AI for the whole team, in one place.
Most AI tools are one person and one chat. Toolbelt is built for teams, yours and your customer's: everyone works with the same assistants, in the same chats, at the same time, and sees the same results. Connect a tool once and share it with the whole organization, and every assistant can use it without anyone handing over a password.
Northgate Supply / Order desk3 in this chat
DKDana KimNorthgate Supply
Can you check whether the Hollis order shipped and let them know?
ODOrder deskAssistant
It shipped yesterday, two cartons, tracking ending 4417. I've drafted a note to Hollis from your inbox. Want me to send it?
orders.lookup ran on Dana's connection
RMRiley MossYour firm
Looks right. Dana, it's your inbox, so it's your call.
Shared assistants.
Share one with the whole organization or with chosen people, to use or to edit. Change its instructions once and everyone has the new version.
Shared chats.
Your team, the customer's team and the assistant in one conversation. Each person's actions run on their own logins, so nobody borrows the owner's inbox.
Shared connections.
Connect a service, share it with the organization, and every shared assistant and workflow can use it. The credentials stay hidden, and you can take it back any time.
Roles and switches.
Owners, admins and members, with switches for who may publish, upgrade, share chats, manage services or change tool permissions.
Assistants that work together.
One assistant hands a task to a specialist and brings the result back into the same thread.
Wherever people already are.
In Toolbelt, in Slack, on a phone, or from Claude, ChatGPT and Cursor.
A space for every customer.
Each customer gets its own organization on Toolbelt: its own people, logins, connections, files, memory, skills, plan and history, separated from every other customer. Create one in a moment or link one that already exists. It runs under the customer's own agreement with Apexti, so if you ever part ways, nothing has to be moved.
Its own plan and allowance.
Paid usage stops at the allowance unless you allow more, up to a dollar limit you set.
Your people and theirs.
Both teams are members with roles. Give the customer's staff a simplified view, chat and home only, if that's all they need.
What it learns stays home.
Memory, skills and files live in the customer's space, not in your firm's and not in a model provider's account.
All of them on one screen.
Org Ops shows every customer's plan, spend against allowance, assistants and last use. The ones over 85% of their allowance are named at the top. We set up Org Ops with you.
Connected to everything. Allowed precisely.
More than 2,000 connections through MCP, from the CRM and the inbox to the order system, plus any remote MCP server by its address and custom MCP servers you build in chat. Connections are made in the customer's own name and stored encrypted. Every tool on every connection is set to allow, ask or deny, so access to a CRM is not permission to do everything in it.
Connections / CRM / Tool permissions
Read onlyAsk on writesAllow all
| contacts.search | AllowAskDeny |
| deals.read | AllowAskDeny |
| contacts.update | AllowAskDeny |
| deals.create | AllowAskDeny |
| contacts.delete | AllowAskDeny |
Allow, ask or deny, per tool.
Start from a preset: read only, ask on writes, or allow all. An asked tool pauses for a person to approve it.
Each person on their own logins.
A schedule or workflow can run once for each member, using that member's own connections.
Behind the firewall too.
The Bridge reaches tools and models on a private machine through an outbound connection.
Locked where it matters.
A tool switched off at the organization level cannot be switched on by an assistant below it.
Built to run, not to re-think.
A model is for working things out. Once a step is worked out, Toolbelt turns it into something that runs without the model: saved code, typed decisions, SQL over large results. Runs get faster, cheaper and steadier the longer a solution lives.
Wrenches.
Saved, versioned code that calls tools in a fixed order, with no model call inside. The hundredth run costs a fraction of the first and gives the same result for the same input.
Workflows.
Steps, branches and parallel groups, started by a schedule, a webhook or a person. A tool-call step runs with no assistant turn at all.
Classifiers.
Typed decisions, a label, a score or yes or no, answered by a small decision model without waking an assistant. Triage, routing and gates.
Large results into a database.
When a tool returns more than a conversation should hold, the result lands in a database and the model asks for what it needs with SQL. Nothing is silently cut.
Delegation.
Sub-chats do a piece of work in their own context, on the model you choose, and hand back only the result.
Design.
One map of an assistant: its jobs, schedules, wrenches, files, databases, dashboards and services, with a planner that drafts a job from a sentence and flags what is missing.
How this keeps costs down
Any model. A setting, not a rebuild.
Pick the model per assistant, per sub-task and per schedule, and change it in the middle of a conversation. When a provider retires a model, Toolbelt maps it to its successor. Because memory, skills and files live in the customer's space rather than in a provider's account, nothing moves when the model does. Limit which models an assistant or a role may use, and bring your own endpoint when a customer requires it.
OpenAIAnthropicGooglexAIInceptionAny OpenAI-compatible endpointLocal models through the BridgeFree models hosted by Toolbelt
Toolbelt-hosted models cost no credits, and keep working when credits run out.
Delivered where people already work.
A solution is only finished when the customer can use it without thinking about where it runs. Toolbelt hosts the whole thing.
Dashboards and web apps.
On the customer's own domain, open or behind a one-time sign-in link sent by email, carrying your logo and colors. HTTPS is set up for you.
Slack.
Chat with an assistant by command, direct message or mention, and let it post results to a channel.
The Toolbelt app.
On a desktop or as a home-screen app on a phone, with a simplified view for people who only need chat.
Schedules and triggers.
Work that runs at 7 a.m. Monday whether anyone is online or not, with every run kept as its own chat.
Your other tools.
Every assistant doubles as an MCP server for Claude, Claude Code, ChatGPT, Cursor, Windsurf and VS Code, and there is a REST API with webhooks into workflows.
See what ran. Prove it. Undo it.
The retainer is for noticing before the customer does. Toolbelt keeps the evidence in one place.
Spend, by anything.
Every customer's credits against allowance, by assistant, by model, by person and by day, with a projected end of month. Export the ledger as CSV.
Failing first.
Across every customer, failing schedules sort to the top of one screen, and the schedule's owner is notified inside Toolbelt.
Every run, kept.
Each scheduled run is its own chat you can open and read. Each wrench run logs its inputs, outputs and console. Each workflow run shows the result of every step.
Every change, undoable.
Who changed what, when, and what it was before, for assistants, schedules, connections, wrenches, files and members, with an Undo button.
Publish once. Roll out with care.
Publish an assistant to your catalog as a version. Try it on a beta ring of the customers who like being first, then upgrade everyone who is behind with one button. Instructions, model and services update; new wrenches and dashboards are added; what a customer built on top stays theirs. Nobody moves until you move them.
Secure by construction.
The questions a customer's reviewer asks have answers that are true by design, not by policy.
Checked before it ships.
Scarlett, Toolbelt's built-in penetration-testing agent, runs the first pass on an app and reports what to fix.
No passwords to hold.
Connections are made in the customer's own name, stored encrypted, and never shown to other users. Sign-in is federated through OAuth and OIDC.
Separated by organization.
Each customer's data is separated by organization and assistant, with permission checks on every request.
No training on your data.
Not by Apexti, and every model provider on the platform is engaged under terms that prohibit it. Providers receive the minimum context the request needs.
Hosted primarily in the United States.
Access to production is limited to authorized Apexti engineering and security staff. The data processing addendum and the list of sub-processors are published.
Build it here, or where you build.
Describe what you want in chat and Toolbelt builds the assistant, the wrench, the workflow or the dashboard, with a preview link and every version saved. Or keep your editor: every assistant has an MCP address, so Claude, Claude Code, ChatGPT, Cursor, Windsurf and VS Code can build and run against it, with the assistant's instructions enforced on every call.
The hundredth customer is easier than the first.
Every solved step becomes code you keep. Every customer's context stays in its own space, ready for the next model. Every connection is already scoped, every permission already set, every run already logged. Your catalog grows, your morning check stays one screen, and the work that used to be the retainer's burden becomes the retainer's proof.
- Credits are checked before each turn and before each paid tool call.
- Code runs in a sandbox: JavaScript, TypeScript, Python and shell.
- Databases are DuckDB files in the assistant's own storage, next to its files.
- Every assistant has an MCP address and a scoped API key of its own.
- A public page can live on a custom domain; HTTPS is provisioned automatically.
- Plans start free. Paid plans are priced in dollars, with usage in credits you can see.
Plans and what they include