lufigir All external integrations go through Executor (mcp__executor__execute), never a direct MCP server, the public API, or a CLI shortcut. Call skills({ name: "execute" }) inside Executor first if you're unsure how to write the sandboxed code.
Discover, don't memorize. Connections and tool names change over time. Use tools.search({ query: "...", namespace: "<integration>_mcp" }) (or the bare integration name, e.g. "vercel") to find the exact tool path and argument shape. Any MCP added to Executor shows up here automatically — this skill does not keep an inventory of them.
Never construct a tool path or an argument list from memory. tools.search returns the path; tools.describe.tool({ path }) returns the exact inputTypeScript. Calling a guessed path costs a tool_not_found round trip, and a guessed argument name costs a validation error — both of which the two discovery calls would have avoided.
Absence from ToolSearch is not absence of the integration. These MCPs are not exposed as mcp__<name>__* tools; they live inside Executor. Searching the harness tool list for mcp__supabase__* and finding nothing proves only that there is no direct server. Before telling the user an integration is unavailable, run tools.search inside Executor, or list what is actually connected with tools.executor.coreTools.connections.list({}).
Accounts
Every tool path carries a .user.<account> segment, whether the integration has one connected account or several: context7_mcp.user.context7.resolve_library_id, notion_mcp.user.felipegiraldo, supabase_mcp.user.centrodeprototipado. There is no un-suffixed form — context7_mcp.resolve_library_id is a tool_not_found, not a shortcut. Take the whole path from tools.search; never assemble it from the integration name.
Which integrations are split across several accounts changes over time, so check with tools.search rather than assuming a given integration has only one.
felipegiraldois the default;centrodeprototipadois only for that project.- Use the account the user tells you to use — it's the source of truth.
- If the user didn't specify and it isn't obvious, ask before writing (reading from the wrong account is harmless; creating/editing, migrations, deploys, or branches are not).
Context7 (context7_mcp)
Two-step flow, worth remembering because it isn't obvious from the tool list:
tools.context7_mcp.user.context7.resolve_library_id({ libraryName, query })— both required; omitting either one fails validation.queryis the user's full question, and improves relevance ranking.tools.context7_mcp.user.context7.query_docs({ libraryId, query })— use the library ID from step 1 (prefer exact name match, higher benchmark score, and version-specific IDs when the user names a version).
Use it when the user asks about libraries, frameworks, or API references, or needs current code examples instead of relying on training data.
Firecrawl (firecrawl_mcp)
The general research and web-access surface. Not one tool with one use: nine that answer different questions, and picking wrong is what makes it feel redundant with things you already have.
Against Context7
| Indexes | Answers | |
|---|---|---|
| Context7 | Official library docs, resolved by version | "How do I use this API in the version installed here?" |
firecrawl_developer_search |
GitHub issues, merged pull requests, READMEs, curated docs | "Does this actually work, or did someone report that it doesn't?" |
Context7 to build, Firecrawl to find out whether you can build on top of it.
The worked example: oxlint's rules reference lists eslint/no-restricted-syntax. Context7 confirms it, because the docs say so. The binary does not have it. And the negation semantics the whole boundary table depends on were broken for a year, recorded in an issue and in no documentation anywhere. Docs describe intent; issues record reality.
Reading the web
Three tools, and the difference is how much you already know about where the answer is.
| You know | Tool | Notes |
|---|---|---|
| The exact page | firecrawl_scrape |
One URL to markdown, or straight to fields against a JSON schema. Prefer it over a plain fetch on heavy or JS-rendered pages |
| The site but not the page | firecrawl_map |
URL inventory without downloading bodies. search narrows it. Use it to find the page, then scrape that one |
| Neither | firecrawl_search |
Web, news or images. Supports site:, -term, inurl:, quoted phrases. Pass categories: ["developer"] to get the developer index alongside the web results in one call |
firecrawl_crawl is the fourth, and the one to reach for last. It walks a whole site and the result is enormous; bound it with limit, includePaths and maxDepth every time. Reasonable use: pulling a full documentation set once. Unreasonable use: anything map plus two scrape calls would have answered.
Documents
firecrawl_parse takes a file, not a site: PDF, Word, HTML and more, out as markdown, a summary, targeted answers to a question, or JSON against a schema. This is the one to use for a spec, a contract, a bank statement, a paper somebody sent — anything that would otherwise mean reading a PDF by hand. The schema output is what makes it worth calling instead of skimming: ask for the five fields you need and get them typed.
Research that can wait
firecrawl_agent starts an asynchronous job from a prompt, optional seed URLs and an optional JSON schema. It searches, navigates, reads and assembles a structured result across many sources. It returns only a job ID; poll firecrawl_agent_status until completed or failed, which commonly takes several minutes.
Use it when the question is genuinely a synthesis ("compare how these four libraries handle X", "what's the state of Y in 2026") and the answer can wait. Do not use it for anything a developer_search answers in one call, and never start one whose result the task cannot afford to wait for.
The live browser
firecrawl_interact opens a remote browser session: navigate, click, fill, or run Bash/Python/Node against the page. Two things to keep straight:
- It is not the user's browser. The Chrome DevTools MCP drives Chrome on this machine, which is what you want for debugging the local app, inspecting the console, or anything needing the user's logged-in session. Firecrawl's is a clean remote browser, which is what you want for a site that blocks scraping or needs a form filled to reveal content.
- It acts on the live site. Submitting a form is a real, external side effect. Confirm before anything that writes.
Close it with firecrawl_interact_stop when done.
Monitors
firecrawl_monitor_create schedules a recurring scrape, crawl or search and diffs each check against the previous one, returning a unified text diff for markdown targets or changed field paths for JSON. It can judge whether a change is meaningful against a plain-language goal, and notify by email or webhook.
The simple form takes page/pages or queries plus a required goal. Retention, schedule and change-tracking format come from the advanced body form.
This is the surface with no equivalent anywhere else in the toolkit, and the obvious use here is the VERIFY: blocks in project-architecture: pointed at the docs a reference depends on, they stop being "remember to re-read this" and become "you get told when it changes". monitor_run forces a check now; monitor_check and monitor_checks read results; monitor_update and monitor_delete manage them. A monitor schedules real future network work, so create one deliberately, not to satisfy a passing curiosity.
Papers
firecrawl_research_search_papers searches abstracts and metadata across biomedical, life-science, clinical and arXiv literature. research_read_paper pulls passages from one paper relevant to a question, research_inspect_paper returns canonical metadata for a DOI, PMID, PMC or arXiv id, and research_related_papers walks the citation graph from up to ten seeds.
Distinct from firecrawl_search with categories: ["research"], which only restricts web results to research-affiliated sites. These read the literature itself.
Feedback
firecrawl_feedback rates a completed search, scrape, parse or map job. Send the endpoint, job id, rating and issue codes; never the page contents. Worth doing when a result was badly wrong, since it is what improves the index.