keys, the network, the record and the boundaries between teams. ConvOps governs the work; it does not take any of these.
Every run is a pod in your Kubernetes. Self-hosted, the engine, shared memory and the audit live there too.
Helm chart, with Enterprise
Bring your own model provider and credentials. Isolated runs use Claude Code or OpenCode; people work locally with any AI client that connects.
bring your own
Stored through a one-time link, never pasted in chat. Each run gets its own Secret, masked in every log and transcript.
never in the chat
Your network plugin enforces it: run pods take no inbound traffic and cannot reach the rest of your cluster. No service-account token, non-root, every capability dropped.
pod security: restricted
Every task and every workflow instance keeps its history, step by step. Every run keeps tokens, cost, duration and result.
tasks + workflow instances
One per team, business unit or customer, each walled off. Three roles in each: owner, admin, member.
walled off by default
Start with one real process. Add checkpoints where your risk team wants them, and let the rest run on schedules.
One team, one workspace, one real process. Put a checkpoint on every step if your risk team wants to start that way. Every run leaves a record.
Refund review · pilot · a checkpoint on every step, by choice
the same workflow runs in Cursor, ChatGPT and Codex
week one
now
example workspaces · each walled off · same governance
Describe how the work goes, in the AI tool you already use. ConvOps turns it into a workflow that every AI runs the same way.
Refund review · pilot · a checkpoint on every step, by choice
the same workflow runs in Cursor, ChatGPT and Codex
week one
now
example workspaces · each walled off · same governance
A checkpoint can wait for a person, or require a metric or a report before the run moves on. Keep them where production, money and customers are at stake. Leave them out where the work can run by itself.
Refund review · pilot · a checkpoint on every step, by choice
the same workflow runs in Cursor, ChatGPT and Codex
week one
now
example workspaces · each walled off · same governance
Nightly and weekly work runs on its own, each run in its own pod, at the times and in the time zone you set. Each run leaves a record.
Refund review · pilot · a checkpoint on every step, by choice
the same workflow runs in Cursor, ChatGPT and Codex
week one
now
example workspaces · each walled off · same governance
Your platform team installs one Helm chart, delivered with Enterprise, in your Kubernetes. Your models, your keys, your network. Nothing has to leave.
Refund review · pilot · a checkpoint on every step, by choice
the same workflow runs in Cursor, ChatGPT and Codex
week one
now
example workspaces · each walled off · same governance
Engineering, support, finance, each customer. One workspace each, walled off, with the same governance in every one.
Refund review · pilot · a checkpoint on every step, by choice
the same workflow runs in Cursor, ChatGPT and Codex
week one
now
example workspaces · each walled off · same governance
Refund review · pilot · a checkpoint on every step, by choice
the same workflow runs in Cursor, ChatGPT and Codex
week one
now
example workspaces · each walled off · same governance
Gates the engine checks on every move. A history for every task. Three roles. A workspace per team, business unit or customer.
A gate holds the step until a condition is met: an explicit approval is given, a required metric is recorded, or a required report is filed. The engine checks it, not the agent.
GovernanceEvery task and every workflow instance keeps its history. What ran, which step, and where it stopped.
GovernanceOwner, admin, member. Plain to grant, plain to review. No maze of permissions to audit.
WorkspacesOne per team, business unit or customer. Each one walled off in the database itself.
Workspacespolicies guide · gates hold Policies are guidance the AI reads. Gates, required metrics and required reports are what the engine checks. A gated step holds until an explicit approval is given, recorded with the account that gave it.
Examples, not customer stories. Each card names the feature that does the work, with a link to how it works.
Each customer file records a completeness check with a link to the document as evidence. The batch step holds until the check is recorded, and the trend shows the completion rate against the target your compliance team set on that step.
Metrics
One program groups the close schedules under one owner and goal. Each run files a reconciliation report, and the program tells you which runs are healthy, late, silent or failing.
Programs
Each control check files the same structured report: the control, the result and the evidence. The step holds until the report is filed, so a missing check is visible, not assumed.
Governance
Three schedules on one workflow, each at 08:00 local time in Singapore, Dublin and Chicago. Each schedule keeps its own time zone, and daylight saving does not move it.
Unattended runs
Your own runners claim work from ConvOps over HTTPS. Runs and secrets stay inside your network, with no inbound ports.
Executors
The Helm chart delivered with Enterprise takes one values file: the run namespace, pod posture, network policy, time limit and model credentials.
# install (example): the chart is delivered with Enterprise
helm install convops ./convops \
--namespace convops --create-namespace \
-f values.yaml
# values.yaml (example keys)
runs:
namespace: convops-runs
podSecurity: restricted
networkPolicy:
ingress: none
egress: limited
timeLimit: 4h
models:
existingSecret: model-credentialsOne call names the engine, a pinned version, the model, the time limit and the provider connections it runs with. A key on the executor is refused; each key lives on a connection, handed over once.
executors_manage({
"action": "create",
"workspace": "engineering",
"slug": "isolated-sonnet",
"runner": "convops",
"connections": ["anthropic-team"],
"config": {
"backend": "isolated",
"isolated": {
"agent": "claude-code",
"agent_version": "2.1.4",
"model": "sonnet",
"time_limit_s": 14400
}
}
})
// response (abridged)
{
"slug": "isolated-sonnet",
"connections": [{ "name": "anthropic-team", "provider": "anthropic" }],
"active_check": { "status": "queued", "trigger": "create" }
}A schedule fires on a cadence you pick, in the time zone you set. It creates the task and runs it on the executor you name.
{
"tool": "schedule_create",
"arguments": {
"title": "KYC file review {date}",
"vertical": "operations",
"project": "customer-onboarding",
"cadence": { "kind": "weekdays", "at": "01:00" },
"timezone": "Europe/London",
"executor": "isolated-sonnet"
}
}Members join a workspace with one of three roles. Invites wait to be accepted and expire after seven days.
{
"workspaces_members_manage": {
"action": "create",
"workspace": "northwind",
"email": "lead@northwind.example",
"role": "member"
}
}The controls, the boundaries and where each one stops. Written down before the first call.
No big-bang migration. We start on one real process of yours and grow from there.
We walk through one process you want to run on ConvOps, and where your models, runs and data need to live.
One workspace, checkpoints where your risk team wants them, your team in the loop. Weeks, not quarters.
Your platform team installs the Helm chart delivered with Enterprise. Teams move over one workspace at a time.
Bring one process and your security questions. We show ConvOps running it, and where it would live in your cluster.
ConvOps Enterprise is ConvOps, the operations layer for AI agents, run on your own Kubernetes or next to it. Your teams keep their AI clients; ConvOps holds the process as workflows with gates, a shared memory, schedules, programs, metrics and an audit trail. Self-hosting comes as one Helm chart, delivered with Enterprise, with support from the engineers who build it.
Self-hosted, ConvOps runs in your own Kubernetes cluster, in the region and under the controls you already operate. The engine, shared memory, the audit trail and every run live there, and nothing has to leave your cluster. In a hybrid setup, runs and secrets stay in your cluster and your runners claim work from ConvOps over HTTPS, with no inbound ports.
Your own. With ConvOps you bring the model provider and the credentials, handed over through a one-time link and masked in every log. Isolated runs use Claude Code or OpenCode, with the model you pin on each executor. People work locally with any AI client that connects to ConvOps, such as Claude, Cursor, ChatGPT or Codex.
Yes. Self-hosted, the whole of ConvOps runs in your cluster from one Helm chart, delivered with Enterprise. Runs reach only the model endpoints, repositories and connected tool servers you configure. Metric trends and program KPIs read from OpenSearch in your cluster; recording metrics and the gates that require them work without it.
Self-hosted ConvOps runs under a commercial licence, scoped with you on a call around your teams, your workspaces and your rollout. Programs, the Plan board, schedules and metrics are included, as they are on every ConvOps plan. There is nothing to sign before the pilot shows you it works on a real process of yours.
Neomanex, the team that builds ConvOps. You talk to the engineers who wrote the engine, from the first call through the rollout and after it. Support terms are part of the commercial licence and are agreed with you, around the teams, the workspaces and the processes you run on ConvOps.
Weeks, not quarters. A ConvOps pilot is one workspace and one real process of yours, for example a KYC file review, a month-end close checklist or support triage. You place checkpoints where your risk team wants them, watch the process run on your own work, and decide on the rollout after that.
No. Claude, Cursor, ChatGPT and Codex connect to ConvOps as they are, with a browser sign-in. Your teams keep the tools they use; the process, the gates and the record move into ConvOps. Unattended work runs as isolated runs with Claude Code or OpenCode, inside your cluster when you self-host.