8.4 Natural Language-First & Consent Tiers
2026-09-25 · Finished in development (phase one: skills, industry solutions, MCP connectors), not yet in a release. Implementation notes: Session install, phase one.
Core Principle
Natural language drives everything. Ask for permission only when it matters.
This means in your session, you can use natural language, GitHub links, or market resource links (https://market.teloa.ai/<entry>/) to search, add and install skills, and to add industry solutions and MCP connectors (phase one shipped in development). No form navigation or multi-step flows — just say what you want in conversation.
Consent Tier Framework
Different operations require different levels of confirmation:
| Operation | Confirmation | Details |
|---|---|---|
| Browse, search, preview | No | Market content, resource search, detail views |
| Content-only skills | Once | One card covers "add and install": source, license, file manifest, fingerprint |
| Network/script skills | Once | List declared websites and tools from the skill text; mark data egress |
| MCP connectors | Once | List provided tools (read-only / write) and accessed hosts; local-process recipes name the npm package and version to be installed |
| Code execution plugins | Strictest | Most detailed confirmation and permission checks (no session entry point in this phase) |
| Uninstall, delete, change permissions | Yes | All modifications require confirmation. Available in phase one: disable a skill on the capabilities page, delete a connection under Market → Connections, unload a solution from the space; uninstalling skills and deleting content library entries come in phase two |
Special Rules
Keys and Account Authorization
- Never paste API keys, accounts, or tokens in chat
- The session only registers the connection; actual token entry happens in connection settings or the OAuth page
- Any parameter that looks like a secret (common token prefixes or high-entropy strings) is rejected, and the rejection never echoes the value
- This protects credentials from appearing in conversation logs
AI Colleague Limitations
- Phase one: AI colleagues working in tasks (including twins, group chats and sub-agents) cannot search, add, install or connect; they can only say what they need in their reply
- Request records flowing into "Needs you" for the owner to handle come in phase two
- Confirmation and execution authority stays with you
Installation Audit Trail
- Every add, install, connect and load records the pinned version or fingerprint, who confirmed it, when, and a stable request ID
- Undo entry points are only the ones that really exist (see the "Uninstall, delete, change permissions" row above)
High-Risk Operations Get Special Treatment
- Payments, API keys, security settings, permanent deletion
- Even initiated via natural language, completion happens in dedicated UI
- Never fully completable in conversation alone
Why This Design
This approach balances two goals:
- Convenience: Common operations (installing a skill, adding a solution) flow smoothly in conversation without context switching
- Safety: Important decisions (authorizing keys, deleting data) always require explicit human confirmation to prevent accidents or misuse
Consent tiers let your AI assistant help you decide quickly, while important choices stay in your hands.