Flora Open Project

Open a FLORA project

Hosted MCP generation inputs are plural: call flora_create_generations with { "generations": [{ "workspace_id": "ws_…", "project_id": "prj_…", "type": "image", "prompt": "…" }] } (1–20 items). Put per-generation fields, including optional model, params, and reference_node_ids, inside each item. Read generations[] in the response; retain successful entries' run_id and handle failures individually. Poll flora_list_generations with { "run_ids": ["run_…"] }, even for one run; add technique_id for technique runs. Never retry successful items because another item failed.

Use dedicated tools for this workflow, including batches. execute is deprecated; use it only for an SDK operation without a dedicated tool. SDK examples below describe orchestration: use the corresponding dedicated tools, issue independent calls concurrently, retain every run id, and poll in later calls. Do not choose execute just to combine calls.

The law: the sign-in link is a session, not a URL. It opens once, in a tab the agent controls, and appears nowhere else — not in a reply, not in a log.

Starting point for a session on one canvas. Ends with the canvas in front of the agent (in its own browser tab or via the hosted tools) and a short offer of what to do next, so the user's first real request lands on a canvas the agent has already read.

Input: a workspace id (ws_…) and a project id (prj_…). Output: the canvas open and read, plus starter or continuation ideas.

Steps

1. Get the ids. Both a workspace id and a project id are required before anything else. If the user gave a canvas URL, the project id is the /projects/<id> segment, prefixed prj_. If either is missing, resolve it: flora_list_workspaces, then flora_list_projects with the workspace, and match on the user's wording. Ask if the match is ambiguous; never create a project here.

2. Check for an in-harness browser with WebMCP. Some hosts give the agent its own browser tab and a JavaScript tool (Claude Code Desktop and Claude Cowork do; others may). The test is capability, not host name: can you open a URL in a tab you control and run JavaScript in that page? If yes, take the browser path; if no, take the hosted path. Do not guess — a host that only has a fetch/HTTP tool has no browser.

3. Browser path. Call flora_open_project with the project id and, if the user asked for a specific session length, session_hours (1–24, default 8). Open the returned url in a new tab. It signs you in as the user and lands on the canvas with ?webmcp_embed=1. Rules that follow from that:

   • The link works exactly once and acts as the user for the whole session: never print it, paste it into a reply, or reload it. Navigate within the tab instead, and keep webmcp_embed=1 on the URL.
   • Drive the canvas only through the in-page tools on navigator.modelContext. List them with navigator.modelContext.listTools().map(t => t.name). Calls are async and the JavaScript tool may not await, so stash and poll: window.__r = {}; navigator.modelContext.callTool({name: "flora_get_canvas", arguments: {}}).then(x => window.__r.a = x) then read window.__r.a (structuredContent holds the parsed result).
   • Typical flow: flora_get_canvas → flora_add_nodes → flora_generate (pass expected_project_id) → flora_wait_for_generation. The in-page tools have their own names and inputs; hosted flora_create_generations is not an in-page tool.
   • spend_approval_pending means an approval dialog is open in the tab; only the user can accept it. Ask them, then retry.
   • Do not attach a separate FLORA pane or use look-alike plugin tools (flora_show_canvas, flora_pane_*); the tab is the canvas.

4. Hosted path. Tell the user in one sentence how you are working: through FLORA's hosted MCP, so every change lands in their real project and they can watch it at the project's canvas_url (from flora_get_project). Then use the flora_* tools as usual.

5. Read the canvas. flora_get_canvas for the graph; on the hosted path add flora_list_canvas_nodes for media and asset URLs. Say what is there in a sentence or two — node count, what kinds of media, any obvious workflow.

6. Offer next steps.

   • Blank canvas: offer two or three concrete starters. Draw on what you already know about the user and their work (their brand, product, recent requests) first. If you know nothing about them, call flora_discover_skills without a name and offer the two or three skills that fit best, each as a one-line prompt they could type.
   • Existing content: offer exactly two ways to continue the work that is already there, grounded in what you read — extend a workflow, resize or restyle an existing asset, fill a gap the graph suggests. Name the nodes you mean so the user can confirm.

   Then stop and wait. Do not generate anything before the user picks.

Rules

• Ids first. No canvas read, no link, no ideas until both ids are known.
• The link is a secret. It is a signed-in session as the user. It goes into a browser tab you control and nowhere else.
• Read before proposing. Ideas come from the canvas and the user, not from the tool list.
• Generation spends credits. Quote a cost before running anything the user picks.