Agent Mode
Use natural language to accomplish any task in the terminal
Last updated
Was this helpful?
Use natural language to accomplish any task in the terminal
Last updated
Was this helpful?
Agent Mode is a mode in Warp that lets you perform any terminal task with natural language. Type the task into your terminal input, press ENTER, and Warp AI runs highly accurate commands tailored to your environment.
Agent Mode can:
Understand plain English (not just commands)
Execute commands and use that output to guide you
Correct itself when it encounters mistakes
Learn and integrate with any service that has public docs or --help
Utilize your saved workflows to answer queries
.
You may enter Agent Mode in a few ways:
Type any natural language, like a task or a question, in the terminal input. Warp will recognize natural language with a local auto-detection feature and prepare to send your query to Warp AI.
Use keyboard shortcuts to toggle into Agent Mode CMD-I or type ASTERISK-SPACE.
Click the “AI” sparkles icon in the menu bar, and this will open a new terminal pane that starts in Agent Mode.
From a block, you want to ask Warp AI about. You can click the sparkles icon in the toolbelt, or click on its block context menu item “Attach block(s) to AI query”.
This will put you in Pair mode by default. While pairing with Warp, you can write out questions and tasks in an ongoing conversation.
When you are in Agent Mode, a ✨ sparkles icon will display in line with your terminal input.
The feature Warp uses to detect natural language automatically is completely local. None of your input is sent to AI unless you press ENTER in Agent Mode.
If you find that certain shell commands are falsely detected as natural language, you can fix the model by adding those commands to a denylist in Settings > AI > Auto-detection denylist.
You may also turn autodetection off from Settings > AI > Input Auto-detection.
The first time you enter Agent Mode, you will be served a banner with the option to disable auto-detection for natural language on your command line:
Warp input occasionally shows hints within the input editor in a light grey text that helps users learn about features. It's enabled by default.
You can quit Agent Mode at any point with ESC or CTRL-C, or toggle out of Agent Mode with CMD-I.
Once you have typed your question or task in the input, press ENTER to execute your AI query. Agent Mode will send your request to Warp AI and begin streaming output in the form of an AI block.
Unlike a chat panel, Agent Mode can complete tasks for you by running commands directly in your session.
If Agent Mode finds a suitable command that will accomplish your task, it will describe the command in the AI block. It will also fill your terminal input with the suggested command so you can press ENTER to run the command.
When you run a command suggested by Agent Mode, that command will work like a standard command you've written in the terminal. No data will be sent back to the AI.
If the suggested command fails and you want to resolve the error, you may start a new AI query to address the problem.
If Agent Mode doesn't have enough context to assist with a task, it will ask permission to run a command and read the output of that command.
You must explicitly agree and press ENTER to run the requested command. When you hit enter, both the command input and the output will be sent to Warp AI.
If you do not wish to send the command or its output to AI, you can click Cancel or press CTRL-C to exit Agent Mode and return to the traditional command line.
Once a requested command is executed, you may click to expand the output and view command details.
In the case that a requested command fails, Warp AI will detect that. Agent Mode is self-correcting. It will request another command until it completes the task for you.
Warp lets you choose from a curated list of LLMs for use in Agent Mode. By default, Warp uses Claude 3.7 Sonnet, but you can switch to other supported models, including:
OpenAI (General Purpose): GPT-4o, GPT-4.1
OpenAI (Reasoning Models): o3-mini, o3, o4-mini
Anthropic: Claude 3.7 Sonnet, Claude 3.5 Sonnet, Claude 3.5 Haiku
Google: Gemini 2.0 Flash, Gemini 2.5 Pro
You can also select "auto" to let Warp automatically choose the best model for your task based on factors like query type and context.
When you start an agent mode conversation, you will be able to see the model being used.
To change the model being used, click the current model name, 'claude 3.5 sonnet' in the example image above, to open a dropdown menu with the supported models. Your model choice will persist in future prompts.
Conceptually, a conversation refers to a sequence of AI queries and blocks. Conversations are tied to panes and you can have multiple Agent Mode conversations running at the same time in different panes.
You will get more accurate results from AI queries if the conversation is relevant to the query you ask. When you start an AI query unrelated to the previous conversation, start a new conversation. When you start an AI query related to the previous conversation, ask a follow-up and stay in the same conversation.
Agent Mode can gather context from your terminal sessions and tailor every command to your session and environment.
You can supply a block of context to your conversation with Agent Mode as part of your query. From the block in the terminal, click the AI sparkles icon to "Attach as Agent Mode context."
The most common use case is to ask the AI to fix an error. You can attach the error in a query to Agent Mode and type "fix it."
If you're already in Agent Mode, use the following ways to attach or clear context from your query:
Attach a previous block
To attach blocks to a query, you can use CMD-UP to attach the previous block as context to the query. While holding CMD, you can then use your UP/DOWN keys to pick another block to attach.
You may also use your mouse to attach blocks in your session. Hold CMD as you click on other blocks to extend your block selection.
Clear a previous block
To clear blocks from a query, you can use CMD-DOWN until the blocks are removed from context.
You may also use your mouse to clear blocks in your session. Hold CMD as you click on an attached block to clear it.
By default, if you ask an AI query right after any interaction in Agent Mode, your query will be sent as a follow-up. The follow-up ↳ icon is a bent arrow, to indicate your query is continuing the conversation.
To enter follow-up mode manually, press CMD-Y.
If there is no follow-up ↳ icon next to your input, this indicates a new conversation. If you ask an AI query after running a shell command you will be placed in a new conversation. Agent Mode will also kick you out to a new conversation after 3 hours.
To start a new conversation manually, use CMD-Y or BACKSPACE.
Agent Mode now includes advanced coding capabilities directly within your terminal, triggered when it detects an opportunity to generate a code diff. This powerful feature allows for seamless code generation, editing, and management tasks, all within your terminal environment.
For a more tailored editing experience, you can attach context blocks directly from the terminal, providing Agent Mode with specific input to guide its diff suggestions.
Agent Mode responds to prompts related to code generation, editing, and analysis. Here are some examples:
Code creation: “Write a function in JavaScript to debounce an input”
Based on error outputs, suggest fixes: “Fix this TypeScript error.”
Modify code within a file: “Update all instances of ‘var’ to ‘let’ in this file.”
Apply changes across multiple files: “Add headers to all .py files in this directory”
When Agent Mode generates a code diff, you can review, refine, and decide whether to apply the changes.
When Agent Mode generates a code diff, it automatically triggers a built-in text editor diff view, which visually displays the changes as distinct hunks.
You can navigate through the highlighted hunks using the UP and DOWN arrow keys or mouse clicks. Agent Mode also supports multi-file changes, enabling you to view and manage hunks across several files. To switch between files, use the LEFT and RIGHT arrow keys.
Once satisfied with the changes, you can apply them by pressing ENTER or selecting the “Accept Changes” button. These modifications will not be applied to the files until you explicitly accept them.
For refining or customizing the changes, Agent Mode allows for further interaction. You can refine the query (and diff) using natural language by pressing R or the “Refine” button, which will generate an updated diff based on your follow-up input.
If you wish to make direct edits within the text editor, press E or the “Edit” button to open the editor view. You can exit the editor by pressing ESC.
To cancel a pending action, use CTRL-C (on both Mac and Linux systems).
Agent Mode’s built-in text editor supports a wide range of programming languages and syntax highlighting, including: Python, JavaScript, TypeScript, Rust, Golang, Java, C, C#, C++, HTML, CSS, Bash, JSON, YAML. We are also continuously working on adding support for more languages.
In order for Agent Mode to be able to answer questions about about the code files in your codebase, Warp can generate an outline containing relevant information about each file. This outline can then be used when querying Agent Mode to find relevant files for a requested query.
No code indexed with codebase context is ever stored on our servers.
Outline generation works in the following way:
When opening a new terminal pane or changing to a new directory in the terminal, check if this directory is a part of a Git repository.
If it is in a Git repository, extract metadata from each code file in the repository.
Only file types for languages supported by Agent Mode are processed.
Files in .gitignore are not processed.
This metadata currently includes function names in the file, but this may expand.
The first time that a user opens a directory after the Warp application is opened, the outline for the repository is fully generated. Creating an outline for repositories can take a few minutes for large repositories. Codebase context will not be used in Agent Mode queries until the outline is generated.
Outline generation is completely local. The outline is only sent to the server when an Agent Mode request is made that could use the codebase context and the user approves sending codebase context for that directory.
After the outline is generated in full, Warp watches for changes to any file in the directory and the outline is updated as needed.
These outlines can then be used by Agent Mode to search for relevant files to answer questions about a codebase, or find relevant files to edit.
Dispatch is a form of Agent Mode that carries out complex tasks automatically. When you make a Dispatch query, the AI will:
Use a reasoning model to create a plan to carry out the task. You can refine the plan with AI, or edit it directly.
Automatically carry out the approved plan.
Dispatch is currently in beta. Its capabilities will improve over time.
You can enter Dispatch in a few ways:
Press CMD-SHIFT-I to toggle between Dispatch and the terminal input, or to switch from pairing to Dispatch.
Agent Mode supports configurable autonomous command execution under Settings > AI > Autonomy. You can customize this by:
Using a command allowlist to specify which commands can auto-execute
Using a command denylist to specify which commands require confirmation
Letting the Agent Mode model automatically determine if a command is safe to execute based on whether it's read-only
Agent Mode comes with default allowlist entries for common read-only commands that can be automatically executed without user confirmation.
which .* - Find executable locations
ls(\s.*)? - List directory contents
grep(\s.*)? - Search file contents
find .* - Search for files
echo(\s.*)? - Print text output
You can add your own regular expressions to this list in Settings > AI > Autonomy > Command allowlist. Commands in the allowlist will always auto-execute, even if they are not read-only operations.
Agent Mode comes with default denylist entries for potentially risky commands that always require explicit user permission before execution. A couple of examples include:
wget(\s.*)? - Network downloads
curl(\s.*)? - Network requests
rm(\s.*)? - File deletion
eval(\s.*)? - Shell code execution
The denylist takes precedence over both the allowlist and model-based auto-execution. If a command matches the denylist, user permission will always be required, regardless of other settings. You can add your own regular expressions to this list in Settings > AI > Autonomy > Command denylist.
Agent Mode can dynamically analyze command safety for automatic execution of read-only commands. This provides intelligent command safety analysis but follows a strict precedence order.
Denylist (highest priority) - Commands always require confirmation
Allowlist - Commands always auto-execute
Model-based analysis (lowest priority) - Agent Mode determines if a command is read-only safe
This behavior can be toggled in Settings > AI > Autonomy > Model-based auto-execution.
When performing coding tasks, Agent Mode can automatically read files. This allows Agent Mode to analyze code without requiring explicit permission for each file access.
This behavior can be toggled in Settings > AI > Autonomy > Coding read permissions.
When a Warp Drive object is pulled as context, it will be displayed in the conversation as a citation under "References" or "Derived from".
Warp AI does not have up-to-date information on several commands’ completion specs
You cannot continue any conversation that’s before an existing conversation.
Agent Mode works better with Warp's default prompt settings, where the prompt starts on a new line, than it does with a same-line prompt. If you are using the same-line prompt, the cursor will jump from the end of the single line to the start of the input box when you switch to Agent Mode.
Lite is our basic AI model that serves two purposes:
Fallback model: If you hit your Turbo token or AI request limits, Warp automatically switches to Lite so you can keep using AI without interruption — at no additional cost.
Standalone option: You can also choose to use Lite before hitting your limits. In that case, usage will still count toward your monthly token and request limits — but once those limits are reached, Lite remains available with unlimited usage.
This error means your input (plus attached context) exceeds the maximum context window of the model you're using. For example, GPT-4o has a context window limit of 123,904 tokens. If you exceed that, you may receive no output.
To fix this, try:
Starting a new conversation
Reducing the number of blocks or lines attached to your query
Agent Mode has replaced the Warp AI chat panel. Agent Mode is more powerful in all of the chat panel's use cases. Not only can Agent Mode run commands for you, it can also gather context without you needing to copy and paste. To start a similar chat panel, click the AI button in the menu bar to start on a new AI pane.
No, Warp nor its providers OpenAI or Anthropic train on your data.
The following LLMs are currently supported in Warp:
OpenAI:
GPT-4o
GPT 4.1
o3-mini
o3
o4-mini
Claude:
3.7 Sonnet
3.5 Sonnet
3.5 Haiku
Google:
Gemini 2.0 Flash
Gemini 2.5 Pro
R1
V3
No, DeepSeek is never enabled by default. By default, Agent Mode uses Claude 3.7 Sonnet. To use DeepSeek, you would need to manually select it from the model selector inside Agent Mode.
Warp AI is tailored for the terminal so you can get optimal results and performance. It's suitable for AI power users and professional use cases.
For organizations with strict security requirements, a “Bring Your Own LLM” option is available on the Enterprise plan. At the Enterprise plan level, we can work closely with your team to ensure quality and compliance for your LLM of choice.
Toggle this feature Settings > AI > Show input hint text or search for "Input hint text" in the or Right-click on the input editor.
DeepSeek: R1, V3 (hosted by in the US)
When using "Pin to the top" , the direction for attaching or detaching is reversed (i.e. CMD-DOWN attaches blocks to context, while CMD-UP clears blocks from context).
When using "Pin to the top" , the direction for attaching or detaching is reversed (i.e. CTRL-DOWN attaches blocks to context, while CTRL-UP clears blocks from context).
When using "Pin to the top" , the direction for attaching or detaching is reversed (i.e. CTRL-DOWN attaches blocks to context, while CTRL-UP clears blocks from context).
If you have questions or feedback about this recent feature, feel free to contact us at .
Gather context about the task, using your , , and followup questions.
While Dispatch is executing a plan, it automatically runs commands and applies code changes. However, it will still obey your . The border along the left of the session will change color to indicate that Dispatch is running autonomously:
If you're using the , you can also click the Dispatch context chip:
Agent Mode can leverage your contents to tailor responses to your personal and team developer workflows and environments. This includes any Workflows, Notebooks, Environmental Variables, etc..
Every Warp plan includes a set number of Warp AI requests per user per month. Please refer to to compare plans.
AI Request limits apply to Agent Mode, , , and AI autofill in Warp Drive. When you have used up your allotted requests for the cycle, you will not be able to issue any more AI requests until the cycle renews.
Every time you submit an AI query from your input box, this counts as one Warp AI request. do not count as billable Warp AI requests. suggestions are counted separately to Warp AI requests as noted in our docs.
You can monitor your request usage under Settings > AI > Usage.
Tokens are chunks of text, such as words, parts of code, or characters, that large language models (LLMs) break down to analyze and generate responses. LLMs have a maximum number of tokens they can process at once. Warp AI Requests and Suggestions are not the same as Tokens, which are limited separately regardless of which plan you're on. Please learn more about Tokens and refer to for the current monthly token limits on each plan.
Allotted AI requests refill every 30 days from your signup date. When you upgrade to a , you will be given more requests immediately. You can follow along with your refill period by referencing Settings > AI > Request Usage.
For more FAQs about pricing, visit .
Note that Agent Mode blocks are not shareable during . Participants will be able to share regular shell commands that are run, but will not be able to share AI interactions (requested commands, AI blocks, etc.).
Block actions such as are not available on Agent Mode AI blocks.
Lite is more token-efficient than other premium models and supports core AI workflows. Learn more in in our pricing documentation for more details.
Once you exceed your monthly token limit or AI requests on the Turbo plan (see for current limits), premium models will be disabled, and Warp will automatically switch you to Lite. This allows you to continue using AI features with a more token-efficient model until your quota resets at the start of your next billing cycle.
If you have questions or need extended access, feel free to reach out to us at .
See our for more information on how we handle data with Agent Mode.
DeepSeek (hosted by in the US):
We take privacy and security very seriously when it comes to models developed by foreign companies or hosted outside the US. DeepSeek models in Warp are hosted exclusively on US servers through our trusted provider, . No requests are routed to servers outside the US.
