@sheetrender/mcp
MCP server for SheetRender: turn rows of a
spreadsheet into one PDF each, such as certificates, letters, donation receipts
and offer letters.
The quickest start needs no account and no install. Add the hosted server
https://mcp.sheetrender.com/mcp as a connector in ChatGPT, Claude or any
client that takes a Streamable HTTP URL, then paste or describe your rows. It
fills SheetRender's built-in templates and returns a preview and a PDF per row
(details).
With a SheetRender API key you get the full tool set: your own saved templates,
HTML templates, datasets and batch jobs, either through the npm package below or
the same hosted URL.
Setup
Get an API key from sheetrender.com under Settings →
API keys, then add this to your MCP client config:
SHEETRENDER_API_URL is also read, and defaults to https://sheetrender.com.
Set it only if you're pointing at a self-hosted or staging instance.
Hosted endpoint
The same server runs at https://mcp.sheetrender.com/mcp over Streamable
HTTP, so clients that can't spawn a local process can use it too. Nothing is
installed; each request carries your API key:
Requests without that header get the smaller no-key tool set described
below. The key goes straight through to the
SheetRender API for that one request and is never stored — the server keeps no
sessions, so every request stands alone.
Where the key goes depends on the client:
-
Claude Code:
claude mcp add --transport http sheetrender https://mcp.sheetrender.com/mcp --header "Authorization: Bearer sr_live_..."
-
Cursor, Windsurf, VS Code and other clients with an mcp.json:
-
Claude API (the Messages API's MCP connector): add
{"type": "url", "url": "https://mcp.sheetrender.com/mcp", "name": "sheetrender", "authorization_token": "sr_live_..."}
to mcp_servers.
-
claude.ai, Claude Desktop and ChatGPT custom connectors take a server URL
and an OAuth client, not a static header. Until the endpoint speaks OAuth,
use the stdio package above there — it's the same tools, with the key in
env — or bridge with npx mcp-remote https://mcp.sheetrender.com/mcp --header "Authorization: Bearer sr_live_..."
as the command.
Two differences from the stdio server follow from the process not running on
your machine. Rendered PDFs come back inline as a base64 resource (up to 8 MB;
larger ones are reported with their size and left for the web app) instead of
as a temp-file path, and upload_dataset is not offered because there is no
local file to read — send rows with create_dataset instead.
GET /healthz answers 200 without credentials. Request bodies are capped at
25 MB; anything larger is a 413.
A key that merely starts with sr_ proves nothing, so the server waits for
the SheetRender API to accept it. Until a call made with the key has
succeeded (normally the first tool call, e.g. list_templates), its requests
get the same bounds as requests without a key, described below: bodies up to
2 MB, at most 4 JSON-RPC messages per batch, and every message counted toward
the per-network budgets. Once the API has accepted the key, the server
remembers a hash of it (never the key) for an hour after its latest
successful call, and its requests get the 25 MB body cap and batches of up to
20 messages. MCP clients send one message per request, so the batch caps only
matter to hand-written clients. If the very first call is a body over 2 MB
(a large create_dataset), it gets a 413 that says to make a small call
first and retry. Every request, with a key or without, counts toward the
server-wide in-flight cap.
One tool call's arguments may hold at most 600,000 values (array elements and
object fields together) from an accepted key, enough for create_dataset's
500,000 cells, and 20,000 from a key not yet accepted. Anonymous calls are
capped at a little over the largest real call.
Bodies over 2 MB also take one of a few server-wide large-body slots, 1 by
default (LARGE_BODY_MAX_IN_FLIGHT). A slot is taken before reading when
Content-Length declares such a body, or as soon as a streamed body passes
2 MB, and is held until the response ends. When all are taken, the request
gets a 503 with Retry-After: 5 and the rest of its body is not read. One 25 MB
body can take around 100 MB of heap while it is decoded, parsed and passed
on, so raise this only with the heap (the hosted container runs a 192 MB
heap).
Without a key (ChatGPT and Claude directory listings)
When the server runs with SHEETRENDER_DEMO_API_KEY set, a request that
sends no Authorization header gets a different, smaller tool set
instead of a 401. These tools fill SheetRender's built-in templates
(certificate of completion, letter, donation receipt, job offer letter) from
rows in the chat, through a dedicated rendering account:
list_document_templates: the templates, their fields (key, label,
required, example, character and line limits).render_documents: up to 25 rows per call, one PDF per row; returns a PNG
preview and a PDF link per document (both expire after an hour), the rows
that missed a required field, and how many of the month's 50 documents per
user (in Claude, per conversation) remain. When a limit shared with other
users is what cut or refused the call, it says the limit is shared instead of
showing a count. A refused or cut call also says the rows can be taken
to the website with a continue link. A value over
its field's character or line limit is refused before the API is called, and
the API's own reason for refusing rows (row number, field and rule, never the
value) is passed on.create_continue_link: up to 100 rows, stored for 7 days, and a link to the
template's page on the website with those rows loaded. Called only when the
user wants to continue there.
render_documents comes with an MCP Apps
view, ui://sheetrender/documents.html (text/html;profile=mcp-app): a strip
of previews with PDF links and a "Continue in SheetRender with these rows"
button. Its script, src/widget/documents.ts, is bundled by esbuild into
dist/widget/documents.js during deno task build and inlined into the page
when the resource is read.
Calls are counted per user: ChatGPT's anonymised openai/subject, else the
client IP. Any caller can send openai/subject, so it is believed only from
OpenAI's published egress ranges (chatgpt-connectors.json,
overridable with OPENAI_EGRESS_CIDRS); from anywhere else the call is
counted by IP. The hosted server fetches that list at startup and every 24
hours, keeping the built-in copy (then the last good fetch) whenever a fetch
fails or returns anything malformed or empty; it logs one line per refresh. It
also logs, at most once an hour, how many calls sent an openai/subject from
outside the list, so a stale list shows up.
The server allows 30 render or continue calls per subject or IP per hour
(ANON_CALLS_PER_HOUR). Callers counted by IP also share a second bucket per
IPv4 /24 or IPv6 /48, 300 calls an hour (ANON_NETWORK_CALLS_PER_HOUR), so a
block of cheap addresses counts as one caller. All traffic from Claude's
160.79.104.0/21 network (Anthropic's published outbound range, which the
Claude API's MCP connector also uses) shares a separate 3,000-call hourly
bucket (CLAUDE_CALLS_PER_HOUR); each Claude conversation that sends its
Mcp-Session-Id also gets the 30-call per-subject bucket inside it. Neither
User-Agent text nor _meta can claim or leave these buckets.
The SheetRender API still applies the monthly document volume against the
hashed subject or IP. The subject and IP are sent to the API only as SHA-256
hashes, and the request log carries a fingerprint, the detected client (chatgpt,
claude or other) and the row count, never the rows.
Anonymous HTTP bodies are capped at 2 MB, enough for 25 rows at every
field's limit; a continue link's rows are capped at 256 KB. A private socket peer is
treated as a proxy and only the final valid IP in X-Forwarded-For is
trusted. Keep this listener behind Caddy, with no publicly exposed container
port, as in the deployment's Compose topology.
Anonymous requests have three more bounds, so a small request cannot make
the server generate a lot of output. A JSON-RPC batch may carry at most 4
messages (MCP dropped batching in protocol version 2025-06-18, and ChatGPT
and Claude send one message per request); a larger one gets a 400 before
anything in it runs. Messages other than tool calls (initialize,
tools/list, resources/read, ...) count 600 an hour per IPv4 /24 or IPv6
/48 (ANON_RPC_PER_HOUR); traffic from OpenAI's and Claude's egress ranges
is not counted there, since each of those addresses carries many users. And
at most 64 requests, with a key or without, are answered at once (ANON_MAX_IN_FLIGHT),
8 per network outside those ranges (ANON_NETWORK_MAX_IN_FLIGHT); past
either, the answer is a 503 or 429 with Retry-After: 5. Requests with an
API key the SheetRender API has not yet accepted get all three bounds, with
tool calls counted toward the 600 as well; once the key is accepted they
count only toward the overall in-flight cap, and their batches may carry 20
messages.
A request carrying a SheetRender bearer key uses the API-key tools above.
A malformed Authorization header is rejected. The stdio server never
offers the anonymous tools.
Running it yourself
sheetrender-mcp-http is a second bin in the package. It reads PORT
(default 8080), HOST (default 0.0.0.0), SHEETRENDER_API_URL,
MAX_BODY_BYTES and IDLE_TIMEOUT_MS (default 60000), and logs one JSON
line per request to stdout — method, path, status, duration, the JSON-RPC
method and tool name, a fingerprint of the key (never the key), and
key_verified, whether the API had accepted that key yet.
For the anonymous tools, and the request bounds above (which also apply to
keyed requests when no demo key is set), it also reads:
With SHEETRENDER_DEMO_API_KEY set, the server refuses to start if the view
bundle (dist/widget/documents.js) is missing. The Dockerfile in this repo
builds a non-root runtime image for it:
Tools
design_template: Design from exactly one of dataset_id, inline rows, data_base64 plus data_filename, or stdio-only data_path, and an example, brief or style. Data files must be CSV or XLSX, up to 10 MB. Waits up to 3 minutes and returns status, template details and an authenticated preview URL. Examples use example_base64 plus example_filename, or stdio-only example_path.get_design: Poll a design_id for its status, template id, name, column mapping and preview URL.
render_pdf
Renders one PDF from HTML you supply.
Returns the path of the saved PDF and its size. Under 512 KB it's also attached
inline as a base64 resource, so clients that display attachments show the
document itself.
HTML over 2 MB is rejected, and that's measured both on
what you send and on the document after data is substituted in, so a template
that expands a long dataset can cross the line even when the markup you wrote
doesn't. A "Made with SheetRender" footer is added when required by the
rendering account's settings. That applies to render_pdf and
render_template alike.
list_templates
No arguments. Returns each saved template's name, id and last-updated date. Call
it to turn a template name the user mentioned into the id the other tools want.
render_template
Renders one PDF from a template already saved in the account.
Same return as render_pdf.
create_dataset
Turns JSON rows into a dataset a batch job can render. This is the usual way to
start a batch: assemble the rows, send them, get back a dataset_id.
The header is the union of every row's keys in first-seen order, so rows don't
have to agree on their keys — a missing one is a blank cell rather than a
shifted row. Values have to be scalars: strings, numbers, booleans or null.
Nested objects and arrays are rejected, and so are NaN, Infinity and whole
numbers past 2^53 (send those as strings to keep them exact). The caps are
50,000 rows and 500,000 cells per call.
Returns the dataset id, row count and, for each column, the sanitized key.
That key is what template placeholders, filename_template and group_by
address, and it's often not the header verbatim — Invoice No becomes
invoice_no. Read it off this result instead of guessing.
Creating a dataset does not consume document volume; only rendering does.
upload_dataset
The same thing from a file that already exists.
The first row has to be the header. Files over 20 MB, the wrong extension and
empty files are refused locally, before anything is uploaded. Same return as
create_dataset.
list_datasets
Takes template_id and lists every dataset in that template's project, newest
first, with ids, row counts and column keys. Use it to find data the user
already loaded, or to read a dataset's column keys before writing a
filename_template or picking group_by.
create_batch_job
Queues a background job that renders one PDF per row of a dataset. The whole
loop runs from here — list_templates → create_dataset or upload_dataset →
create_batch_job → get_job → get_document.
Returns the job id to poll with get_job. Worth knowing: filename_template
and group_by are persisted to the template and the project respectively, so
they change the defaults for later runs too.
If the server predates the public batch endpoint, the tool reports that batch
jobs are unavailable rather than failing obscurely. The three dataset tools do
the same for a server that predates the dataset endpoints.
get_job
Takes job_id. Returns the status, rows done and failed, and the id and
filename of every rendered document. That document list stays empty while the
job is queued, retry_queued or running, and fills in once the job reaches
succeeded, partial, failed or cancelled. Those document ids are what
get_document takes.
get_document
Takes document_id and downloads that single rendered PDF. The ids come from
get_job on a finished batch, and there's no other way to get one. Same return
as render_pdf: path, size, and an inline blob under 512 KB.
If you want a whole batch, the merged PDF and ZIP in the web app beat fetching
each document in turn.
page_settings
Shared by both render tools. Every field is optional:
page_size is lowercase: a3, a4, a5, letter, legal or tabloid.
Margins are plain numbers in millimetres, not CSS lengths.
Rendered PDFs are written to the system temp directory. API errors like a bad
key, a missing template or a rate limit come back as tool errors carrying the
server's own message.
The public API allows 120 requests per minute per API key. Past that it returns
429, and the tool reports that you're rate limited and should retry shortly.
Batch job creation is metered separately and more tightly.
Development
You don't need node or npm on the host: scripts/dev.sh install, then
scripts/dev.sh deno task build and scripts/dev.sh deno task test.
scripts/dev.sh node dist/http.js runs the HTTP server on the host network.
MIT licensed.
