AGENT OASIS 3.0 · OPEN SOURCE FOR macOS
Private decision intelligence for the products you sell, the agents you run, and the money you spend — now with Hermes Fleet as a first-class module for a real agentic-operations install.
Earlier releases synced a Hermes install for a handful of aggregate token counts, tucked behind a button in the Agents tab. 3.0 makes it a first-class module: live kanban shape and its oldest blockers, the open decision queue by authority level with an ack-deadline urgency indicator, roster, gateway process liveness, and structural fleet integrity — all read over one hardened SSH round trip per sync.
Kanban card bodies and decision context, recommendation, and options text — which on a real install carry internal strategic content — are never decoded into the app. The Swift models simply have no property to hold that data, so the decoder drops it before a value exists. Two tests re-encode a parsed model built from a real sensitive-looking payload and assert the sensitive text is verifiably absent, not just that a differently-named field is missing.
This is not Shadowfetch-specific. Any Hermes install works, provided its host answers to ssh <host> and its profile, tools, and state directories match Settings — every path is validated against the same injection-resistant charset already used for the hostname. Read the full contract:docs/HERMES-FLEET.md.
If you run AI agents, someone will eventually ask what they are worth. The honest answer usually contains a few real invoices and several educated guesses — an estimated hourly rate, an estimate of the hours saved, an estimate of the revenue influenced. Most tools add those together and print a single confident figure.
That figure inherits the weakness of its worst input while looking as precise as its best one. Agent Oasis refuses to produce it.
Most tools that price AI work add three guesses to one invoice and print an ROI. Agent Oasis computes them separately and never sums them. Money that moved is one number; capacity you modelled is another. They sit side by side, labelled, and the app will not blend them for you.
An agent whose value inputs were all typed by hand scores zero confidence, no matter how many tasks it completed today. Being busy is not being measured. Staleness can lower the score; nothing can inflate it.
Record a confounder on an experiment and the app will not report a lift. A number a reader cannot separate from seasonality is worse than no number, because it gets repeated without the caveat that was the whole finding.
A new workspace is empty. There is no sample data in the shipped app at all - the generator lives in the test target, so the binary cannot fabricate a record even by accident, and a test fails if that ever changes. Hermes Fleet holds itself to the same rule: there is no fleet-wide “duty success rate” anywhere on a real Hermes install, so the app does not report one.
This section exists because the product argues that unverifiable claims should be labelled. A page about it that hid its own gaps would be arguing against itself.
The workspace key is bound to owner presence, so Touch ID guards the key rather than just the window. The official download is signed and carries the entitlement required for macOS's data protection keychain. A locally built unsigned copy falls back to the older keychain rather than refusing to open your ledger - real protection either way, but not the strongest one.
80 tests, zero warnings, and continuous integration that fails on any Swift warning or a debuggable release build. That is a floor, not a guarantee. Export an encrypted backup and keep the recovery key before you rely on it for anything that matters.
The Hermes Fleet panel reads profiles, tools, and state over SSH from configurable paths. If your install is organised differently it will find nothing for that section - and it says exactly which directory it searched, rather than showing an empty panel with no explanation.
Two hosts appear anywhere in the source, and both are opt-in. There is no analytics, no telemetry, no crash reporting and no update check. You do not have to take that on trust — the repository is public, and one command checks it:
grep -rhoE "https?://[a-zA-Z0-9.-]+" AgentOasis | sort -ucat, list, stats, --version, or systemctl list-units. Nothing on the remote host is started, stopped, written, or deleted.The workspace is encrypted with AES-256-GCM before it is written to disk, at file mode 600, with the key held in your Mac's Keychain. An exported CSV is deliberately plaintext; encrypted backups need the recovery key you stored separately.
Download the signed universal DMG, open it, and drag Agent Oasis into Applications. It requires macOS 14 or later and runs on Intel and Apple silicon. No Xcode, Terminal, account, or API key is required.
The universal DMG is signed with a Developer ID certificate, notarized by Apple and accepted by Gatekeeper. SHA-256:
d3ef31e8bb98d15afe168cd95d0a17ef8cfced59cd05f392852160ef447cfec7brew install xcodegen
git clone https://github.com/ShadowfetchAI/agent-oasis.git
cd agent-oasis && xcodegen generate
open AgentOasis.xcodeprojPress ⌘R. No account, no API key and no Apple team ID required to run it.