Per-project Claude agent launcher (Pickism × Claude integration)

Pickism × Claude integration: per-project agent launcher

Vision

A button on a Pickism project ("Open in Claude" / "Launch Concierge") opens a Claude conversation pre-seeded with:

1. The Travel Concierge agent system prompt (portable template)
2. The user's profile (Yan-style preferences layer)
3. The specific project's full context (project + options + siblings)
4. The Pickism MCP already connected

Effectively: one click → kickstarted agent ready to go. Replaces the manual "open Travel Concierge project → new chat → type 'kickstart Prague'" flow with zero ceremony.

Why

• Today's flow (Claude Project + manual kickstart) works but has friction: user must remember to create a project, paste instructions, attach city knowledge, and type the kickstart command each time.
• Pickism already has 100% of the structured trip context. Forcing the agent to re-discover it via MCP calls every chat is wasteful — we could hand it the bundle on launch.
• A per-project launcher is the natural Pickism feature: "this project has an AI concierge attached to it." Maps to how users already think about their trips.

Probable implementation paths (pick one in design)

Path 1: Claude.ai deep link with prefilled message

• Pickism builds a URL: https://claude.ai/new?q={prefilled_message}&project={travel_concierge_project_id}
• Prefilled message = the kickstart prompt + serialized project context
• Pros: zero backend work on our side, uses existing Claude Projects
• Cons: requires user to have the Travel Concierge project already created; URL length limits on the prefilled message; doesn't auto-attach MCP

Path 2: Anthropic API + custom chat UI inside Pickism

• Pickism hosts a chat surface, calls the Anthropic API directly
• We control system prompt, can inject project context as initial messages, can wire MCP server connections per session
• Pros: full control, no Claude.ai project required, can be embedded inline ("chat with your trip"), works on mobile web without app-switching
• Cons: real backend work (auth, billing, message storage), need to handle streaming, file uploads, photos
• Subpath 2a: build minimal MVP — text-only, no photos, no file uploads
• Subpath 2b: full parity with Claude app (photos, voice, files)

Path 3: MCP-prompt bundle (anthropic-side feature, may not exist yet)

• If/when Anthropic ships a "launch with pre-filled context" mechanism for MCP-aware clients, this becomes trivial
• Worth tracking but not blocking on

Open product questions

1. Per-trip vs per-user agent? Each project gets its own conversation, or one rolling conversation per user across all trips? Per-trip is cleaner mentally but loses cross-trip learning.
2. Read-only or write-back? Should the agent's recommendations (locked plans, added activities) write back to the Pickism project automatically, or only on explicit user request? Today's design is explicit-only — keep that.
3. Multi-user / collaborative? If a Pickism project has multiple members, do they share one conversation or each have their own? Trip companions seeing each other's "where should I eat" chats has obvious privacy concerns.
4. Pricing model? Anthropic API usage isn't free. Charge users, eat the cost, or require user to bring their own API key?
5. Mobile-first vs desktop? The use case is overwhelmingly mobile (user is in-city, on the move). Whatever surface we ship has to work one-handed on a phone.

Prerequisites / dependencies

• Pickism MCP server is already shipped and stable (✅ done, v1.12.0)
• Travel Concierge agent prompt is stabilized and battle-tested (✅ in progress, Vienna trip was first real test, retro complete)
• User profile schema decided — what fields belong in the user-level profile vs the per-project context (partially answered by current agent template's {{USER_PROFILE}} slot)
• City knowledge handling — accumulated as project knowledge in Claude Projects today, would need a different surface in Path 2

Suggested first slice (if/when prioritized)

Path 1 with a single hardcoded Travel Concierge project ID. Pickism adds "Open in Claude" button on any project with start_date set. Button opens claude.ai/new with a prefilled kickstart message including the project slug/ID. Agent (already configured in the Claude Project) reads project via MCP and runs kickstart.

Smallest possible step that proves the concept. ~half a day of work.

Status

Backlog — no commitment. Documenting because we just designed the manual Option A flow and Option C is the obvious end-state, but not the priority right now. File for future product planning.