Cycle on Sep 17, 2026

I'll start this research cycle by reviewing the open threads and existing notes for this topic, so I can build on prior findings rather than start over.

I have a rich existing note set. The two open threads most relevant to this cycle's brief are the **destination-vs-backend question** (does MCP-first distribution commoditize research platforms — core to Hey, Lefty's positioning) and the **Scholar Labs GA/premium watch**. The HN map also has Panel (already tracked) and the Goedecke essay (already tracked). Let me drill into the Panel story first — it's the direct bottom-up signal in Hey, Lefty's layer — and check the Scholar Labs watch in parallel.

Scholar Labs is still free/experimental per Katina Magazine's 2026 library-market review — no GA or premium tier, so that watch hasn't fired. The Panel story is rich: local-first dockable workspace, agent-built panes, a typed "Module" protocol, Claude Code as the only harness — and the top user asks are exactly the destination-vs-backend tension (persistence of agent-built layouts; "can I add it with an MCP server?"). Now let me gather evidence for the destination-vs-backend question thread: are OpenAI/Anthropic shipping native research workspaces, and is there usage/adoption signal on MCP-called vs destination research tools?

Strong leads. The Scholar Labs watch stays open (still free, no GA). For the destination-vs-backend question, I now have three concrete evidence threads: OpenAI absorbing deep research into ChatGPT Enterprise/Projects/Canvas, specialist tools (Elicit/Consensus/Scite/Undermind/CASRAI) exposing MCP servers, and Panel's users asking for MCP-first integration. Let me fetch the two substantive market-map sources to get verbatim material.

My next step: I still need to verify the CASRAI pricing/MCP claim (I only have a search snippet so far — I won't cite it unverified), then write the two findings, resolve the question thread, register the tripwire watch, and close the cycle with `finish_cycle`. First, the CASRAI fetch:

The CASRAI page is verified and rich — a five-class procurement taxonomy, quantified recall data (Elicit ~39–40% sensitivity vs ~94.5% for traditional searches), and citation-trust rules by class. I now have everything needed. Writing the two findings:

The call failed — I need to pass the parameters explicitly. Retrying note 1:

Note 1 saved. Now note 2 — the pricing/segmentation snapshot with verified capability limits:

Both findings are saved. Now resolving the destination-vs-backend question thread and registering the tripwire watch that emerged from it:

Thread resolved and tripwire registered. Closing the cycle:

Cycle complete. Here's what this research cycle produced for the AI research-tools market map: ## Findings written **1. [[destination-vs-backend-research-layer]]** — resolves the open question thread on whether MCP-first distribution commoditizes research platforms: - **Hosts are absorbing the casual destination layer**: ChatGPT Projects, Claude Projects ("a long-running research workspace"), and Perplexity Spaces all ship workspace persistence inside $20/mo subscriptions. - **Specialists are becoming MCP backends**: Cypris exposes its R&D intelligence layer via MCP server and launched Cypris Q for Microsoft Copilot (Aug 2026) — "the general-purpose model supplies reasoning and interface, while the domain layer supplies the standardized method and the grounded corpus." - **Bottom-up destinations are harness shells**: Panel (Show HN) is local-first and MIT-licensed but runs on Claude Code only — and its first users immediately ask for it to be MCP-callable *from* their own agents. Both directions of MCP-nativity are table stakes. - **Positioning language for Hey, Lefty**: sell control/reproducibility/audit, not autonomy ("The difference between them is not sophistication. It is control"), with the enterprise wedge being the governance gap (only 21% of orgs have mature agent governance; deployment pulled back 42%→26% in Q4 per KPMG). **2. [[research-tool-pricing-segmentation-2026]]** — pricing and buyer snapshot: - Individual tier is a **$12–20/mo commodity band** (Elicit $12, Consensus $15, Phind $17, Perplexity/ChatGPT/Claude $20); enterprise is custom-quoted and compliance-gated. - CASRAI's five-class procurement taxonomy plus **quantified capability limits**: Elicit ~39–40% recall sensitivity vs ~94.5% for traditional searches; extraction quotes reproduced only ~46% of the time. Grounding reliability — not model quality — is the buyer's selection criterion, and NIH/NSF/Cochrane disclosure rules define a compliance feature set none of the consumer tools ship natively. ## Thread & watch housekeeping - **Resolved**: the destination-vs-backend question thread. - **Checked, not fired**: Scholar Labs remains free/experimental — no GA or premium tier yet, so that watch stays open. - **New tripwire registered**: a host platform (OpenAI/Anthropic/Google) shipping a full native research workspace (files + notebooks + citation library beyond chat Projects) — the main compression risk to Hey, Lefty's layer from above.