swarmfs

An fsspec backend for Ethereum Swarm, talking to a Bee node (or public gateway) over its HTTP API. Installing it makes Swarm a first-class storage backend for the Python data ecosystem — pandas, dask, zarr, xarray, pyarrow, DuckDB — via URLs like bzz://<reference>/path/to/file.parquet.

Status: v3. Read-only bzz:// access, transactional copy-on-write writes (postage stamps, every commit a snapshot), mutable feed-backed bzzf:// mounts, encryption and ACT access control, a read-only FUSE mount, and a local-first mode: commits land on local disk instantly and sync to Swarm in the background — offline is the normal mode, fs.sync() is the certainty barrier. See the roadmap and the local-first design in docs/localstore-design.md.

New to swarmfs? This README is the quick tour — the User Guide walks through a worked example for every library above (pandas, all three Dask collection types, Zarr, xarray, PyArrow, DuckDB) and explains the content-addressing model in plain terms. For lookup — a storage option, a signature, an error, a default — use the Reference: definition-first tables, pinned against the code by the test suite, and the right document to hand to an AI agent.

Install

pip install swarmfs            # or: pip install "swarmfs[feeds]" for signed feeds
                               #     pip install "swarmfs[fuse]"  for `swarmfs mount`

Upload and download a file

You need a running Bee light node (http://localhost:1633 by default) and, for uploads, a usable postage stamp:

import fsspec

fs = fsspec.filesystem("bzz", stamp="auto")

ref = fs.upload("photo.jpg")                    # → "c0ffee…" (64 hex chars)
fs.download(f"bzz://{ref}/photo.jpg", "copy.jpg")

On Swarm the address of new content is the result of a write, not its input — upload returns the new reference, and that reference is permanent: it names this exact content forever. Directories work the same way and come back as a single reference for the whole tree:

ref = fs.upload("dataset/")
fs.ls(f"bzz://{ref}")
fs.download(f"bzz://{ref}", "dataset-copy/", recursive=True)

upload accepts content_type= (otherwise guessed from the filename), encrypt=True (files and directories; the returned 128-hex reference includes the decryption key), and redundancy=0–4 (erasure coding, default 2). The stamp is validated before any byte moves, so a missing or expired stamp fails immediately with an actionable error.

The data ecosystem

The point of being an fsspec backend: everything that speaks fsspec now speaks Swarm, with zero extra code.

import pandas as pd

df = pd.read_parquet("bzz://<64-hex-reference>/data.parquet")

# local caching via URL chaining
df = pd.read_parquet("simplecache::bzz://<reference>/big.parquet")

import fsspec

fs = fsspec.filesystem("bzz")          # api_url=..., default $BEE_API_URL or localhost:1633
fs.ls("bzz://<reference>/")            # client-side Mantaray trie walk
fs.find("bzz://<reference>/dataset/")  # recursive listing (dask uses this)
fs.cat("bzz://<reference>/hello.txt")

with fs.open("bzz://<reference>/big.parquet", block_size=2**20) as f:
    f.seek(-8, 2)                      # range requests: only the bytes you touch
    f.read(8)

Same story for any other fsspec-based tool — Intake, DVC, Kedro, pyxet, Hugging Face Datasets, petl, and more (see the User Guide).

Transactional writes

For anything beyond a one-shot upload — building a dataset in place, changing one file inside a large collection — writes are copy-on-write commits: each commit patches the manifest trie client-side, re-uploads only what changed, and yields a new root. Old roots are untouched, so every commit is a snapshot.

fs = fsspec.filesystem("bzz", stamp="auto")
with fs.transaction:
    fs.pipe_file("bzz://new/dataset/a.parquet", data_a)
    fs.pipe_file("bzz://new/dataset/b.parquet", data_b)
root = fs.latest("new")          # share this reference; it never changes

Local-first writes

Add local_store= and the network leaves the write path entirely: commits land in a local store directory instantly — they work on a plane — and a background worker pushes them to Swarm and confirms arrival peer-to-peer. Reads of anything the store holds are served from disk too (offline read-your-writes, including ls and ranged reads); foreign references still read through the node.

fs = fsspec.filesystem("bzz", local_store="~/.myapp/store", redundancy=0)
with fs.transaction:
    fs.pipe_file("bzz://new/dataset/a.parquet", data_a)   # instant, offline-safe
root = fs.latest("new")
fs.sync()                        # optional barrier: confirmed ON the network
print(fs.sync_status())          # pinned vs evictable bytes, batch expiries

No postage stamp is needed at commit time (the push owns postage), local disk becomes a budgeted working set (unpushed data is pinned; only Swarm-confirmed blobs evict, and evicted reads heal by verified re-fetch), and on bzzf:// mounts the feed update publishes only once the network provably serves the new root. redundancy=0 is required — erasure coding would fork the node's references from the local address space. Full design: docs/localstore-design.md.

Mutable feeds (bzzf://)

A feed gives you a stable URL whose contents you can update — the mutable filesystem on top of immutable commits:

ffs = fsspec.filesystem("bzzf", stamp="auto", signer="<private key hex>")
ffs.pipe_file(f"bzzf://{owner}/my-app/config.json", b'{"v": 2}')
# readers need no keys — and the URL never changes

Access control (ACT)

Swarm's ACT lets a publisher decide who can read a reference — by their nodes' public keys — and change that list later. In swarmfs it is a storage option:

pub = fsspec.filesystem("bzz", stamp="auto", act=True)
root = pub.upload("private/")           # an ACT reference; encrypted by default
history = pub.act_history               # SAVE THIS — it is what unlocks the content

# a grantee's node (or the publisher's own) reads with the history; the
# publisher's key is needed too and defaults to the reading node's own
reader = fsspec.filesystem("bzz", act_history=history, act_publisher=pub.publisher_key())
reader.cat(f"bzz://{root}/report.parquet")
fsspec.filesystem("bzz").ls(f"bzz://{root}")   # anyone else: FileNotFoundError

gl = pub.create_grantees([friend_public_key])       # a grantee list + its history
pub2 = fsspec.filesystem("bzz", stamp="auto", act=True, act_history=gl.history)
gl = pub.patch_grantees(gl.reference, gl.history, revoke=[friend_public_key])

Facts that shape the design, all measured live: an ACT reference is the real reference encrypted, same length, so it looks like any other; only the root is wrapped (children resolve normally, and the content itself stays plaintext-addressable — which is why act=True implies encrypt=True); reading needs the history and the publisher's key, and happens through a node holding the publisher's or a grantee's private key — so ACT works only against your own node, never a gateway. Losing the history loses the content, for the publisher too.

Mount it as a folder

For everything that is not Python — a shell, an editor, rsync, a tool that only takes a local path — a reference or a feed can be mounted as an ordinary directory:

pip install "swarmfs[fuse]"                    # plus libfuse2 (below)
mkdir ~/mnt/dataset
swarmfs mount bzz://<reference> ~/mnt/dataset  # or a bare 64-hex reference
ls -la ~/mnt/dataset; head ~/mnt/dataset/data/part-00000.parquet
fusermount -u ~/mnt/dataset                    # or Ctrl-C the mount

The mount is read-only by default: a bzz:// reference is immutable by construction, and a bzzf://<owner>/<topic> mount is a live view of the feed — it follows updates (--feed-ttl, default 15 s) without ever changing URL. Files are 0444, directories 0555, sizes are real, reads are ranged (the kernel and fsspec both cache), and simplecache::bzz://<ref> mounts with a local disk cache for free. From Python the same thing is swarmfs.fuse.mount(url, mountpoint); a gateway works too (--api-url … --allow-gateway, with chunk verification on).

--rw makes it writable — every file you save is one commit:

swarmfs mount --rw bzz://<reference> ~/mnt/work    # needs a usable stamp (checked first)
cp report.parquet ~/mnt/work/data/; rm ~/mnt/work/old.csv; mv ~/mnt/work/a ~/mnt/work/b
fusermount -u ~/mnt/work                            # prints the new root: bzz://<new>

A bzz:// mount keeps showing the latest state while mounted (read-your-writes) and prints the final root at unmount — the original reference is untouched, every commit was a snapshot. A bzzf:// mount with --rw --signer <key> publishes the feed on every commit, so the URL never changes. The commit happens on close, so cp and editors see a refused commit (no stamp, a rejected write) as an error instead of losing it; mkdir gives you an empty directory that becomes real when a file lands in it (Mantaray has no empty directories); chmod/touch are accepted and ignored. Other fsspec filesystems can be mounted through the same mounter with fs=, writable too — that is how ontodag-fs mounts its lattice view.

Caveat: FUSE support comes from fsspec's generic wrapper over fusepy, which needs a system libfuse 2 — apt install libfuse2 (libfuse2t64 on Ubuntu 24.04+), macFUSE on macOS; no Windows. If either is missing, swarmfs mount says so and exits; the rest of swarmfs is unaffected.

Addressing content offline, and buying stamps

Two things you can do without uploading anything:

import swarmfs

ref = swarmfs.content_address(open("photo.jpg", "rb").read())   # no node needed
root, chunks = swarmfs.split(data)   # the whole chunk tree, keyed by address

content_address computes the reference Bee would return for those bytes, so you can check whether the network already has content, name blobs by their Swarm address in a local store, or build test fixtures at real addresses. It needs the feeds extra (keccak256).

One caveat that bites in practice: this is the reference for a plain upload. Erasure coding changes every intermediate chunk and therefore the root, and parity is the node's to generate — so a redundant upload's reference cannot be predicted offline. swarmfs writes with redundancy=2 by default and many nodes default to redundancy too, so pass redundancy=0 if you want the uploaded reference to match the one you computed.

Stamps can also be handled programmatically — selection is automatic (stamp="auto" picks the usable batch with the longest life), and spending is available but never implicit. Every plan_* call is a pure question; only the verbs move money:

from swarmfs.stamps import StampManager, depth_for_addresses, stamped_chunks

plan = await mgr.plan(size_bytes, ttl_secs)   # depth, amount, cost in xBZZ
batch = await mgr.buy(plan.amount, plan.depth)  # spends the node wallet's xBZZ

# depth depends on how you will upload: erasure parity and encryption add
# stamped chunks, and a batch dies of a full *bucket*, not of full capacity
plan = await mgr.plan(size_bytes, ttl_secs, redundancy=4, encrypted=True)
# ...or skip the statistics entirely — a plain upload's addresses are known
root, chunks = swarmfs.split(data)
plan = await mgr.plan(size_bytes, ttl_secs, depth=depth_for_addresses(chunks))

# renewal: extend BY a duration, TO a total, or for at most a budget
plan = await mgr.plan_topup(batch, ttl_secs=30 * 86400)
print(plan.cost_bzz, plan.total_ttl_secs, plan.warning)
info = await mgr.topup(batch, plan.added_amount)   # waits until the node applies it

# capacity, not time: dilution costs gas, and is paid for in TTL
print((await mgr.plan_dilute(batch, 20)).ttl_after_secs)   # ~halved per step

Four things about renewal that are easy to get wrong, so swarmfs encodes them: a topup adds to the remaining life rather than restarting it; a nearly-full immutable batch should be diluted before topping up, or the dilution halves away part of what you just bought (plan_topup().warning says so); remaining life is the node's batchTTL and never amount / currentPrice (that field describes lifetime from the creation block, and is local bookkeeping that can revert); and an expired batch cannot be revived, so renew while it lives. The node also takes ~40 s to index a topup — topup() polls, because reading straight after the transaction shows the old value and looks like a silent failure.

For monitoring, mgr.list_batches() plus StampInfo.problem(min_ttl) turns "still usable" into "needs renewing" at whatever threshold you choose, and mgr.buckets(batch) reports the true per-bucket headroom that bounds the next upload — utilizationRatio only summarises it.

plan sizes the batch for your upload (bucket-overflow-aware) and prices it from the chain; buy purchases and waits until the batch is usable. Nothing in swarmfs buys on its own — deciding to spend is the caller's.

Which API should I use?

Three tiers, all backed by the same endpoint resolution (api_url=... → $BEE_API_URL → http://localhost:1633):

• SwarmFileSystem / fsspec URLs — the default. Filesystem semantics, transactions, verification, and the whole data ecosystem for free.

• swarmfs.SyncSwarmClient / swarmfs.SwarmClient — direct calls against the Bee API (upload a blob, fetch bytes, post a feed update) without filesystem semantics. SyncSwarmClient is the blocking twin for plain scripts; SwarmClient is the same surface as coroutines for asyncio code:

  from swarmfs import SyncSwarmClient
  
  with SyncSwarmClient() as client:            # async? use SwarmClient + await
      ref = client.bzz_post(open("photo.jpg", "rb"), stamp=batch_id)
      data = client.bzz_get(ref, "photo.jpg")
  

• Raw HTTP — the Bee API is plain HTTP; no library needed:

  curl -X POST -H "Swarm-Postage-Batch-Id: <batch>" \
       --data-binary @photo.jpg http://localhost:1633/bzz?name=photo.jpg
  

  What the library adds over this: stamp validation up front, chunk verification, gateway policy, better errors — the edge cases.

Nodes, gateways, verification

The recommended setup is a local light node — reads then come straight from the network with nothing to trust in between. Pointing api_url at a public gateway is discouraged and requires an explicit allow_gateway=True — on that path swarmfs verifies every fetched chunk client-side against its BMT address (a Swarm reference is the content hash), so even an untrusted gateway can't tamper with what you read. Verification can also be forced on/off with verify=True/False.

How it works

Swarm has no server-side directory listing today, so swarmfs parses the binary Mantaray manifest trie itself, fetching nodes on demand via /bytes (see swarmfs/mantaray/ — a self-contained pure-Python codec). File reads resolve the path to its data reference once, then use HTTP range requests against /bytes, which is what makes Parquet predicate pushdown and zarr chunk reads viable. When Bee grows a server-side listing endpoint (ethersphere/bee#5535) it will slot in behind the existing capability seam with no API change.

Compared to ipfsspec

ipfsspec, the closest analog in the fsspec ecosystem, is read-only by its own admission. Postage stamps make paid writes tractable on Swarm, so swarmfs adds a transactional write path and, via bzzf:// feeds, a stable URL you can actually mutate — not just read.

Development

pip install -e ".[test]"
pytest                                   # 412 tests; the live ones skip with no node
SWARMFS_TEST_BEE=http://localhost:1633 \
SWARMFS_TEST_STAMP=<batch-id> pytest tests/test_integration.py