First-party chat panel — BYOK LLM chat without a coding-assistant sub

Framework. Game-agnostic. A bring-your-own-key (BYOK) chat panel directly inside the vibesmith dev tooling — for the slice of indie devs who have an LLM API key (Anthropic / OpenAI / Google / OpenAI-compatible local like Ollama or llama.cpp) but no monthly coding-assistant subscription. The framework supplies the substrate — threaded conversations, per-thread provider selection, token meter, streaming render — and routes through the same LLM-call capability the cmd+P quick actions + proactive synthesis use. Not a Claude Code competitor; the on-ramp for people without one.

What this is

A chat panel that lives next to the viewport. Three regions:

• Thread list — every persisted conversation, click to switch, hover to delete. Threads survive across reload via .vibesmith/threads/<id>.jsonl.
• Active thread — message transcript with per-message token meter (input / output / cached); the assistant response streams in as the LLM emits it; newly-arrived proactive tips surface beneath the response as ↳ tip: lines.
• Composer — textarea + ⌘↩ to send; per-thread provider + model selector in the header; a settings cog opens the provider-config modal.

Why a first-party panel, not “use Claude Code”

The framework’s MCP server + the vibesmith mcp install cookbook already cover users with a paid coding-assistant sub (Claude Code / Codex / Copilot drive the framework via MCP). But a meaningful slice of indie devs have:

• An Anthropic / OpenAI / Google API key (or a local Ollama)
• No monthly coding-assistant sub

For that audience, “go buy a sub” is friction that loses them. The first-party panel uses the key they already have, in a UI that already knows the project’s canon.

Explicitly lighter than full coding assistants

The panel is not trying to match Claude Code / Cursor / Copilot feature-for-feature. It’s an on-ramp:

• No agentic loops, no multi-file refactor superpowers, no per-line completions
• Strong at canon-aware chat, MCP tool invocation, scenario authoring, recipe / provider discovery — the things the framework has a unique edge on
• Users who outgrow the panel migrate to a paid harness; exported conversation history travels with them

Configuring providers

Open the chat panel and click the cog in the header. The settings modal lets you add / edit / delete provider entries. Each entry has:

Field	Notes
Kind	anthropic, openai-compatible, or gemini. The picker shows a one-line hint per kind.
Id	Stable identifier (letters / digits / _ / -). Immutable after creation — used by the per-thread provider selector + future routing chains.
Label	Friendly name for the dropdown. Falls back to the id.
API key	Your provider key. Stored in this browser’s localStorage; never sent anywhere except the provider. Reveal toggle on the field.
Base URL	Required for openai-compatible (https://openrouter.ai/api/v1, http://localhost:11434/v1, etc.).
Default model	Optional. Falls back to the per-kind default (e.g. claude-haiku-4-5, gemini-2.0-flash, gpt-4o-mini).

Saves take effect immediately — no reload required. The panel’s provider selector populates from the same registry.

Today: localStorage. Tomorrow: ~/.config/vibesmith/config.toml

The localStorage path lands end-to-end today. The long-term shape — [llm.providers.*] in ~/.config/vibesmith/config.toml with api_key_env indirection — comes in a follow-up alongside the Tauri-side file-read + permissions plumbing.

Per-thread provider override

Each thread carries its own providerId + model. Change either via the header dropdown / text input on an active thread; subsequent turns dispatch against the new selection without affecting other threads. Leave the provider blank to fall back to the chain’s default routing.

Token meter, no dollar conversion

Every assistant message displays its token usage inline:

in 1248 · out 372 · cached 980

The meter is token-only. vibesmith does not convert to dollars because subscription assistants don’t publish $/tok rates honestly and don’t formally disclose plan allowances — any conversion would be misleading. The meter exists so you can see what each turn cost, in the only currency the framework can measure without lying about.

Streaming render

Long responses appear as the LLM produces them — a preview row with a pulsing caret renders the partial text. When the stream completes, the placeholder clears and the persisted message lands in the thread.

Inline ↳ tip: rendering

When you send a turn, the panel captures the current timestamp. Any proactive tip the framework’s proactive-advice queue emits during the turn surfaces under the assistant response as an ↳ tip: line. Dismissed tips don’t appear (the queue already filters them).

The framework never synthesises the tip’s prose — it surfaces the structured candidate (category, summary). To get LLM-rendered prose for a tip, click Synthesize advice in the proactive-tips panel; that dispatches the framework’s synthesize-advice task through the same LLM-call chain the chat panel uses.

When to use the chat panel vs cmd+P quick actions

Surface	Best for
Chat panel	Multi-turn conversation; back-and-forth refinement; canon-aware Q&A; one-off scripted authoring tasks
Cmd+P quick action	One-shot context-aware prompt that already has a known shape — “summarize this scene”, “describe this asset”, “lookup pattern”
External assistant via MCP	Heavy lifting: multi-file refactors, agentic loops, full coding-assistant superpowers

The three surfaces share the LLM-call chain, the task-context contract, and the MCP Tier-1 tool surface. Switching between them costs nothing.

Persistence model

Each thread is one append-only JSONL file at .vibesmith/threads/<thread-id>.jsonl:

{"kind":"header","id":"thr_…","title":"Tune the tavern lighting","providerId":"anthropic","model":"claude-haiku-4-5","createdAt":"2026-05-18T…"}
{"kind":"message","message":{"id":"msg_…","role":"user","text":"how do I add a directional light?","createdAt":"…","taskId":"lookup-pattern","stateRef":{"kind":"panel"}}}
{"kind":"message","message":{"id":"msg_…","role":"assistant","text":"…","createdAt":"…","tokens":{"input":1248,"output":372,"cacheRead":980}}}

The header is line 0; subsequent lines are messages. Partial writes never corrupt the whole thread. Malformed lines are skipped on load.

Persistence is opt-in — it requires the platform-bridge to be wired with fileSystem capability (Tauri host). In the zero-install browser host, the panel still works; threads just don’t survive a refresh.

Privacy + key handling

• API keys live in this browser’s localStorage only. No framework code sends them anywhere except the provider’s API endpoint.
• No telemetry on chat content. The framework doesn’t read, log, or transmit your messages.
• No prompt persistence by us. The JSONL files are yours — under your project directory, never synced anywhere unless you choose to commit them.

Related

• AI assistant — the four-tier interaction model this panel completes.
• MCP tiered surface — the Tier-1 tools the panel + external assistants share.
• Task context — every chat turn is a task invocation against the shared context contract.
• Install MCP into your coding assistant — if you have a paid harness, use that.