Show HN: A durable filesystem layer for AI agents

CelestoAI released SmolFS, an open-source durable filesystem layer for AI agents that persists workspace data across agent process restarts. The tool mounts as a standard directory, supports local and cloud storage backends including Redis and S3, and provides Python and TypeScript SDKs wrapping a Rust core. SmolFS aims to simplify state management for short-lived agent runtimes by decoupling storage setup from agent code.

SmolFS gives agents a workspace folder that can survive after the agent process stops. You mount it like a normal directory, write files into it, unmount it when the job is done, and mount it again later from another process. SmolFS owns the developer experience around creating volumes, checking the machine, mounting, flushing, unmounting, and inspecting status. The storage backend is installed and managed for you. | Agents can keep files across short-lived runtimes without each runtime managing storage setup directly. | Use | | Use Redis plus S3-compatible object storage when the same workspace needs to outlive one machine. | Run | | Python and TypeScript bindings call the same Rust core so agent tools can use SmolFS without shelling out. | Cloud metadata, buckets, and credentials stay explicit so durable agent data is easier to audit. | Keep agent work across turns. Mount the same workspace again after an agent process exits. Share a workspace across runtimes. Put metadata in Redis and file contents in S3-compatible storage. Test locally before using cloud storage. Start with --dev , then switch to explicit metadata and object storage settings. Wrap storage in agent tooling. Use the Python or TypeScript SDK from an agent runner instead of teaching every agent process about storage internals. Install SmolFS: curl -fsSL https://raw.githubusercontent.com/CelestoAI/smolfs/main/scripts/install.sh | sh The installer downloads the latest CLI release asset for your platform and installs SmolFS' managed storage backend. If no release asset exists yet, use the source checkout flow in Development development . Set SMOLFS INSTALL BACKEND=0 if you only want to install the CLI. Set SMOLFS VERSION=dev to try the latest successful build from main ; tagged v releases remain the stable channel. Check the machine and try a local volume: smolfs doctor smolfs init demo --dev smolfs mount demo ./workspace echo hello ./workspace/hello.txt smolfs flush demo smolfs unmount demo smolfs mount demo ./workspace cat ./workspace/hello.txt SmolFS needs local mount support on the machine that mounts volumes. Mount support lets SmolFS provide a folder your tools can read and write. Install the Python SDK with the CLI curl -fsSL https://raw.githubusercontent.com/CelestoAI/smolfs/main/scripts/install.sh | SMOLFS INSTALL PYTHON=1 sh The installer runs uv add smolfs from a directory with pyproject.toml , or uv pip install smolfs inside an active virtualenv. Set SMOLFS PYTHON MODE=user to use uv pip install --user smolfs . The normal SmolFS flow is: check the machine, create a volume, mount it, use the directory, flush important writes, unmount it, and mount it again when you need the same files later. Run doctor before creating or mounting volumes: smolfs doctor doctor checks whether SmolFS has its managed storage backend and whether the machine can mount local directories. Useful options: smolfs doctor --install installs SmolFS' managed storage backend. smolfs doctor --json prints the same report as JSON for scripts. SmolFS looks for its home directory in SMOLFS HOME . If it is not set, SmolFS uses ~/.smolfs . The home directory stores SmolFS config, volume records, logs, managed binaries, and local dev-volume data. A volume is the named workspace that SmolFS can mount later. smolfs init demo --dev --dev creates a local-only volume. It uses SQLite for metadata and local files for object data under the SmolFS home directory. This is the simplest path for trying SmolFS on one machine. Cloud volumes need explicit metadata and object storage settings. Metadata stores the file tree. Object storage is where file contents live. smolfs init agent-workspace \ --metadata redis://localhost:6379/1 \ --storage s3 \ --bucket https://my-bucket.s3.us-east-2.amazonaws.com You can pass object storage in either of these forms: --store s3://bucket/prefix , --store gs://bucket/prefix , or --store file:///path/to/objects . --storage TYPE --bucket BUCKET , which is useful for S3-compatible services that expect an endpoint-style bucket URL. For Cloudflare R2 or another S3-compatible service, keep credentials in the environment used by SmolFS. Do not put access keys in command arguments or logs. set -a . ./.env set +a export SMOLFS HOME=/tmp/smolfs-r2-home VOL="r2demo-$ date +%Y%m%d%H%M%S " smolfs init "$VOL" \ --metadata "$SMOLFS R2 METADATA" \ --storage s3 \ --bucket "$SMOLFS R2 BUCKET URL" Mounting makes the volume appear as a normal local directory: smolfs mount demo ./workspace echo hello ./workspace/hello.txt cat ./workspace/hello.txt SmolFS creates the mount directory if it does not exist. After the mount succeeds, programs can read and write files through that directory. Useful options: --check-storage asks SmolFS to test object storage access before the mount completes. --foreground runs the mount process in the foreground instead of starting it in the background. Run flush when you want a best-effort check that recent writes have reached the mounted filesystem: smolfs flush demo Run status to see known volumes: smolfs status smolfs status demo smolfs status --json Unmount when the job is done: smolfs unmount demo unmount asks SmolFS to flush before detaching the mountpoint. Use smolfs umount demo if you prefer the shorter alias. Add --force when the mountpoint is busy and you want SmolFS to force the unmount. After unmounting, you can mount the same volume again and read the files: smolfs mount demo ./workspace cat ./workspace/hello.txt | Command | What it does | |---|---| smolfs doctor | Checks SmolFS storage, local mount support, and configuration. | smolfs init NAME --dev | Creates a local development volume. | smolfs init NAME --metadata URL --storage TYPE --bucket BUCKET | Creates a cloud volume with explicit metadata and object storage. | smolfs mount NAME PATH | Mounts a volume at a local directory. | smolfs flush NAME | Probes the mounted volume and syncs a small file through it. | smolfs status NAME | Shows known volumes and current mountpoints. | smolfs unmount NAME | Unmounts a mounted volume and asks SmolFS to flush. | smolfs umount NAME | Alias for smolfs unmount NAME . | Every command has its own help page: smolfs help smolfs init --help The Python package is SDK-only. Install it with uv : uv add smolfs For local development from this checkout: uv run --isolated --with-editable ./bindings/python python -c "from smolfs import doctor; print doctor " Use the SDK from any Python agent runner: python from pathlib import Path from smolfs import SmolFS, doctor report = doctor if not report "storage backend" "found" or not report "mount support" "found" : raise RuntimeError f"SmolFS is not ready: {report}" fs = SmolFS.from env volume = fs.ensure volume "demo", dev=True mount = fs.mount volume.name, "./workspace" workspace = Path mount.mountpoint workspace / "hello.txt" .write text "hello from SmolFS\n" try: fs.flush volume.name finally: fs.unmount volume.name Cloud volumes use the same API with explicit metadata and object storage: fs.ensure volume "agent-workspace", metadata="redis://localhost:6379/1", storage="s3", bucket="https://my-bucket.s3.us-east-2.amazonaws.com", For S3-compatible services such as MinIO or Cloudflare R2, pass the service bucket URL and provide ACCESS KEY and SECRET KEY in the environment used by SmolFS. The TypeScript package is a native Node.js binding over the same Rust core. The npm package ships prebuilt native bindings for Linux and macOS on x86 64 and arm64. Install it with npm: npm install @celestoai/smolfs For local development from this checkout, use Node.js 18 or newer: cd bindings/node npm ci npm run build:debug npm test Use the SDK from a Node.js agent runner: js import { SmolFS, doctor } from "@celestoai/smolfs"; import { writeFile } from "node:fs/promises"; import { join } from "node:path"; const report = doctor ; if report.storageBackend.found || report.mountSupport.found { throw new Error SmolFS is not ready: ${JSON.stringify report } ; } const fs = SmolFS.fromEnv ; const volume = fs.ensureVolume { name: "demo", dev: true } ; const mount = fs.mount { name: volume.name, path: "./workspace" } ; try { await writeFile join mount.mountpoint, "hello.txt" , "hello from SmolFS\n" ; fs.flush volume.name ; } finally { fs.unmount volume.name ; } Cloud volumes use the same options object: fs.ensureVolume { name: "agent-workspace", metadata: "redis://localhost:6379/1", storage: "s3", bucket: "https://my-bucket.s3.us-east-2.amazonaws.com" } ; Work from a source checkout when you are changing SmolFS itself or when a CLI release asset has not been published yet. Build and check the CLI: cargo build -p smolfs-cli ./target/debug/smolfs doctor Run the normal quality checks: cargo fmt --all -- --check cargo clippy --workspace -- -D warnings cargo test --workspace Run the R2-style lifecycle from this checkout: cargo build -p smolfs-cli set -a . ./.env set +a export SMOLFS HOME=/tmp/smolfs-r2-home VOL="r2demo-$ date +%Y%m%d%H%M%S " MOUNT="/tmp/smolfs-r2-workspace" ./target/debug/smolfs init "$VOL" \ --metadata "$SMOLFS R2 METADATA" \ --storage s3 \ --bucket "$SMOLFS R2 BUCKET URL" ./target/debug/smolfs mount "$VOL" "$MOUNT" echo "hello from smolfs r2" "$MOUNT/hello.txt" ./target/debug/smolfs flush "$VOL" ./target/debug/smolfs unmount "$VOL" ./target/debug/smolfs mount "$VOL" "$MOUNT" cat "$MOUNT/hello.txt" Run the MinIO integration test path when the SmolFS storage backend, Redis, and a MinIO bucket are available: SMOLFS RUN INTEGRATION=1 cargo test --workspace -- --nocapture Build the Python wheel: uvx maturin build --manifest-path bindings/python/Cargo.toml --interpreter python Develop the Python binding locally: uvx maturin develop --manifest-path bindings/python/Cargo.toml Test the TypeScript SDK: cd bindings/node npm ci npm test SmolFS stores agent workspace data outside the sandbox lifecycle. Treat it like durable infrastructure. - Do not log credentials, S3 access keys, Redis URLs with secrets, or mount tokens. - Prefer explicit object-store configuration over hidden global state. - Make mount and unmount behavior idempotent where possible. - Fail loudly on missing storage backend, missing metadata URLs, missing object-store config, or missing local mount support. - Avoid changes that weaken persistence guarantees without calling them out. The smolfs command is built from the Rust CLI crate. The GitHub workflow at .github/workflows/publish-cli.yml builds Linux and macOS release binaries for x86 64 and arm64 targets, smoke-tests smolfs --help , and attaches tarballs to v GitHub releases. Pushes to main publish a rolling dev prerelease after deleting the previous dev release and tag, so the dev channel only exposes the latest successful main build. Python packaging uses uv and maturin . The GitHub workflow at .github/workflows/publish-python.yml builds wheels for Linux and macOS, builds an sdist, and publishes to PyPI when a v tag is pushed. Node.js packaging uses npm and napi-rs . The GitHub workflow at .github/workflows/publish-node.yml builds native TypeScript SDK bindings for Linux and macOS, checks the npm package contents, and publishes to npm when a v tag is pushed. Before the first Python release, configure PyPI Trusted Publishing: - Create or claim the smolfs project on PyPI. - Add a trusted publisher for repository CelestoAI/smolfs . - Set the workflow name to publish-python.yml . - Set the environment name to pypi . Before the first npm release, configure npm trusted publishing for @celestoai/smolfs : - Add a trusted publisher for GitHub Actions. - Set the organization or user to CelestoAI . - Set the repository to smolfs . - Set the workflow filename to publish-node.yml . - Set the environment name to npm . - Allow npm publish . Release: git tag v0.1.0 git push origin v0.1.0 - Add type stubs for the Python package. - Add a Linux CI job that mounts a local dev volume when mount support is available. - Add release notes and a changelog before the first non-draft release. Apache 2.0. Celesto AI https://celesto.ai