Abilities
In many environments, an MCP server exposes dozens of tools. Passing all of them into every request increases token usage, degrades reasoning quality, and makes tool selection less reliable. Abilities address these problems by using a RAG-based semantic retrieval approach to activate only the small, domain-relevant subset of tools needed for each task, keeping the model focused. Abilities answer a simple question: Which tools should the agent activate for the current task?
In Agentforce Vibes, each Ability represents a domain-specific context, such as Apex, Lightning web components (LWC), DevOps, or Mobile, and automatically activates the right tools. Abilities are enabled by default. You can toggle it on or off using the Abilities setting in VS Code.
To enable or disable the Abilities feature from Settings:
- Select File > Preferences > Settings (Windows or Linux) or Code > Settings > Settings (macOS).
- Under Agentforce Vibes, select Enable Abilities.
Abilities are enabled by default.
When you toggle this setting, the DX MCP server configuration is updated automatically, unless you've customized your configuration.
When you enable Abilities for the first time, the DX MCP server configuration file is updated or overwritten to use the --toolsets "all configuration.
After the first time, toggling the Abilities setting in the UI updates the configuration file each time, switching between the --toolsets "all and the current default configuration.
Exception: If you've configured the --toolsets or --tools flags in your DX MCP server configuration file (after the first-time upgrade), the Abilities setting respects your configuration and doesn't overwrite the file.
To return to either the --toolsets all configuration or the defaults on your own, disable and then delete the DX MCP server. The configuration is then recreated with either --toolsets all or default configuration, depending on the current value of the Abilities setting.
Abilities are designed to manage context intelligently by controlling tool activation. This table highlights the core characteristics that make Abilities efficient, composable, and optimized for quality execution.
| Feature | Benefit / How It Works |
|---|---|
| Context-Aware Activation | Only the tools relevant to your request are enabled automatically. |
| Automatic Execution | Activates on every request without manual setup or configuration. |
| Composable Modules | Multiple Abilities can be active at the same time (for example, Apex + LWC + DevOps). |
| Token-Efficient | Minimizes token usage and cost by limiting context to what’s necessary. |
| Quality Optimized | Models perform better when operating within a focused and coherent execution context. |
Each time you start a new task, the Ability Coordinator interprets your intent and automatically activates the relevant Ability (or combination of Abilities) without any manual setup or configuration. It determines which Abilities to activate by using signals from the prompt intent and keywords in the prompt.
Depending on the type of request, the Ability Coordinator activates the appropriate domain-specific Ability, for example:
Build a component – If you ask the agent to create a Lightning web component with tests, the LWC Ability is activated.
Update metadata – Requests such as adding a new value to a Lead picklist trigger the Platform Ability.
Write triggers – Requests to create an Apex trigger and handler class activate the Apex Ability.
Abilities analyzes your request, selects the minimal set of tools required, and ensures that the agent operates in a focused, relevant execution context. Multiple abilities can be combined to form the final execution context for a single request.
To decide which tools to activate, the Abilities Coordinator:
-
Converts each MCP tool’s definition—including its name, description, and intended usage—into vector representations stored in a database.
-
Converts the user’s prompt into a vector using the same embedding model.
-
Compares the prompt vector with the tool embeddings to identify the most relevant tools.
-
Picks the top K tools (around 10) based on similarity scores.
-
Automatically includes any dependent tools that are required to ensure complete execution.
Some MCP tools function as orchestrators, which means they require more tools to complete a task successfully. To support this, Abilities apply a post-selection expansion step after the initial tool match. This step ensures that the final execution context includes all required dependencies—without exposing unnecessary tools.
Prompt : “Deploy the Apex changes and run all relevant tests.”
Activated Ability: Apex Ability.
Available Tools (DX MCP Server) The DX MCP server exposes a broad set of tools across multiple domains, including Apex, LWC, Mobile, DevOps, and Platform:
Tools Activated for This Request Based on detected intent and the active Ability, only the tools required for Apex deployment are included in the execution context:
[deploy_metadata, run_code_analyzer, apex_test_run]
With Abilities enabled, the model reasons only over the tools relevant to Apex deployment, while tools from unrelated domains such as LWC, Mobile, DevOps, and Platform are excluded. This reduces token usage, improves tool selection accuracy, and ensures that execution remains intentional, predictable, and high quality.
At the same time, all Salesforce DX MCP server tool definitions are written to a file in the .sfdxdirectory, and the model is explicitly instructed that it can discover other tools there. This lets Agentforce to look up and learn how to use other DX MCP tools when they’re referenced during a task even if they aren’t part of the active tool set, without expanding the initial context unnecessarily.
Each Agentforce Vibes Ability bundles a curated set of MCP tools for a specific development domain. Here's an example of possible active tools for each of the Abilities, depending on the prompt:
| Ability | Includes MCP Tools |
|---|---|
| Apex Ability | run_code_analyzer, deploy_metadata, create_scratch_org |
| LWC Ability | create_lwc_component_from_prd, guide_figma_to_lwc_conversion, create_lwc_jest_tests |
| Mobile Ability | create_mobile_lwc_biometrics, create_mobile_lwc_document_scanner |
| DevOps Ability | create_devops_center_pull_request, checkout_devops_center_work_item, list_devops_center_work_items |
| Platform Ability | get_username, assign_permission_set, list_all_orgs |
Abilities work together with Skills to enable governed, scalable, and trustworthy agentic development.
| Feature | Primary Role | What It Addresses |
|---|---|---|
| Abilities | Context activation | "What should the agent use?" |
| Skills | Procedural guidance | "How should the agent do the work?" |
- You submit a request.
- Abilities automatically activates the relevant tools based on intent.
- The agent plans its approach within this constrained environment.
- Agentforce invokes as needed to guide execution.
- The agent executes using the selected tools.