MCP File Transfers Are Still a Draft: A Safe Pilot for OpenClaw Workflows
MCP’s proposed file-transfer extension tackles a familiar personal-agent problem: moving a PDF into storage or bringing a large CSV into a sandbox without making the language model carry the file’s bytes.
The practical opportunity is useful, but the status matters. The official MCP blog introduced the work on November 10, 2026; the linked specification is still marked Draft when checked on November 19. This is an optional, experimental extension—not a feature you should assume is available in your OpenClaw installation or on Clawly.
For technical users, now is a good time to design the transfer boundary and a small compatibility pilot rather than rewrite production workflows around an unfinished API.
What the proposal changes
The official explainer and draft specification describe an extension named io.modelcontextprotocol/files, targeting MCP revision 2026-07-28 and later.
Its out-of-band upload flow separates the tool request from the transfer:
- The host obtains user approval or applies an operator policy for unattended work.
- The client calls a tool with file metadata, such as its identifier, media type, and size.
- The server checks what it can before receiving bytes, including destination permissions and quota.
- If accepted, the server returns
upload_required, transfer targets, and an opaquerequestStatevalue. - The client transfers the bytes directly over HTTPS, outside the JSON-RPC exchange.
- After successful transfers, the client repeats the tool call with the same arguments and the returned state. The server verifies the upload before executing the tool.
The model does not need to generate a base64 payload or handle the upload URL. Small inline files are also contemplated, subject to advertised limits; out-of-band transfer is the relevant path for larger artifacts.
Downloads take a different shape: the tool can return download information in result metadata, and the client fetches the artifact without another tool call just to complete the download. That representation remains an open design question.
Two permissions, not one
The most important distinction is between permission to use a tool and permission to send a particular file to a particular destination.
The proposal’s credentials field controls credential handling:
omitmeans the transfer does not inherit the client’s MCP credential or ambient cookies. A signed storage URL can supply its own authorization.includeallows the MCP credential only within the draft’s same-origin or MCP-address path restrictions.
Both the explainer and draft warn that keeping a token private does not, by itself, keep the file private. An upload target could receive confidential bytes without receiving any MCP token at all.
For a pilot, adopt a stricter operator policy: allow only explicitly approved transfer origins and approved file classes. Also reject redirects and targets that attempt to supply their own Authorization header, as the proposal requires. Treat signed URLs as sensitive capabilities; do not put them in chat replies or ordinary diagnostic logs.
These are transfer controls, not protection against malicious document contents. A successfully delivered PDF is still untrusted input when an agent later reads it.
A concrete pilot: save one synthetic report
Start with a generated, non-sensitive PDF and a disposable storage folder. The goal is to prove the boundary, not benchmark model intelligence.
Record the client, server, and draft revision being tested. Confirm that both sides advertise the relevant extension support. A server supporting the base MCP revision is not sufficient evidence that it implements this optional extension.
Use a short acceptance checklist:
| Test | Expected outcome |
|---|---|
| Approved PDF and storage origin | Upload completes; the tool confirms the saved artifact. |
| File exceeds the configured per-file limit | Transfer is rejected before bytes are sent where the metadata check can establish the violation. |
| Target points at an unapproved origin | Client blocks the transfer. |
| Transfer responds with a redirect | Client refuses to follow it. |
| Upload finishes but stored size differs | Server rejects finalization rather than reporting success. |
| Download has a supplied digest that does not match | Client discards the downloaded file. |
| Client does not support the extension | Server omits the extension-dependent tool or offers a documented alternative. |
For the successful case, independently inspect the destination artifact and compare its size and hash with the original. Keep the test report to artifact identifiers, byte counts, verification outcomes, and sanitized errors—not file contents or signed URLs.
This is a proposed test plan, not a claim that these tests have been run against OpenClaw or Clawly.
Plan for partially transferred files
A failed workflow does not necessarily mean that no bytes reached storage. The draft explicitly distinguishes a tool that did not run from bytes that may already have been stored.
Before testing, define who owns temporary uploads, how long staging objects survive, and how cleanup is verified. If an upload succeeds but finalization fails, the operator should be able to locate the staged object without asking the model to reconstruct an upload URL from conversation history.
For a CSV download, add a separate local-processing boundary: save to an approved directory, enforce size limits, verify the artifact, and only then let a sandboxed analysis step read it. Moving bytes outside model context is not permission to execute their contents.
What OpenClaw operators should do now
Do not assume there is a configuration switch that enables this draft across existing MCP connections. Check the actual client or adapter, server implementation, and advertised capabilities first. If support is missing, keep your current approved transfer mechanism instead of asking the model to improvise one.
The useful architectural lesson is already actionable: let the model request the work, let the client move the bytes, and let explicit policy decide where those bytes may go. Adopt the experimental wire format only when your implementation and tests justify it.
Sources and status
- MCP: Exploring File Uploads and Downloads — official explanation of the experimental design.
- File Uploads and Downloads draft — cross-checked for capability negotiation, transfer flow, validation, and failure semantics.
- MCP release index — secondary chronology placing the announcement on November 10, 2026.
Status checked November 19, 2026. Draft fields and requirements may change; no production availability or Clawly rollout is implied.
Protect your AI agent with Clawly
Deploy your OpenClaw agent in an isolated, hardened container with encrypted credentials and managed updates. No DevOps required.
Deploy Your Agent