What Oraclous doesn't do — on purpose

Most platforms tell you what they do. This page tells you what Oraclous doesn’t — what’s deliberately out of v1, what’s coming when a customer actually needs it, and the lines it will never cross. The disclosure is the point: the guarantee you can verify is the one with its boundaries named.

Citable answer — Oraclous publishes its scope openly: a focused v1 (text, documents, structured data, and code; three model-provider protocol shapes; per-harness task boards), a clear “later, on real demand” list, and permanent commitments it won’t break — never removing human approval for consequential agent actions, never hidden behaviour, never mining customer data. Honesty about limits is treated as a credibility asset, not a gap to hide.

Coming later — when a customer actually needs it

These are things Oraclous should have eventually. They’re named so they’re built on real demand, not speculatively:

• More data modalities. v1 covers text, documents, structured data, code, and time-series. Images, audio, video, and design files come when a customer has the concrete use case — the architecture already allows for them.
• Higher-fidelity export to other runtimes. Today your work exports to the open OHM format, to Claude Desktop and Claude Code, and over MCP. Deeper exporters for specific external agent frameworks arrive on real demand.
• More model providers. v1’s three protocol shapes — Anthropic-native, OpenAI-compatible, and Gemini-compatible — cover the large majority of needs (the OpenAI-compatible shape alone reaches many providers and local runtimes). More vendor-native integrations are added as customers ask.
• Cross-team task boards. v1 gives each Harness its own task board; boards that span an entire workspace are a later read-side addition.
• Pricing model + chargeback tooling. The platform meters usage today (tokens, tool calls, storage, time); the specific cloud pricing model and pre-built internal-chargeback reports come as the offering matures.

Probably never — deliberate non-goals

Not on the roadmap, by design — documented so the answer is unambiguous:

• A drag-and-drop workflow editor. Oraclous is prose-first: you describe the goal and the platform compiles the Harness. A visual node editor would reintroduce exactly the framework-style modelling the platform avoids.
• Real-time collaborative manifest editing. OHM is small and versioned; single-author editing with version control is enough.
• Framework-compatibility layers. Interop is at the protocol level (MCP, OHM, OpenAPI), not by impersonating another framework’s interface.
• A long list of first-party SDKs. The platform speaks REST, MCP, and standard formats; any language with HTTP can integrate.

Never — commitments it won’t break

These are architectural commitments. The answer to “could the platform do X?” is a firm no:

• No unbounded agent autonomy. Consequential changes — an Agent expanding its own access, rewriting its own limits — always route through human approval. There is no setting that removes the gate, and no path to one.
• No hidden behaviour or “platform magic.” Everything is either inspectable open-source code or a capability you own and control. “The platform does it silently” is treated as a bug, not a feature.
• No platform-imposed content moderation. Your policies govern your workspace, not the vendor’s. Oraclous gives you the mechanisms (redaction, human gates, custom skills); it doesn’t impose its own.
• No vendor-accessible credential store. Your credentials live in your own credential broker under per-organisation encryption — even in cloud mode, Oraclous-the-company cannot decrypt them. There is no shared secret pool.
• No cross-customer data sharing or data-mining. Data sovereignty is absolute in both modes. The platform improves through community contributions and feedback — never by reading across organisations.
• No pretending Agents are people. Agents are software that use a model as a resource. “Second mind” is a metaphor for human + AI working together — not a claim that an Agent is a mind or a legal person.

Frequently asked questions

Q: Why publish what you don’t do? A: Because honest scope is how serious teams evaluate infrastructure. Naming the limits up front — what’s deferred, what’s a non-goal, what’s a permanent commitment — lets you decide with full information, and means there are no surprises at adoption or exit. The disclosure is the credibility.

Q: Will Oraclous ever run agents without human approval? A: Not for consequential actions. Bounded learning with human-in-the-loop approval on changes that expand access or rewrite limits is the platform’s safety floor. The most permissive setting lets an Agent propose changes that a human reviews — there is no setting that removes the review.

Q: Does Oraclous train on or aggregate customer data? A: No. Data sovereignty is absolute in both self-hosted and cloud modes. Oraclous-the-company has no code path that reads across organisation boundaries and does not mine customer workflows. The platform improves through open-source contributions and feedback, not customer data.

Q: Is there a visual workflow builder? A: No, by design. Oraclous is prose-first: you write the goal in plain language and the platform compiles a governed Harness you can review. You can edit the resulting OHM manifest directly; a drag-and-drop editor would reintroduce the framework-style modelling the platform deliberately avoids.