Skip to main content
    Log inBook a Working Session
    PRODUCT · ACTIONS

    Let AI act in your systems, under your rules.

    Actions give chat and agents the tools to do work in connected systems, not just answer about them. Built-in tools cover web search, code execution, and image generation, and custom actions call your own APIs through OpenAPI or MCP, all inside the access and approval rules you set.

    Book a Working Session See it work
    Starts withA task that needs a change in a system
    Grounds inBuilt-in tools and allowed custom actions
    ProducesPrepared or completed work with visible state
    BoundaryTool access and approval policy apply
    THE PRIMARY JOB

    Answers are useful. Completed work is better.

    Actions extend chat and agents from reading to doing. An agent can search the web, run sandboxed code, or call a custom action that updates a ticket, checks a service, or moves a deal, while consequential writes stay visible and approval-gated.

    1. 01

      Context is copied by hand from one system into another.

    2. 02

      The person approving a change cannot see its exact target or payload.

    3. 03

      Partial failures leave the team unsure which steps actually ran.

    LActionsProject AtlasApproval required
    Request

    Follow up on Atlas Retail and assign the missing implementation note.

    Exact payload · Gmail · Atlas Retail thread
    To
    alex@atlas-retail.example
    Subject
    SSO implementation note and Friday patch
    Escalation threadSSO Runbook v4
    Approval required2 prepared writes are waiting for an authorized reviewer.
    HOW THE PRODUCT WORKS

    Three visible parts of Actions.

    01

    Use the built-in tools out of the box.

    Internal search, web search, code execution, and image generation are available to chat and agents once an administrator enables them.

    Prepare and executeFour built-in tools, admin controlled
    02

    Connect your own systems as actions.

    Define custom actions over OpenAPI or MCP so an agent can read or write in your tools through their real APIs.

    Prepare and executeCustom actions via OpenAPI and MCP
    03

    Run code in a sandbox.

    The code execution tool runs Python in an isolated sandbox for analysis, calculation, and visualization inside a chat.

    Prepare and executeSandboxed code interpreter
    04

    Keep consequential writes gated.

    Where an action changes a system, it is prepared with its target and payload visible and held for approval where policy requires.

    Prepare and executeApproval-gated execution with visible payloads
    ACTIONS CONTROL BOUNDARY

    The product handles the work. Your team keeps the decision.

    Linkence handles

    • Expose the built-in tools an admin enables
    • Call custom actions through OpenAPI or MCP
    • Prepare typed payloads from the request context
    • Record the state of each attempted step

    Your team controls

    • Which tools and custom actions are available
    • Connected systems and granted API scopes
    • Approval policy and the authorized reviewer
    • How a failed or partial run is resolved

    Not designed for

    • Unattended screen automation
    • Unsupported or unreviewed system writes
    • Hiding partial or failed outcomes
    AVAILABILITY AND LIMITS

    What is available, and where the boundary sits.

    available

    Required setup

    • Built-in tools enabled by an administrator
    • Any custom actions defined over OpenAPI or MCP
    • An approval policy for consequential writes

    Known limitations

    • Available reads and writes differ by connector and custom action.
    • A multi-step run may stop or complete partially when a downstream system rejects a step.
    • Post-write verification is reported only when the system returns a verifiable result.
    Consequential supported writes can be held for explicit approval.Verified product claimCustom actions can be defined over OpenAPI or MCP to call connected system APIs.Verified product claim
    CONCRETE USES

    Three ways teams use Actions.

    Customer escalation follow-up

    A cited escalation summary

    GmailJiraConfluence

    A reply draft and ticket update held for review

    An authorized reviewer approves or rejects

    Service status check

    A question that needs a live system answer

    Custom actionJiraSlack

    A real-time status pulled through a custom action

    The on-call owner reviews the response

    Analysis with code execution

    A data question over retrieved results

    Code executionDriveConfluence

    A computed table or chart inside the chat

    A person validates the underlying data
    CONTINUE THE WORK
    Understand

    AI Chat

    Ask across company knowledge. Verify every answer.

    Explore product     Delegate

    Agents

    Package repeat work into an agent your team trusts.

    Explore product
    JiraavailableSalesforceavailableGmailavailableSlackavailable
    ACTIONS QUESTIONS

    Direct answers about Actions.

    Availability and connector-specific behavior are stated directly, without universal claims.

    What is the difference between a tool and a custom action?+

    Built-in tools (internal search, web search, code execution, image generation) come with the platform and are enabled by an administrator. Custom actions are defined over OpenAPI or MCP so an agent can call your own systems' APIs.

    Will an action write to a system without review?+

    Consequential writes stay approval-gated. The target and payload are shown before execution, and the run follows the approval policy configured for the workspace.

    What happens when one step fails?+

    The failed step is shown separately from completed work. The run may stop or report a partial outcome, depending on the dependency and the system's response.

    SEE ACTIONS ON YOUR DATA

    Start with one real question or responsibility.

    Connect the relevant sources and show Linkence the approval rules. See the path from evidence to completed work on your own data.

    Book a Working Session View all products
    Representative data is enough for the first demo. No production access required.