renfield-mcp-filesystem
Watch-folder MCP server for Renfield: itwatches folders (local / SMB) for settled new files and pushes them intoRenfield over REST (POST /api/folder-ingest/document), which ingests them intothe knowledge base and Paperless.
This dedicated server is the sole access boundary to the shares — theRenfield backend never mounts them, and credentials/clients live only here. A newoff-cluster share or a per-user folder is added at runtime by editing theroots config — no redeploy, no static volume.
Principles
- Event-driven, never polling. Local roots use
watchdog(inotifyCLOSE_WRITE= the settle signal); SMB roots use SMB2CHANGE_NOTIFY+ anevent-debounce timer. No periodic filesystem scan. (One exception: a singleenumeration at startup catches files that already existed before the watchbegan — a one-shot catch-up, not a poll.) - Create-only. Acts on settled new files; ignores in-place rewrites of filesit has already handled.
- The backend's 4-state response drives the move.
ingested|duplicate→processed/,failed→failed/,retry→ left in place and re-attempted ona bounded backoff.401/403is a fatal token error (the file is never moved). - Local safety gates. Size ceiling + extension allowlist are enforcedbefore any push (an oversized / disallowed file goes straight to
failed/).
Quickstart (local folder)
Point it at a local inbox and an existing Renfield backend:
mkdir -p ./inbox
cat > roots.yaml <<'YAML'
roots:
- name: inbox
type: local
path: /watch/inbox
YAML
docker run --rm \
-e RENFIELD_URL=http://renfield-backend:8000 \
-e RENFIELD_INGEST_TOKEN=<token-from-POST-/api/folder-ingest/token> \
-e FILES_ROOTS_YAML=/config/roots.yaml \
-v "$PWD/roots.yaml:/config/roots.yaml:ro" \
-v "$PWD/inbox:/watch/inbox" \
-p 8080:8080 \
registry.treehouse.x-idra.de/renfield/filesystem-mcp:latest
Drop a PDF into ./inbox → it appears in Renfield's /wissen and Paperless, thenmoves to ./inbox/processed/. A rejected file moves to ./inbox/failed/.
Mint the token on the backend (admin): POST /api/folder-ingest/token. Thebackend feature must be on (FOLDER_INGEST_ENABLED=true).
Configuration
Global settings come from the environment; the watch roots come from amounted roots.yaml (see config/roots.example.yaml).
| Env var | Default | Meaning |
|---|---|---|
RENFIELD_URL |
— (required) | Renfield backend base URL |
RENFIELD_INGEST_TOKEN |
— (required) | folder-ingest Bearer token |
FILES_ROOTS_YAML |
— | path to the mounted roots.yaml (reloaded on change) |
FILES_ALLOWED_EXTENSIONS |
pdf,docx,... |
local extension allowlist |
FILES_MAX_FILE_SIZE_MB |
50 |
size ceiling (enforced before push) |
FILES_SETTLE_SECONDS |
2.0 |
SMB settle-debounce window |
FILES_MCP_HOST / FILES_MCP_PORT |
0.0.0.0 / 8080 |
MCP server bind |
FILES_NOTIFY_WEBHOOK_URL / _TOKEN |
— | optional failure/disconnect webhook |
roots.yaml (creds referenced by env-var name, never inlined):
roots:
- name: documents
type: smb
server: nas.example.lan
share: Documents
path: Inbox
username_env: DOCS_SMB_USER
password_env: DOCS_SMB_PASS
- name: local-inbox
type: local
path: /watch/inbox
Each root takes an optional processed_subdir / failed_subdir (defaultsprocessed / failed). After a file is handled it is moved out of the inbox:ingested/duplicate → processed, rejected → failed, retry → left in place.
Where the processed/failed dirs live differs by provider:
- SMB — at the share root, as siblings of the watched
path. A rootwithpath: Inboxproduces<share>/{Inbox, processed, failed}(not<share>/Inbox/processed). Withpath: ""(watch the share root) they aresimply the two top-level dirs. The watched inbox + both dirs are auto-createdon connect. - local — nested inside the watched
path(<path>/processed,<path>/failed), since a local root is self-contained.
Dry-run (preflight)
Validate config + credentials + the matched/skipped files before the daemontouches anything (pushes nothing, moves nothing):
renfield-mcp-filesystem-scan --dry-run
# root documents (smb):
# would push (2): invoice.pdf (12345 bytes), letter.pdf (6789 bytes)
# skipped (1): notes.exe (extension_not_allowed)
Interactive MCP tools
Registered as mcp.files.* (the files stanza in Renfield'sconfig/mcp_servers.yaml). The agent uses these to browse + ingest on demand(the watch loop is automatic + event-driven):
list_watch_folders()→ roots +connected+last_errorlist_files(root, pattern?)→ files, each with a qualifiedpath"<root>/<relpath>"get_file_info(path)·read_file(path, truncate?)·move_file(path, subdir)
The Renfield agent tool internal.ingest_file({path}) pulls bytes viaread_file(path, truncate=False) and runs them through the same ingest bridge.
Deploy (k8s)
Manifests in k8s/ (ConfigMap roots + Secret creds + a single-replicaDeployment + Service). Single replica by design — two would double-push. Buildon the build box → Harbor → kubectl apply -f k8s/.
Develop
python3 -m venv .venv && .venv/bin/pip install -e '.[dev]'
.venv/bin/python -m pytest
Core modules (config/contract/providers/pusher/engine/gate/daemon/tools/scan)are mcp-free and fully unit-tested; the live inotify/SMB CHANGE_NOTIFY wiring andthe cross-repo push are verified by the Renfield .159 E2E. NFS is deferred — ithas no native change-notification, so it cannot be event-driven without polling.