Skills are Agent Skills: a directory
holding a SKILL.md of instructions for one kind of task, with optional
scripts/, references/ and assets/. Mob reads the same format as other
agents, so a skill written for them works here. The code is in
apps/mob/src/skills.
Where skills come from
Section titled “Where skills come from”Mob looks in two directories, in order:
data_dir/skills/, Mob’s own;~/.agents/skills/, shared with other agents that follow the convention.
Each skill is a directory directly inside one of them; a link to a directory counts, as installers often link skills in. They are read again on every turn, so adding or removing a skill needs no restart. At most 100 skills are offered from each directory.
SKILL.mdstarts with YAML frontmatter.name(up to 64 lowercase letters, digits and single hyphens) anddescriptionare required;compatibilityis shown when the skill is loaded. Other fields,allowed-toolsincluded, are ignored and grant nothing.- Frontmatter that is not valid YAML only because a value holds an unquoted colon is read anyway. A name that differs from its directory is a warning.
- A skill without a description or a valid name is skipped. A name already found in an earlier directory is skipped too, so Mob’s own skills win.
/skillslists the skills found and every warning.
How the model uses them
Section titled “How the model uses them”- While any skill is found, the owner’s turns are offered the
skilltool, and the system prompt lists each skill’s name and description. Only found names are accepted. - With more than 20 skills the prompt only says how many there are, the
skilltool no longer lists every name, and the model finds the ones a task needs withtool_search, which returns each match’s description and theskillcall that loads it. skillwith a name returns the instructions without frontmatter, wrapped in<skill_content name="...">, and lists the skill’s other files (up to 100, hidden entries left out). With apathit returns one of those files as text, up to 256 KiB. Paths and links may not lead out of the skill.- Loading a skill runs at once, without approval: skills are installed by the owner, as MCP servers are configured by the owner.
- With the sandbox on, loading a skill also copies it to
/workspace/.mob/skills/<name>in the conversation’s workspace, replacing an earlier copy, and the model runs its scripts there with thesandboxtool. Links and hidden entries are not copied, and a skill over 512 files or 16 MiB is not copied at all. Reads and copies enforce byte limits while reading, including when files grow; copying cannot follow a path outside the source skill or destination workspace. With the sandbox off the model is told the scripts cannot run. Scripts never run on the host. - A loaded skill is a tool output like any other: when it is cleared to fit the context window the model is told to load it again.
Web management and installation
Section titled “Web management and installation”Open Plugins & cloud services → Skills to search installed skills, read their
instructions, and see scan warnings. Discover offers a small selection and
public GitHub repositories: enter owner/repo, optionally a branch or tag, then
find and select a SKILL.md or enter its path. Preview shows the instructions,
files and resolved commit before installation. A skill’s format being compatible
does not mean tools specific to another agent are available in Mob.
Installation downloads that exact commit into data_dir/skills/, never runs
installation scripts, and refuses links, unsafe paths, incomplete repository
trees, more than 512 files or more than 16 MiB. Files are staged before becoming
visible; cancelled or failed downloads are discarded. Remote operations have
a 45-second overall deadline. An existing skill, including one in the shared directory, is never
overwritten. Removal is available only for skills Mob installed in its own
directory; it deletes those files, including local edits. Mob-installed skills
remain listed and removable under their installation name even if their
SKILL.md is missing, invalid, or renamed in frontmatter. Shared or symlinked
skills remain managed by their original installer.
These operations run on the connected Mob service, including when the browser is on another machine. They are unavailable in the browser-only runner.
Lock files and update checks
Section titled “Lock files and update checks”Mob writes data_dir/.skill-lock.json using the npx skills global lock format
v3: source, source type, source URL, optional ref, path to SKILL.md, GitHub folder
tree SHA, and installation/update timestamps. This is a separate lock for Mob’s
own files; it does not change where npx skills installs or discovers skills.
Unknown fields in a compatible Mob lock are preserved when entries change.
For shared skills Mob reads the CLI’s global lock at
$XDG_STATE_HOME/skills/.skill-lock.json when XDG_STATE_HOME is set, otherwise
~/.agents/.skill-lock.json. Shared locks are never written. A shared entry is
not applied to a same-named skill in Mob’s directory.
Check for updates compares the recorded folder tree SHA with GitHub at the recorded ref, or the default branch when no ref was recorded. Changes to scripts or references count, while unrelated changes elsewhere in the repository do not. Requests for the same repository and ref are shared within a check. Results distinguish current, changes available, unknown source, and request failure. Checking does not modify files, refresh the baseline hash, or install updates; it compares upstream with the recorded installation, not with local edits.
Only public GitHub sources with usable global v3 lock entries can be checked.
The project skills-lock.json v1 format uses a different, content-based hash;
Mob does not load project-scoped skills or compare that hash to a Git tree SHA.
Invalid and unsupported locks are reported rather than overwritten. GitHub
requests currently use its unauthenticated API and can hit its rate limit;
failures are shown rather than treated as an up-to-date result.
Not yet
Section titled “Not yet”The mob skills add CLI, applying updates in place, private repositories,
project-scoped skills, the agent installing skills itself, and skills served by
MCP servers.