No Export, No Exit: How to Check for Lock-In Before You Build a Library Anywhere

The five-minute checklist

We read Recall's developer documentation, Rescroll's homepage and Heptabase's MCP support article on 4 September 2026 and put all three through the same five questions. The questions matter more than the answers: you can run them yourself, on any tool, before you commit five years of saved material to it.

Each check is answerable from a vendor's own public pages, or from one test item in the app.

1. The export test. Can you get everything out, today, on the tier you are actually on? Not "email support", not "available on the paid plan". You want a button in the app or a documented endpoint. If you cannot find one in a minute of looking, that is itself a finding.

2. The format test. Open whatever the export produces. Markdown, plain text, CSV and JSON still open in software nobody has written yet. A bundle that only reimports into the same product is not a copy of your library; it is a copy that one company can read.

3. The API test, which is really two tests. Nearly everything in this category now advertises an API or an MCP server. Ask separately whether it reads and whether it writes. Read-only means anything can query your library but only the vendor's own app can feed it — fine if you always save by hand, awkward the moment you want an assistant to file things for you.

4. The source test. Save one web page, then look at what your copy actually contains. If the tool kept a link and a summary, your archive depends on the original staying online. If it kept the text, it does not. Test this with one item instead of reading about it.

5. The third-party reader test. Does anything other than the vendor's own app read the library — an MCP client, a sync target, a folder on your own disk? One reader is a dependency. Two independent readers is a library.

Running the checklist on Recall

Recall answers the read half of check three plainly, because it publishes the answer rather than implying one. Its MCP documentation, read live on 4 September 2026, says: "The MCP server currently provides read-only access to your knowledge base. Write capabilities are coming soon." The same page documents setup for Cursor, Claude, ChatGPT, Perplexity and Gemini, needs no API key because authentication happens in the browser, and requests a single kb:read scope, described on the page as "Read your saved content" — exactly the shape a read-only server ought to have.

The write half is not there yet, on their own account. Their separate REST API reference at https://docs.recall.it/developer/api is read-only as well — three GET endpoints for listing cards, reading one card and searching — and says in its own words that "a write API is on the roadmap". Their public changelog carries no entry announcing a shipped write API. The 23 July 2026 entry describes a rebuild of social saves as the "first step in the development of the write API", which is a company telling you where it is in the work rather than where it would like you to think it is. We went through the practical consequences in can Claude save to Recall.

Check five, Recall passes: its pricing page lists API and MCP access on the Free tier, not behind Plus or Max. Checks one and two are answered in the changelog rather than the docs — the 19 August 2026 entry notes that exporting a PDF card to Markdown now includes the original PDF in the zip, and the 4 June 2026 entry added multi-select for export. Markdown plus the original file is a good answer to the format question.

Running it on Rescroll, where the evidence runs out

Rescroll — the brand served on the thesecondbrain.io domain, with the page title "Chat with Social Media Content for Creators | Rescroll" and the H1 "Chat with social media content." when we read it on 4 September 2026 — is the case where the checklist returns nothing rather than a bad answer. The string "MCP" appears zero times on that page. There is no export offering on it either. Its plans render as Basic at 500 credits a month, Pro at 2,500 and Ultimate at 10,000, and the homepage is precise about what the product does do: "Paste any YouTube channel, pull 50 videos in one click."

Be careful with what that permits you to conclude. An absence on a marketing homepage is an absence of evidence, not a statement by the company. A homepage is a sales document — chosen, not exhaustive. There may well be an export in the app and an API documented somewhere we did not read; we found neither on the page we fetched, and we are not going to convert that into a claim that they do not exist. The useful conclusion is narrower: if a check cannot be answered from the public pages in five minutes, ask the vendor before you commit, and treat the answer as unknown until they reply.

Heptabase, and what a first-party MCP proves

Heptabase runs a first-party MCP server, documented in a support article that returned 200 when we fetched it on 4 September 2026. First-party is the part that matters for check five: a community bridge is one maintainer's spare time, while a vendor-run server is the vendor accepting that software outside its own app will read the library, and carrying the maintenance for it.

One practical note if you test it yourself: the MCP endpoint is not a page you can open in a browser. A plain GET returns 405, which is ordinary behaviour for an MCP endpoint and tells you nothing about whether the server works. Read the support article and connect a client instead of judging by poking the URL.

What a first-party MCP does not settle is checks one, two and four: it says nothing about whether you can pull a full archive out, in what format, or whether your saved copy survives the original going offline. We did not run those checks against Heptabase, so we are not going to describe its export options here.

Where BrainTube stands

BrainTube runs one read-write MCP server covering the whole library, at https://mcp.brain-tube.com/mcp over streamable-http. Reading the connect page on 4 September 2026 gives 44 tools, among them ingest_content, bulk_ingest, search_knowledge, export_corpus and export_claude_md. Claude.ai connects by OAuth with no API key; a bt_ token in an X-BrainTube-Token or Authorization: Bearer header also works. MCP access and API keys are included on every plan, including Free, so check three and check five can be run on a $0 account.

On check one and check two, the published description of export_corpus is: "Export your entire BrainTube knowledge corpus as a ZIP of Markdown files. Returns a signed download URL." In the app the export formats are JSON, Obsidian-flavoured Markdown, Notion-flavoured Markdown, txt, Anki TSV and ZIP. Export is auth-only, with no tier gate — the same on Free as on Pro+.

Now the half that goes the other way, because a checklist you only pass is not a checklist. The in-app JSON export is capped to the 500 most-recent items, so JSON is a sample, not your archive; the ZIP is the full-archive route. There is no Zotero export and no outbound Notion export, and neither is on the way as anything we can promise you a date for. On the import side, ZIP import exists but has not been tested against a real Recall export, so we will not call it a migration path until we have run one end to end and can say what broke. And on check four we would rather you save one item and look at what BrainTube kept than take our description of it.

How these figures were checked

Recall's two developer pages were read on 5 September 2026: the MCP documentation at docs.recall.it/developer/mcp for the read-only quote and the scope, and the API reference at docs.recall.it/developer/api for the three read-only endpoints. The remaining Recall facts were read on 4 September 2026 — the changelog at feedback.recall.it/changelog for the 19 August, 23 July and 4 June 2026 entries, and www.recall.it/pricing for the Free tier's API and MCP access. Rescroll facts — title, H1, plan credit allowances, the channel-import line, and the absence of any MCP mention or export offering — were read from www.thesecondbrain.io on the same date. Heptabase's first-party MCP server was verified on 17 August 2026 against the support article, which we re-fetched live on 4 September 2026; its endpoint returned 405 to a browser GET on that date, as expected.

Claims about BrainTube describe functionality running in production on 4 September 2026, with the tool count and export descriptions read from brain-tube.com/connect.md and the plan details from brain-tube.com/pricing on that date. Nothing here rests on a source we could not re-fetch on 4 September 2026. If you spot something here that has gone stale, tell us and we will correct it with the date attached.