User guide

Sign in

Your administrator sends you an invitation link, which lasts 48 hours. Open it, create an account with a password of at least 12 characters, and accept. On a hosted server you can also sign up yourself.

Two-step verification. The first time you sign in with a password, you set up an authenticator app, such as one on your phone. You also get backup codes. Each works once, so keep them safe. Every password sign-in then asks for a six-digit code. Hivelock does not remember devices.

Lost your authenticator? Choose Use a backup code at sign-in and tick I lost my authenticator. You then enter your password and a second backup code, and set up a new app, which gives you new backup codes. You need two unused codes for this, so do it before you run out. Two-step verification cannot be turned off.

Single sign-on. If your company uses it, choose Sign in with SSO and enter your work email. Your company's sign-in page handles the rest.

Confirming your address. Open the confirmation link in the browser you signed up with. If you open it anywhere else, or more than a day later, the password chosen at sign-up is removed, because anyone could have signed up with your address. You then get a link by email to choose a new one.

Forgot your password? Use the link on the sign-in page. The reset link lasts an hour, and works only if mail is set up. A session ends after about 8 hours without use.

Workspaces and roles

A workspace is a separate area with its own agents, workflows, connections, files and people. Workspaces cannot see each other. Switch between yours from the menu at the top of the side panel.

Your role in a workspace decides what you can do. Each role can do everything in the rows above it.

One setting is narrower: only an admin or owner can use an action that reads a mailbox, or add a mailbox trigger, because the results show whole messages.

Two things every workspace should know:

• People can approve their own runs. Anyone whose role can approve may approve or decline a run they started themselves. Hivelock does not require a second person. If you need two people, give the person who starts the run a role that cannot approve, or agree this as a rule in your team.
• Everyone in a workspace can read its chats. Any member, viewers included, can open and read every chat with an agent, including chats other people started. Chats are not private to the person who started them.

The Overview page shows live workflows, active runs, waiting approvals and recent activity.

Agents

An agent is an AI assistant with a job description and a list of things it may use. Open Agents and describe what you want in your own words. Hivelock drafts the agent, and you review it before it is saved. You can also start blank.

• Instructions are the agent's job description. Say what it does, how, and what to do when it is unsure.
• Model picks a role: Main for reasoning and tools, Fast for simple high-volume work, Coding for writing code.
• What it can use lists workspace files, actions on your connections, reviewed MCP tools, skills, document search and memory. An agent gets only what you tick.

The platform sets the limits and safety rules. A custom agent has at most 12 steps and 60,000 tokens per run. Agents never see credentials, and an agent you create cannot create or manage other agents.

Each workspace also has built-in agents, such as one that plans a request and hands parts to other agents.

Skills are reusable instructions that any agent in the workspace can follow.

Memory lets an agent keep notes for future chats. You approve each note before it is saved. To see or remove them, open what the agent remembers and choose Forget this.

Tests check an agent still does its job after a change. You give a message, and words the answer must or must not contain. Tests use real model calls, and any action that needs approval is refused. A model never grades itself, because fixed rules score the answer.

Chat

Open an agent to chat with it. Replies appear as they are written, unless your workspace removes personal data from answers, in which case they appear when finished. Earlier chats are listed under Past chats. You can attach a file, which is added to the workspace's files.

If the agent wants to do something that changes data, the chat pauses and asks. Someone with approval rights decides. If nobody decides within 15 minutes, it counts as declined. A chat turn is never retried automatically, because its tools may already have acted. If a worker stops mid-turn, send another message.

Approvals

Approvals are how a person stays in charge. There are two kinds:

• In a chat, the question appears in the chat itself.
• In a workflow, an approval step pauses the run. It appears on the Approvals page with the run's input, and waits until someone decides.

Operators and above can decide. Read what the agent wants to do, and decline if you are unsure. Decisions are recorded in the audit log.

Workflows and the builder

A workflow is a chain of steps that runs by itself. Open Workflows, create one, and drag steps onto the canvas. Connect the dots to join them.

Passing data along. Use placeholders such as {{input.customer}} or {{steps.agent_1.output}}. Values from outside are marked as data, so an agent is told not to obey instructions inside them. That helps, but a model can still be fooled. Placeholders are refused in SQL, SOQL, JQL and Airtable formulas, because outside text could change the query.

Rules. Steps must move forward, so loops are not allowed. A workflow can have up to 100 steps.

Running it. You can run a workflow by hand at any time. New workflows are drafts, so turn a workflow on before its triggers count. Follow each run under Runs, and cancel it there if you need to. A step that fails is tried up to three times, and a step that changes data is never repeated on a guess. If a worker dies, another takes over within a couple of minutes, and the run continues. A run always keeps its own copy of the workflow.

Version history

Every save keeps a numbered version, with who saved it, when, and a note if you added one. Open the history panel in the builder to see a version, or to compare two versions step by step. Moving a step on the canvas does not count as a change.

Restoring saves a new version and discards unsaved changes. Nothing is overwritten. The latest 100 versions of each workflow are kept. Older ones go overnight, unless a run in progress used them.

Triggers

Triggers start a workflow while it is turned on. Add them in the builder.

Schedule

Uses a five-field cron schedule and a time zone. For example, 0 9 * * 1 means 9:00 every Monday.

Webhook

Another system starts a run by calling a web address. It must sign each call:

• Send a POST to /hooks/<trigger id> with a JSON body of up to 1 MB.
• Set x-webhook-timestamp to the time in Unix seconds.
• Set x-webhook-signature to the hex HMAC-SHA256 of <timestamp>.<body>, using the signing secret.

The signing secret is shown once, so copy it straight away. Calls older than five minutes and repeated calls are refused.

Email

Watches one folder of an email connection over IMAP. An admin adds it.

• It checks every 1, 2, 5, 10, 15, 30 or 60 minutes.
• Each new message starts one run, up to 20 per check. The rest wait for the next check.
• The run gets input.from, input.subject, input.text and input.attachments. Attachments are names only, and long text is cut.
• The mailbox is opened read-only, so nothing is moved, flagged or marked as read.
• Only mail arriving while the workflow is on counts. Mail from the mailbox's own address is ignored, so a workflow cannot start itself.

Sender filters narrow mail down, but they do not prove who sent it, because a sender can be forged. Treat all mail as untrusted input. Microsoft 365 connections cannot be used here. Use an email connection with an IMAP server.

Templates

Templates holds 74 ready-made agents and workflows for 11 teams: sales, customer support, finance, people and HR, operations and purchasing, quality and compliance, legal, IT and security, engineering, marketing, and executive. 25 are agents, and 49 are an agent with a workflow. Search them, or filter by category or role.

Each one shows what it does, what it connects to, what needs approval, and its steps. Answer a few questions and choose your connections, then choose Install as a draft.

• Installing creates no triggers, and the workflow stays off. You turn it on when you are happy.
• Read the agent's instructions, and test on made-up data first.
• Templates marked A person decides feed decisions about money, people or legal matters. The agent prepares a draft, and a person decides.
• On a hosted server, the 15 templates that use local file tools cannot be installed.

Connections

A connection links a workspace to an outside service. An admin creates it under Connections, and builders can use it. See Connectors for each service.

Credentials are encrypted when saved and cannot be read back. Use Test to check one works without changing anything. Give each connection the least access it needs. For MCP servers, press Check server, read each tool, and turn tools on one at a time.

Files and document search

Upload PDF, Excel (.xlsx), CSV and text files of up to 15 MB under Files. Scanned PDFs and old .xls files are not supported. Self-hosted workspaces can hold 500 files. Hosted workspaces are limited by storage. You can list and delete files in the app, but there is no download button.

New files are split into passages and indexed by themselves. Older files have a Make searchable button. Turn on document search for an agent, and it finds passages and names the file it used.

• Keyword search always works and needs no AI. It is built for English, so other languages may match less well.
• Add an embedding model to a provider, and search also finds meaning. See AI models.
• Files are sent to your AI provider when an agent reads them. Use made-up data until you have checked your provider's terms.

Channels and WhatsApp

A channel lets people message an agent from WhatsApp through Meta's official Cloud API. On a hosted server this needs the Team or Business plan.

1. Create a Meta app with WhatsApp, and add a WhatsApp connection with its token, phone number ID, app secret and any long verify token.
2. Under Channels, choose that connection and an agent, and list the numbers allowed to use it, with country code.
3. Paste the channel's webhook address and verify token into Meta's webhook settings, and subscribe to messages.

Meta must reach the webhook, so this needs a public HTTPS address. The list starts empty, so nobody can reach the agent until you add them. Other numbers get no reply. Only text messages are read.

Actions that need approval wait for a person in the app, and cannot be approved from WhatsApp. Messages pass through Meta and your AI provider. This has not been tested with the real WhatsApp service.

Observability

The Observability page shows the last 7, 30 or 90 days: tokens, cost, speed, errors, refused actions, and tables by agent and model. Each run and chat links to a trace of its model and tool calls.

• Cost shows only if you enter prices for a provider. Otherwise it shows as unknown, not a guess.
• Traces hold measurements such as time, tokens and status. They hold no prompts or results.
• Tool calls to look at lists high-risk, failed and refused calls. The risk score comes from fixed word rules. It informs people and blocks nothing.
• A flagged injection attempt is a warning sign, not a defence. Wording varies easily.

Audit log

Admins and owners can read the Audit log. It records actions such as single sign-on sign-ins, invitations, role changes, changes to connections, runs, approvals and settings. The page shows the newest 200 entries, and whether the chain is intact.

Each entry is linked to the one before it by a hash, so a changed record shows up. The page also shows the latest hash. Copy that somewhere outside Hivelock regularly, because someone with full database access could otherwise rewrite the whole chain.

Settings

Checked against the code on 6 November 2026.