comparison

browser-memory vs browse.sh

Both give AI browsing agents a shared catalog of pre-built skills, so agents stop re-learning every site. The real difference is what a skill actually is: an executable tool that returns data, or a markdown playbook the agent runs itself.

browser-memoryUpdated Jul 2026
browser-memorybrowse.sh
What a skill isA callable tool: DOM selectors + XHRA SKILL.md playbook with steps and gotchas
Per-run workCall the tool, get data backAgent reads the playbook and executes it
Output to the modelStructured fields, ready to useWhatever the agent parses while following steps
When a site changesRepaired once, pushed to every agentPlaybook updated in the catalog
InterfaceMCP tools, for any MCP agentCLI plus an llms.txt index
Where it runsYour existing browser sessionLocal, or a Browserbase cloud session
Installbmem add {site}/{task}browse, then load a skill

Tool vs playbook

browse.sh ships a skill as a SKILL.md: a written playbook with the steps, selectors and gotchas for a task. The agent reads it and executes the workflow with its own reasoning, handling edge cases as documented. browser-memory ships the operation as a callable tool. The DOM selectors and XHR requests are already wired into code that returns structured fields, so the agent calls it and gets data, with no playbook to work through each run.

That is the whole difference in practice: a playbook is instructions the agent still has to run, a tool is a result. Fewer steps to reason through means fewer tokens and less room to drift.

A self-repairing catalog

When a site changes its markup, a browser-memory tool is repaired once and the fix ships to every agent on the shared catalog, as a tested tool. Missing tools are generated and tested right in your browser, then added for everyone. The catalog gets better the more agents use it.

When to use which

Choose browse.sh if you want a broad catalog of markdown playbooks and a CLI that drives local or cloud browsers. Choose browser-memory if you want each skill as an executable, self-repairing tool that returns structured data over MCP, running on your own browser session, so your agent spends its tokens on the task instead of on following steps.

FAQ

What is the difference between browser-memory and browse.sh?

Both are shared catalogs of pre-built browser skills that an agent installs by CLI. The difference is the execution model: a browser-memory skill is an executable tool built from DOM selectors and XHR requests that returns structured data, while a browse.sh skill is a SKILL.md playbook of steps and gotchas that the agent reads and re-executes with its own reasoning on every run.

Which one uses fewer tokens?

browser-memory tends to use fewer tokens per run because the agent calls a tool and receives fields directly, instead of loading a markdown playbook and reasoning through each documented step. The heavier the workflow, the larger the gap.

How do broken skills get fixed?

browser-memory keeps a self-repairing shared catalog: the first agent to hit a broken tool triggers a repair that ships to everyone. browse.sh maintains its catalog of playbooks centrally. Both aim to keep skills current; browser-memory does it as executable, tested tools.

Does browser-memory work with MCP agents like Claude Code?

Yes. browser-memory exposes skills over the Model Context Protocol, so Claude Code and other MCP-capable browsing agents can discover and run them directly, using your own browser session.

Compare more: browser-use, Playwright MCP, Browserless, Playwright, Stagehand, Browserbase, Unbrowse.

Skills as tools, not playbooks

A shared, self-repairing memory of the web that returns data, so your agent stops re-running steps on every visit.