Aggregated File System (Agent FS) - Everything is a file, in RESTful APIs. A tribute to Plan9.
When coordinating multiple AI Agents in a distributed environment, agents need access to various backend services: message queues, databases, object storage, KV stores, and more. The traditional approach requires writing specialized API calls for each service, meaning agents must understand many different interfaces.
The core idea of AGFS is simple: unify all services as file system operations.
Traditional approach AGFS approach
------------------------------------------------------------------
redis.set("key", "value") -> echo "value" > /kvfs/keys/mykey
sqs.send_message(queue, msg) -> echo "msg" > /queuefs/q/enqueue
s3.put_object(bucket, key, data) -> cp file /s3fs/bucket/key
mysql.execute("SELECT ...") -> echo "SELECT ..." > /sqlfs2/.../query
The benefits:
- AI understands file operations natively - Any LLM knows how to use cat, echo, and ls. No API documentation needed.
- Unified interface - Operate all backends the same way, reducing cognitive overhead.
- Composability - Combine services using pipes, redirections, and other shell features.
- Easy debugging - Use ls and cat to inspect system state.
Install:
curl -fsSL https://raw.githubusercontent.com/c4pt0r/agfs/master/install.sh | shOr via Docker:
docker pull c4pt0r/agfs:latest
# Run the server (HTTP API only)
docker run -p 8080:8080 -e SKIP_FUSE_MOUNT=true c4pt0r/agfs:latest
# On Linux, you can enable FUSE mounting with additional privileges
docker run -p 8080:8080 \
--device /dev/fuse \
--cap-add SYS_ADMIN \
--security-opt apparmor:unconfined \
c4pt0r/agfs:latest
# Note: FUSE mounting in Docker is not supported on macOSConnect using agfs-shell:
$ agfs
agfs:/> ls
queuefs/ kvfs/ s3fs/ sqlfs/ heartbeatfs/ memfs/ ...AGFS can be mounted as a native filesystem on Linux using FUSE. This allows any program to interact with AGFS services using standard file operations, not just the agfs-shell.
# Mount AGFS to /mnt/agfs
agfs-fuse --agfs-server-url http://localhost:8080 --mount /mnt/agfs
# Now use standard tools
ls /mnt/agfs/kvfs/keys/
echo "hello" > /mnt/agfs/kvfs/keys/mykey
cat /mnt/agfs/queuefs/tasks/dequeueThis makes AGFS accessible to any application, script, or programming language that can read and write files.
See agfs-fuse/README.md for installation and usage.
The simplest key-value storage. Filename is the key, content is the value:
agfs:/> echo "world" > /kvfs/keys/hello # write
agfs:/> cat /kvfs/keys/hello # read -> "world"
agfs:/> ls /kvfs/keys/ # list all keys
hello
agfs:/> rm /kvfs/keys/hello # deleteA message queue is abstracted as a directory containing control files:
agfs:/> mkdir /queuefs/tasks # create queue
agfs:/> ls /queuefs/tasks
enqueue dequeue peek size clear
agfs:/> echo "job1" > /queuefs/tasks/enqueue # enqueue
019aa869-1a20-7ca6-a77a-b081e24c0593
agfs:/> cat /queuefs/tasks/size # check queue length
1
agfs:/> cat /queuefs/tasks/dequeue # dequeue
{"id":"019aa869-...","data":"job1","timestamp":"2025-11-21T13:54:11Z"}This pattern is ideal for AI Agent task distribution: one agent writes tasks to the queue, another agent reads and executes them.
Query databases through a Plan 9 style session interface:
agfs:/> cat /sqlfs2/mydb/users/schema # view table structure
agfs:/> cat /sqlfs2/mydb/users/count # get row count
# Create session, execute query, read result
agfs:/> sid=$(cat /sqlfs2/mydb/users/ctl)
agfs:/> echo "SELECT * FROM users LIMIT 2" > /sqlfs2/mydb/users/$sid/query
agfs:/> cat /sqlfs2/mydb/users/$sid/result
[{"id": 1, "name": "alice"}, {"id": 2, "name": "bob"}]Manage the liveness state of distributed agents:
agfs:/> mkdir /heartbeatfs/agent-1 # register agent
agfs:/> touch /heartbeatfs/agent-1/keepalive # send heartbeat
agfs:/> cat /heartbeatfs/agent-1/ctl # check status
last_heartbeat_ts: 2025-11-21T13:55:45-08:00
timeout: 30
status: alive
# After 30 seconds without a new heartbeat, the agent directory is automatically removedDifferent filesystems can operate with each other:
agfs:/> cp local:/tmp/data.txt /s3fs/mybucket/ # upload local file to S3
agfs:/> cp /s3fs/mybucket/config.json /memfs/ # copy S3 file to memoryAGFS shell supports scripting with .as files. Scripts use familiar shell syntax and can be executed directly.
task_worker.as - A simple task queue worker:
#!/usr/bin/env agfs
QUEUE_PATH=/queuefs/tasks
POLL_INTERVAL=2
# Initialize queue
mkdir $QUEUE_PATH
while true; do
size=$(cat $QUEUE_PATH/size)
if [ "$size" = "0" ]; then
echo "Queue empty, waiting..."
sleep $POLL_INTERVAL
continue
fi
# Dequeue and process task
task=$(cat $QUEUE_PATH/dequeue)
echo "Processing: $task"
# Your task logic here
doneenqueue_task.as - Enqueue a task:
#!/usr/bin/env agfs
mkdir /queuefs/tasks
echo "$1" > /queuefs/tasks/enqueue
echo "Task enqueued. Queue size: $(cat /queuefs/tasks/size)"Run scripts directly:
./task_worker.as &
./enqueue_task.as "process report.pdf"See more examples in agfs-shell/examples.
A typical agent coordination pattern: multiple agents fetch tasks from the same queue and execute them.
while True:
task = agfs.cat("/queuefs/tasks/dequeue")
if task:
result = execute_task(task)
agfs.write(f"/kvfs/keys/result_{task.id}", result)See task_loop.py for a complete example.