A browser download drops halfway and starts again from zero.
Built to resume it from the first missing byte, with a crash-safe journal and segmented transfers.
Products · Pre-alpha
Your AI asks. Zeus fetches.
You decide.
The session-aware download and file layer for AI agents.
Pre-alpha
Phase 0, contracts and skeleton. Most methods return a typed not-implemented error, and no download can run end to end today.
Exists today
- Written specifications (draft 1) and typed API contracts.
- A compiling skeleton: policy table, handshake, approval validation, workflow validator and the tool catalogue are real code with passing tests.
- The specification examples run as tests.
- The browser extension is unit-tested but has never run in a real browser.
Planned, not built
- The daemon run loop, the transfer engine, the browser-extension link and the agent tool servers are not implemented.
- A first public alpha is the next stage. Roadmap phases carry no dates.
- Before any public launch the name needs trademark clearance.
What it is for
Your AI asks. Zeus fetches. You decide.
People who run AI agents that find files but cannot reliably save them: agent builders first, and later anyone who wants an agent to fetch a file with their say-so.
Agent browser tools can open a page but cannot save the file.
Designed to give agents one tool to queue a download and wait for a verified file and its SHA-256.
The file sits behind a login the agent does not have.
Designed so a human approves a narrow session grant: one site, a short time, held in memory, never stored.
The server only talks to real browsers.
Built to ask your own browser to fetch the bytes. It never spoofs a browser and never solves captchas.
You cannot tell later where a file came from.
Designed to keep a hash-chained audit log and a per-file record of source, hash and licence status.
Text from a remote page tries to steer the agent.
Designed to return remote strings as untrusted data, with no agent path that can approve anything.
See the idea
A dropped download, finished with one approval.
This sequence follows the design. Nothing here runs yet: the daemon and the extension are not built. The file name and bind code are examples.
Download rescue · concept, from the specifications
dataset-v3.tar.zst1.20 GiB
Downloading
Step 1 / 6 A browser download of dataset-v3.tar.zst (1.20 GiB) is part way through.
How it works
5 steps, in plain words.
Steps describe the design. Where a step is not built yet, the status section above says so.
- 01
An agent asks
The agent queues a download through a small typed tool. It gets a job, not a blocking call.
- 02
A person approves
If a login is needed, a prompt shows the facts and a short bind code. The choices are once, this site for 24 hours, or never.
- 03
Your browser fetches
A browser extension hands over a short-lived grant, or fetches the bytes itself.
- 04
The engine finishes
Segments are written through a journal, resumed after a failure, then verified.
- 05
A record is kept
The file is filed with a manifest, and the action goes into a tamper-evident log.
Principles
What we hold it to.
- Agents ask. Only a person approves, on a surface an agent cannot reach.
- Completed means verified: hashes and file type are checked before a job is done.
- It never reads your browser cookie database.
- No telemetry, and no model unless you bring one.
- Never charges for speed. It is not a download accelerator.
- States what its audit log does not detect.
Demo available on request
See Zeus with the team that is building it.
Walkthrough of the specifications and the compiling skeleton.
Pre-alpha. Not released. Source is private.