Skip to content

docs(design): single UMIDE Agent (Claude, API-key) — Phase 2 proposal - #84

Open
dev-josias wants to merge 3 commits into
mainfrom
docs/single-agent-spec
Open

docs(design): single UMIDE Agent (Claude, API-key) — Phase 2 proposal#84
dev-josias wants to merge 3 commits into
mainfrom
docs/single-agent-spec

Conversation

@dev-josias

Copy link
Copy Markdown
Contributor

Proposal — not for merge until we agree on direction. Spec for collapsing the multi-backend agent picker into one "UMIDE Agent."

Full doc: docs/design/single-agent.md

The headline (a compliance finding that changes the premise)

I researched the Claude Agent SDK before speccing. The "power our agent with the user's Claude subscription" path I hoped for is currently not available to third parties — Anthropic's docs say they don't allow third-party apps to offer claude.ai login/rate-limits (the June-15 Agent-SDK subscription-credit change is paused). The Agent SDK just spawns the claude CLI, there's no official Rust SDK, and API-key auth is the sanctioned path.

So the single agent is Claude via the claude CLI (the Agent SDK mechanism), authenticated with the user's Anthropic API key — shaped into a mobile agent via the system prompt + device MCP + per-action approvals UMIDE already has for its Claude Code path.

⚠ Affects the shipped beta

UMIDE advertises "No API key needed — use your Claude login." That's the pattern the policy prohibits → fast-follow to remove that messaging (app + umide.dev + README).

What changes

  • Keep: the Claude/claude-CLI path (prompt + device MCP + approval bridge + sessions).
  • Remove: Codex, Gemini-CLI, and the multi-provider direct-API path (OpenAI/DeepSeek/Gemini); the AssistantBackend picker; the key-free/login UX.
  • One decision to make: drive the claude CLI (Path A, recommended — reuse + mature loop) vs UMIDE's own loop on the raw API (Path B — more control, more build).

This is the foundation for the Phase-2 mobile closed loop (which rides as system-prompt + device tools on this one agent).

…ent SDK mechanism

Proposal for Phase 2: replace the multi-backend picker (Claude/OpenAI/DeepSeek/
Gemini + Codex/Gemini-CLI) with one 'UMIDE Agent' — Claude via the claude CLI
(the Agent SDK's own mechanism), authenticated with the user's Anthropic API key,
shaped into a mobile agent via system prompt + device MCP + per-action approvals.

Includes the compliance basis: Anthropic currently does NOT allow third parties
to offer claude.ai subscription login (the June-15 Agent-SDK subscription-credit
change is paused), so the single agent is API-key-based. Flags a fast-follow to
remove the beta's 'no API key needed — use your Claude login' messaging.

Not for merge yet — pending direction sign-off.
…tchable

Per design discussion: 'one interface' means one CHAT surface, not one provider.
The agent stays provider-agnostic (Claude/OpenAI/Gemini/DeepSeek by API key) on
UMIDE's OWN loop (LlmRunner) — vendor CLIs dropped. The active provider is a
compact, always-available switch (not a per-chat menu), flippable any day;
switching applies to the next turn. All BYO-API-key (subscription login isn't a
clean third-party path for Anthropic or OpenAI). Own-loop also gives Phase-2 the
loop control it needs.
…rs + Auto + override)

Per discussion: v1 ships Claude + OpenAI (Gemini/DeepSeek follow). Add first-class
model tiering so simple requests use cheap models (Haiku / mini) and only agentic
turns hit the deep model (Opus 4.8) — never default to the flagship. Auto router
by default, manual per-session override in the header, per-turn model label for
transparency. Goal: don't burn the user's credits.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant