Your AI team clocks in before you do.
Clovar gives you AI coworkers that hold real roles, work inside your ERP, inbox and spreadsheets, and hand back finished work under rules you set. Built on Claude.
Prefer your own servers? Self-hosting is available too.
A simulated morning using the platform's real tools. Every line is something Clovar agents do today.
Works with what you already run.
Agents don't need a new system of record. They work inside the ones you have.
- HubSpot
- Salesforce
- QuickBooks
- Slack
- Microsoft Teams
- Microsoft 365
- Gmail
- Meta
- X
- Google Ads
- Calendly
- Buffer
- Metricool
- PostgreSQL
- SQL Server
- MySQL
- any MCP server
- any OpenAPI spec
Your ERP, live
Agents query part data, pricing, stock and purchase history straight from your ERP over the Model Context Protocol. They answer with your numbers, not guesses.
Your inbox, both directions
Gmail and Microsoft 365 with real-time push. Agents triage, draft, attach and send. Anything going out can wait for your approval first.
Anything with an API
Paste in an OpenAPI spec and Clovar builds the connector. Every connector runs isolated and code-signed, with its permissions reviewed before it touches anything.
Built on Claude.
Every agent runs on Claude. Pick the model that fits the job, set a default per business, and override it per agent.
- OpusJudgement calls, like the Finance Reviewer weighing a purchase order.
- SonnetThe daily work: lookups, spreadsheets, drafting, coordination.
- HaikuTriage and routing, where speed and volume matter most.
OpenAI, Azure AI and Ollama are supported too.
-
Prompt caching, on by default
An agent's standing context, its role, its contract, its tools and its skills, is cached so turn after turn costs a fraction of the first.
-
Programmatic tool calling for big datasets
For a 2,000-row workbook the agent writes a script that runs in Anthropic's code execution sandbox and calls Clovar's tools. Paged ERP rows flow through code instead of the model's context. The sandbox holds no credentials; every call is executed and authorised on the platform side.
Live pilot: four tool round-trips in a row at zero model tokens.
-
Your ERP over the Model Context Protocol
Part data, pricing, stock and purchase history reach the agent through MCP, scoped to what that agent is allowed to see.
-
Your key or ours
Bring your own Anthropic key and pay Anthropic directly with no markup, or let Clovar Cloud include inference.
Hire a role, not a chatbot.
Each one comes with a persona, a toolset and a contract. They start useful on day one.
- Operations CoordinatorManages the others. Assigns, reviews, escalates.
- Inbox AssistantReads the shared mailbox and answers what it can.
- Procurement BuyerPrices parts against the ERP and prepares requests.
- Finance ReviewerChecks anything that costs money.
- Support EscalationOwns the hard cases until they close.
- Executive BriefingTells you what happened and what needs you.
- Security AnalystReviews access and connector permissions.
- Supplier Feed IngestersFive roles that live on supplier feeds: price updates, datasheets, end-of-life notices, last-time-buy warnings and lifecycle bulletins.
Or write your own. A role is a persona, a toolset and a contract. Nothing more exotic than that.
They work like a team, because they are one.
Managers and direct reports, shared memory, owned work, and a budget that means something.
-
They show up on their own
Agents run on a schedule, react to an event, or answer when spoken to. Work happens overnight without anyone prompting it.
-
They work as a team
Managers delegate down and reports escalate up. They brief their peers and borrow a specialist from another department when the job needs one.
-
They remember
Facts and relationships persist in a knowledge graph, found by meaning rather than keyword. Tuesday's agent knows what Monday's learned.
-
They own the work
Tasks carry dependencies, priorities and locks so two agents never fight over one job. Cases stay open across days and close when they're actually resolved.
-
They show you the plan first
For bigger jobs an agent lays out what it intends to do before it does any of it. Read it, change it, approve it, then watch it execute exactly that.
-
They get better, inside a budget
Lessons from their own work are promoted into how they operate. Test two variants and keep the winner. Spend is tracked per agent, with a ceiling that stops work rather than surprising you.
From trigger to finished work.
Seven steps, every time, with your rules applied before anything touches the outside world.
Something kicks it off
A schedule fires, an email lands, a workflow reaches its turn, a colleague delegates, or someone simply asks.
It gathers the context
Memory, knowledge, open tasks and unread briefings come together into one picture, so the agent starts informed.
It decides what to do
Claude works through the job in the agent's role and persona, planning the steps rather than answering a single question.
It takes action
Every tool call is checked against what that agent is allowed to touch before it runs. Records, files and external systems update for real.
Your rules apply
Anything with a consequence is weighed by risk, cost and how much trust this agent has earned. It proceeds, waits for approval, or is refused.
A human signs off
High-stakes actions queue with the evidence and the reasoning attached. Approve, reject, edit the draft, or ask for more.
It remembers what happened
Facts land in the knowledge graph. The conversation is summarised with the entities that mattered, so the next agent starts ahead.
Full control without slowing agents down.
Agents earn scope. Every consequential action is weighed against your rules, and all of it is on the record.
- Observe only
- Draft only
- Recommend only
- Approval required
- Policy-bound autonomy
- Emergency stop
Trust is earned one level at a time. Agents graduate by proving themselves, and one control halts everything in flight.
-
A written job description, enforced
Each agent works under a versioned contract: its role, the systems it may touch, what it must never do, the evidence it has to produce, and what it may spend. Contracts move through draft, staging, certified and production, and become immutable once published.
-
Ask it why, get a real answer
Every approved or rejected action carries a structured account: what triggered it, how the agent reasoned, which policy applied, what else it considered, and what actually happened.
-
Rehearse before you go live
Run an agent in simulation to see what it would have done. Certification suites test a contract before production, and past runs replay as regression tests.
-
Nothing leaves unrecorded
Every action that touches the outside world is written to a ledger before it happens. Secrets are referenced by name, never seen.
-
Proof it's working
Resolution rate, approval rate, value delivered and a health score, so whether this is paying off has a number attached.
What this looks like on a normal Tuesday.
Three jobs that used to belong to someone's afternoon.
The 2,000-row spreadsheet
A customer sends a stocking-level workbook with half the columns empty. An agent reads it, looks up every part in the ERP for cost, lead time and last sale, fills the columns, and emails the finished file back.
1,984 rows, 9,870 cells, 0 errors, one operation.
The inbox that handles itself
Mail is triaged in seconds. Routine questions are answered from your own documentation, quotes are drafted with live pricing, and anything unusual reaches a person with the full thread and the reasoning attached.
Routine mail answered. Unusual mail escalated with context.
Buying under control
A buyer agent watches demand, prices against live ERP data and prepares the request. Anything over your threshold waits for a human, with the numbers, the alternatives and the reasoning already laid out.
Over threshold: waits. Under threshold: proceeds. Either way: recorded.
Run it your way.
Managed by us, or on servers you control. Same platform either way.
Clovar Cloud
We host it, run it, secure it and keep it current. You bring the work.
- Fully managed. No servers, no upgrades, no on-call.
- Inference included. No API keys to manage.
- Dedicated tenant with isolated data and secrets.
- Guided onboarding: your systems connected, your first roles live.
- Onboarding a limited number of businesses.
Your servers
Runs on your own infrastructure. Your database, your keys, your network.
- A small set of containers your IT team installs and runs.
- Bring your own Anthropic key. No markup on tokens.
- A conventional stack your IT and security teams can audit.