Agent tools

DAIV's agent works through tools: it reads and edits files, runs commands, searches the web and reports back on your git platform. This page lists each tool, when the agent has it, and which settings turn it on or off. The agent is built on Deep Agents and runs on LangGraph.

Tools bound up front

DAIV binds these tools to the model at the start of every run.

Tool	What it does	When the agent has it
ls, read_file, glob, grep	List directories, read files, find files by pattern and search file contents by regular expression	Always
write_file, edit_file	Create a file, or change an existing one	Always
bash	Run a command in the persistent, isolated sandbox, after the command policy checks it	When the run has a sandbox. How an environment is resolved says when it has one.
write_todos	Track the steps of a task	Always
skill	Run a skill	Always
task	Hand work to a subagent	Always
ask_user_question	Ask you 1 to 4 questions and end the turn, and your answer arrives as the next message	On the main agent only, and not on a run nobody can answer: a scheduled run, a pipeline watch run or a delegated job. Answer a question shows it from your side.
delegate_jobs	Start independent jobs on other repositories or on the same one, and resume when they finish	When orchestration is on, which is the default, and the run is not a delegated leg

Tools loaded on demand

The agent loads these tools through tool_search, a tool that finds a tool by what it does and loads its definition. This keeps the list of tools the model sees short. A tool the agent loads stays available for the rest of the session. Setting DEFERRED_TOOLS_ENABLED to false binds every tool up front instead; Deferred tools describes it and its related settings.

Tool	What it does	When the agent has it
gitlab	Inspect and update issues, merge requests, pipelines, job logs and releases	On a GitLab repository
gh	The same, through the gh command line	On a GitHub repository
web_search	Search the web through DuckDuckGo, Tavily or SearXNG	When web search is on, which is the default. Web search describes the settings.
web_fetch	Fetch a URL, convert it to Markdown and answer a prompt about its content	When web fetch is on, which is the default. Web fetch describes the settings.
publish_artifact	Copy a workspace file, such as an HTML report, a CSV or a chart, into DAIV as an artifact and return its viewer URL. With artifact_id, it replaces the file of an artifact the session already published.	Always
fetch_artifact	Copy an artifact the session already published back into /workspace/tmp, so the agent can revise it	Always
check_delegated_jobs, cancel_delegated_jobs	Check, or cancel, the jobs the agent delegated	When delegate_jobs is available
MCP tools	The tools of the MCP servers the run has, such as Sentry or Context7	When an MCP server is enabled for the run

Model priority

The agent runs on the first model that a source sets, from a model you pick for the run, through the daiv-max issue label and the repository's .daiv.yml, to the site's settings. Model overrides lists the order. A subagent follows its own order, described in Subagent model precedence, and Environment variables lists the settings.