Tools let your agent browse, code and run
What your agent can do beyond talking. Memory searches its files and conversations, Scheduler starts subagents and sets schedules, Web reads and searches the web, Browser works pages and Desktop a desktop as a person does, Run and File run programs and change files, each call in a sandbox of its own, and Code asks the project's language servers.
Link's tools, in every harness
Link has eight tools, and a harness reaches each of them the same way.
| Tool | What it does |
|---|---|
| Memory | Its files and conversations, searched by meaning |
| Scheduler | Subagents, schedules and messages from others |
| Web | A page as markdown, and a search with your key |
| Browser | A real browser, sandboxed, read and clicked |
| Desktop | A virtual desktop, its mouse, keys and screenshots |
| Run | Programs, each call sandboxed, long ones behind |
| File | Its files read, written and changed |
| Code | What language servers find wrong, and names |
Each tool runs as an MCP server outside your agent's sandbox, and your agent calls it through its proxy. Each tool's page says what it does in full: words and meaning in Memory's search, the browser's own sandbox, the desktop's mouse and keys, Run's background programs.
Give your agent one
lnk agent tool list shows which your agent has, and lnk agent tool add memory, add web, add browser or add desktop gives it one;
its first lnk agent start offers them.
Which harness has which
Memory, Web, Browser and Desktop work in any harness, Scheduler in any harness that reads who a message is from, and Run, File and Code in any harness that keeps to the files folder's layout. Today those are Link Harness, since the other harnesses bring files, a shell and code of their own.
On in Link Harness
Scheduler, Run, File and Code come with Link Harness, each on unless
its agent's spec says [tool.<name>] enabled = false. A tool knows
nothing of a harness: what it reads of an agent, it reads in a
documented format.
A harness's own tools
A harness may bring tools of its own, such as Hermes Agent's scheduled jobs, and Link's tools sit beside them. Link adds its tools in the harness's own MCP settings and removes only its own. Each harness's page says what it keeps: Harnesses.
Namespaces
Every harness receives Link's actions by namespace. task combines run
and scheduler: run, start, read, agent, schedule, list and
stop. file combines file operations and code lookups. agent offers
send; web, browser, desktop and memory keep their namespace names.
Task kinds
task__list lists runs and schedules together; kind selects run or
schedule. When both providers are present, task__stop requires that
selector so an ambiguous identifier never stops the wrong kind of work.
Provider policies
Each call uses its provider's existing confinement and current approval policy. The namespace does not grant more access. Changing or removing a provider takes effect on the next call, and a changed launch requires a new namespace session.
Docs
- Decisions: Where Link's tools live, and how a team's tool runs once for the team