Security

Who is asking, what they can reach, and what they left behind.

Identity, isolation, and credentials are decided before work reaches a model or a tool. Sensitive steps can wait for a person. Every run stays traceable.

  • No access beyond the person who asked
  • Isolated environment per company
  • Approval before sensitive writes
  • An audit record of every run

Every request

Four checks before work leaves the workspace.

Security and identity decide the door. Observability and governance decide what waits, and what you can reconstruct later.

  1. 1

    Identity

    SSO and MFA, when you use them, confirm who is asking. The agent runs with that person’s permissions.

  2. 2

    Scope

    Workspace, project membership, and the tools that project has turned on decide what the run can reach.

  3. 3

    Checkpoint

    Personal data is redacted before a model call. Workspace data stays off OpenAI, Anthropic, and Google.

  4. 4

    Record

    Writes that matter wait for a person. The run, the decision, and the change stay visible afterward.

One morning, one workspace

Identity, access, secrets, and approvals in the same record.

Joining a project, rotating a token, a denied folder, an approved write, and a finished flow are the same kind of fact: who did it, in which workspace, and what changed.

Workspace activity
  • Ana Silva joined the Finance project

    Invited by Lena Park · project membership

  • Secret rotated: SALESFORCE_TOKEN

    Admin · org secrets · value stays out of chat

  • Access denied: Payroll folder

    Caller is not a member of the People project

  • Approved: update 12 Salesforce records

    Sales lead · ask-first on the Sales ops project

  • Flow finished: Monday invoice reminder

    Read late invoices · mail waited for approval

  • PII redacted in 4 fields

    Email, phone, IBAN · before the model call

Illustrative events from one workspace morning.

Security and identity

Identity and access

Members, roles, and the person behind the agent.

People join a workspace by invite. Admins manage members, roles, and groups. Projects and topics have their own membership, so a finance project stays with the people on that project.

An agent runs with the permissions of the person behind it. Optional SSO and MFA go through your identity provider. Sensitive actions stay reviewable.

  • Workspace roles for admins, members, and groups
  • Project and topic membership on top of the workspace
  • The agent’s door is the same width as the person who asked

Security and identity

Workspace isolation

A separate environment for each company.

A workspace is one company: its people, projects, topics, agents, and connected tools. Work in one workspace stays in that workspace. Another company on Kluff does not share the environment, memory, or secrets.

If you belong to more than one company, you switch workspaces. Billing, team, and audit history belong to the workspace you are in.

  • People, memory, and secrets stay inside the company workspace
  • Switching workspaces switches billing, team, and audit history
  • Project files stay isolated from other projects, including personal drafts

Security and identity

Encrypted secrets

Credentials applied at runtime, kept out of chat.

Integration tokens and other credentials live as encrypted secrets. Agents and flows receive them when a tool needs them. They stay out of prompts, threads, and canvases.

Rotate a secret from Settings if it leaked elsewhere. Least-privilege scopes stay on the integration: only the access that workflow needs, and still reviewable.

  • Encrypted at rest, injected when a tool call needs them
  • Shared MCP servers connect once, with scopes that match your policies
  • A project only gets the connections and MCP tools you enable there

On the way to a model

Data handling around every model call.

Identity decides who may start the request. These controls decide what a model is allowed to see, and which tools the project may use.

  • Private model access

    Use state-of-the-art models through Kluff. Workspace data is not sent to OpenAI, Anthropic, or Google.

  • Redacted PII

    Personally identifiable information is redacted before it reaches a model. That is the default path for chat, agents, flows, and canvases.

  • Least privilege

    Integrations request only the scopes a workflow needs. An agent cannot use a tool the project has not turned on.

  • Shared MCP servers

    Connect an MCP server once at the organization. Projects opt in, and every call stays inside the scopes you set.

Redaction is the default path. You do not turn it on per thread. Output checks and ask-first approvals remain separate controls.

Observability and governance

Approvals and control

Review an action before it runs.

On each project, tools can be ask-first or automatic. Ask first is the right default for anything that emails a customer, writes to a CRM, deploys, or spends money. Automatic fits read-only lookup once you trust the connection.

When an agent needs permission, a card appears in the thread and in Notifications. Approve, decline, or continue with a different approach. Declining stops that action.

  • Ask-first or automatic, set on the project where the work happens
  • The same card in the thread and in Notifications
  • A declined action does not run

Observability and governance

Run history

Follow a workflow from the trigger to the result.

Each flow run has a history you can inspect: what ran, what failed, and what came out. Agent threads leave artifacts — documents, diffs, schedules, pull requests — as the reviewable record of the work.

Everyday review happens in the thread, in Notifications, and in the run you already opened. Admins also have audit history for sensitive actions in the workspace.

  • Flow runs: what ran, what failed, what came out
  • Agent artifacts beside the conversation that produced them
  • Audit history for sensitive actions, scoped to the workspace

Observability and governance

Usage

See what the workspace ran.

Usage answers what the workspace ran. Tokens and cost are grouped by model kind: faster, default, and thinking.

If a project is expensive, look at how often it uses deeper thinking models or long agent threads. Tighten approvals, or move experimental work into a personal project so it stays out of the shared one.

  • Tokens and cost by faster, default, and thinking
  • A view of activity for the workspace, not a model-provider console
  • Audit logs sit next to usage for admins reviewing sensitive actions

For a security review

The question, and where the answer lives.

QuestionAnswer in Kluff
Who can see this project?Workspace membership, plus the people invited to the project or topic.
What can the agent reach?The permissions of the person who asked, and the tools that project has enabled.
Which company is this?The active workspace. Billing, team, memory, secrets, and audit history belong to it.
Did a write wait for a person?Ask-first on that project. The card is in the thread and in Notifications.
What ran overnight?Flow run history, and the artifacts left on the agent thread.
Where did the spend go?Usage by model kind: faster, default, and thinking.

FAQ

Questions for buyers and reviewers.

Identity, data handling, approvals, and the record those actions leave.

Every request carries an identity. The agent uses that person’s permissions inside one company workspace. Personal data is redacted before a model call, secrets stay encrypted and out of chat, and sensitive actions can wait for approval. Runs leave a record.

No. Kluff keeps customer data off those model providers. Personally identifiable information is redacted before a model call, and company context stays in Kluff.

Yes. Kluff supports SSO and multi-factor authentication through your identity provider. They are optional and configured with that provider.

Integration tokens and other credentials are encrypted secrets. Agents and flows receive them at runtime. They are not copied into prompts, threads, or canvases. Access follows organization permissions, and you can rotate a secret from Settings.

No. The agent runs with that person’s permissions. Project membership and the tools enabled on the project narrow it further. A connection someone else made in the workspace stays unavailable until that project turns it on.

Yes. Each company has its own workspace: people, projects, memory, connected tools, and secrets. Switching workspaces switches billing, team, and audit history.

When that tool is set to ask-first on the project. That is the default to use for emailing a customer, writing to a CRM, deploying, or spending money. The card shows in the thread and in Notifications. Declining stops the action.

Open the flow run for what ran, what failed, and what came out. Agent threads keep artifacts — documents, diffs, schedules, pull requests — next to the conversation. Admins can review audit history for sensitive actions in the workspace.

Tokens and cost by model kind: faster, default, and thinking. It is a view of what the workspace ran. If one project stands out, look at deeper thinking models and long threads, then tighten approvals or split experimental work into a personal project.

Unlimited inference.
Company memory.

Connect your tools, ask your first question in a project, and watch memory, spend, and visibility stay in one place—not scattered across personal chats.

$15 / user · Unlimited AI
Q4 launch› Launch dashboard
Sara Castillo

@kluff build a launch dashboard from HubSpot and the Q4 budget sheet. Refresh it every morning.

Kluff

Done. Signups come from HubSpot, spend from the Q4 budget sheet. It refreshes at 7:00 AM.

v2 adds the channel split you asked for.

Maya Ruiz

Can we flag any channel over $40 per trial?

Kluff

Added. Paid social is flagged.

Q4 launch dashboard
Q4 launchHubSpot + Q4 budget sheet · refreshed 7:00 AM
Live
Signups1,284+18% vs last week
Trials started31224% of signups
Spend$8,420of $12,000 budget
Daily signups
MonTueWedThuFriSatSunMonTueWed
Send the renewal to Harbor Foods?

It sends from your Gmail, as you, with the draft attached. Nothing else leaves the project.

SENDGmail · Dana Reyes, Harbor Foods purchasing
ATTACHGoogle Drive · Harbor renewal 2026.docx
Finance check done

Ana Chen confirmed net 45 at 9:18 AM. Your rule: every customer email waits for you.

Sent · logged to the project