One doorway to every system your agents touch.
Nucleus sits between AI clients and the company's systems. Claude, a Dispatch box, or a Slack bot connects to one MCP endpoint and sees only the tools that person is allowed to use. Every call is authorized before it leaves, runs as a named person, is metered, rate-limited and written to an audit trail. Nucleus also runs workflows against those tools in the cloud, with nothing to install.
Clients ask. Nucleus decides. Systems answer.
An agent with a credential can do anything the credential can do. Nucleus is there so no client ever holds one. The client presents its own key, and Nucleus decides whether this person may call this tool, now, at this rate, within this budget. It holds the system credentials itself and makes the call on the person's behalf.
Nucleus
- Who · a named person, from their own key
- Allowed? · default-deny, per install and tool
- Rate · tenant, person and key buckets
- Budget · reserve, call, settle
- Record · ledger, audit, traces
Built on the assumption that tools are powerful.
Whatever connects, it gets the same controls. Nothing here depends on the client behaving well.
Each person sees only their tools
The tool list is built per person from access rules. A tool someone may not call is not on their list, so an agent cannot reason its way toward it.
Default-deny, every call
Access is checked again on each call, not only when the list is built. Revoking access takes effect on the next call, not the next session.
Secrets stay in the vault
System credentials are stored encrypted with each server install. Clients present their own key. No system password ever reaches a laptop, a prompt or a model.
Priced per tool, on a ledger
Each tool carries a credit cost. Calls reserve against a budget before they run and settle afterwards, and every event is recorded on an append-only ledger.
Rate limits at three levels
Token buckets per organization, per person and per API key. A runaway loop hits its own limit first and stops there, before it can affect the shared system or the budget.
Observable in production
Metrics, logs and traces for every call, plus an audit log of security events. When something goes wrong, you can see who called what, and when.
Workflows that run in the cloud, with nothing to install.
Nucleus goes beyond granting access to tools. It also chains them into workflows that run unattended, in the same way Dispatch does, but hosted, so a team can automate the moment it signs in.
In the browser
Lay out the steps in the workflow builder: which tools to call, in what order, and what each step passes to the next. There's no server to stand up and no runtime to maintain.
In the cloud, under the same rules
Every step is a governed tool call, run as the workflow's owner, rate-limited, priced and recorded like any other call. Automation gets no back door.
Onto your own box
When a job needs local models, your own MCP servers or systems only your network can reach, run it on Dispatch instead. It still uses the same tools, rules and record. Cloud or box ›
Start as tools. Grow into an app, without leaving Nucleus.
A new product can begin as a database and a set of MCP tools that Claude drives. When the team needs screens, Nucleus hosts a web app in front of the same tools, with sign-in, permissions and live updates already in place.
Tools and a database
Install it like any MCP server. Claude is the interface while the business rules take shape against real data.
Register an app
Add two fields on the server's catalog entry, and the app appears at /apps/<name>/ behind Nucleus sign-in, with every request permission-checked.
People and agents, live
The app and the tools share one implementation, so a draft Claude rewrites shows up on a person's screen within about a second. See Protolyth ›
Every action has a name on it.
Any meaningful action can be traced to the same five things, which is what makes it billable, auditable and safe to let agents act on.
The same tools from wherever the work starts.
A tool is installed once, in Nucleus, and every client gets it. The person asking at a desk, an automation running overnight on a box, and a teammate typing a command in Slack all reach the same tool, with the same rules and the same record.
Claude
Ask a question about a customer or a project, and Claude answers from live systems through Nucleus, within your own access. Nothing to export, and no credentials to paste in.
Dispatch
When a workflow needs a local model, a local MCP server or a system on your own network, it runs on a Dispatch box instead of in the cloud. It still calls the same Nucleus tools under the same rules. How the two fit ›
Slack
A slash command can start a workflow, and the result is posted back to the channel the team already reads. Nobody has to learn a new screen.
From gateway to running work.
From MCP tools to a full app
Protolyth started as a database and tools in Claude, and two weeks later it was a team app hosted by Nucleus. See the screens.
› ArchitectureNucleus and Dispatch, in unison
Both run workflows. Nucleus runs them in the cloud and is the quickest way to start. Dispatch runs them on your own box, with local models, local MCP servers and the systems on your own network.
› Use caseTranslated pages crawlers actually see as current
99.84% of 20,728 URLs current, against 23.8% of 8,474 on a hand-built cron job. The same cache, kept fresh by Nucleus and Dispatch.
› PlatformWhat Dispatch is
The self-hosted engine for deterministic, agentic workflows, running on your own hardware.
›
nucleus