# Fyle for AI labs

**A shared workspace for researchers and agents, on your own buckets**

Labs already buy storage for training data. What researchers lack is a shared, versioned place on their desktops where they and their agents work on eval sets, experiment configs, notes and reports. Fyle plans to be that workspace, running on top of the lab’s own buckets.

## Cluster storage isn’t a researcher’s workspace

Training storage is built to feed GPU clusters. Researchers still pass eval sets, configs and write-ups around through ad hoc buckets, notebooks and chat, with little record of who, or which agent, changed what.

As more work is done by agents, labs need to know which agent session wrote each file, and to stop agents changing files people own without someone signing off.

## How Fyle fits

Available now:

- Organisations with roles: owner, admin, billing, member and viewer. (Available now)
- Organisation security policies: maximum session age, idle time-out and how often desktop devices must sign in again. (Available now)
- Device sign-in for the desktop app and CLI: approve a short code in the browser, with no password typed into a terminal. (Available now)

Planned:

- Runs on any S3-compatible storage you already have (AWS S3, Google Cloud Storage, R2, MinIO, Ceph); Fyle stores only immutable objects and the coordinator decides which write wins. (Planned: Enterprise)
- Agents run with `fyle agent` in their own session folder; each session records the agent, user, device and task. (Planned: agent layer)
- Agent grants for MCP and API agents: scoped to a workspace and paths, read, write or branch-only, short-lived. (Planned: agent layer)
- Review rules such as /datasets/ to the data platform team; agents never overwrite human-owned files if policy says so. (Planned: review)
- Deployment options: Fyle-hosted dedicated, in your own cloud account, or on-premises. (Planned: Enterprise)

## An agent at work

The agent runs the scoring script, writes a new results file and updates the shared results table. The eval set itself is owned by the data platform team, so the agent’s cleanup of two items becomes a branch.

```
$ cd ~/Fyle/evals/reasoning-v3
$ fyle agent claude --task "Re-run the scoring script on the new eval split and update the results table"

Session ended · 3 changes
  results/2026-10-run-07.json  new
  results/summary.md           modified
  data/split-b.jsonl           branched (human-owned path)
2 files landed. 1 branch waiting for review by @data-platform.
```

Illustration of a planned workflow. Names are examples; commands and output may change before release.

## Data location, agent identity and audit

- Region-locked workspaces cover content, metadata, indexes, previews, logs and backups, not just the bucket. UK-only residency uses your UK bucket and a UK-hosted coordinator.
- Full audit log of commits, reviews, permission changes and agent grants, with SIEM export (planned for Enterprise).
- Fyle’s own search and agent features treat text inside files as data, never as instructions.
- Certifications are not in place yet. Cyber Essentials Plus, then ISO 27001 or SOC 2, are on the Enterprise roadmap.

Not legal advice. Fyle features mentioned here are planned unless marked available now.

## Related comparisons

- [Fyle vs Archil](https://fyle.sh/compare/fyle-vs-archil)
- [Fyle vs Mesa](https://fyle.sh/compare/fyle-vs-mesa)
- [Fyle vs Space](https://fyle.sh/compare/fyle-vs-space)

## Sources

Last checked: October 2026.

1. [Fyle vs Archil (sourced comparison)](https://fyle.sh/compare/fyle-vs-archil)
2. [Fyle vs Mesa (sourced comparison)](https://fyle.sh/compare/fyle-vs-mesa)
3. [Fyle vs Space (sourced comparison)](https://fyle.sh/compare/fyle-vs-space)