What we're thinking about, what we're not.

Inbound asks we think have merit. Not yet scoped or scheduled — read these as 'we're listening, not promising.'

• Native MCP plug-in marketplace inside the workspace

  Under consideration

  Browse, install, and pin third-party MCP servers from a marketplace pane — without leaving the workspace shell. Today MCPs are wired at the harness level; an in-app surface would let non-engineers add tools.

  Scoping the trust + safety model first (what runs in-tenant vs. proxied).

• Cold-start free tier without credit card

  Under consideration

  A workspace that boots with a $0 model budget, capped output, and read-only integrations. Today /pricing requires a card on file; this would lower the activation barrier for evaluators.

  Need a credible abuse-defense pattern before turning the spigot on.

• Multi-region data residency selector at workspace creation

  Under consideration

  EU / US / APAC residency choice baked into the create-workspace flow, with downstream storage (Postgres, GCS, vector store) honoring the selection per-tenant.

  Today Cloud SQL is US-region. EU residency would unblock a handful of enterprise asks.

• Per-domain dashboard customization API

  Under consideration

  Let workspace admins re-arrange or hide /admin/* sub-dashboards (revenue, ops, agent-evals) without forking the codebase. Useful for tenants who only use 2–3 domains and want a cleaner shell.

• Public Mastodon / ActivityPub outbound for the social vertical

  Under consideration

  The /social DomainSpec is intentionally empty pending an OSS leader pick. Mastodon + ActivityPub is the leading candidate (Apache-2.0 friendly, federated, no API-key gate), but adoption shape is still under debate.

  Held until we see signal that customers actually want federated posting, not just centralized scheduling.

Concrete work-in-flight that hasn't shipped: spikes, prototypes, proofs-of-concept. Outcome is uncertain by design.

• Slack-native and Discord-native agent runners

  Researching

  Dispatch the master agent and read step events without leaving a channel. Prototype exists for Slack via slash commands; SSE-to-Slack-thread bridging is the open question.

• BYO-LLM via Bedrock and Vertex passthrough

  Researching

  Today the provider router uses Anthropic / OpenAI / Google / Groq direct keys. Customers with existing Bedrock or Vertex contracts want their spend to flow through their cloud — research is on cost-attribution + latency tradeoffs.

• Compliance template gallery (SOC 2, HIPAA, GDPR starter kits)

  Researching

  Pre-wired policy docs, audit-log retention, redaction rules, and a smoke-test pack per framework. Spike landed on the policy DSL; the gallery shell is still on the bench.

• On-device inference for the desktop client (Apple Silicon)

  Researching

  Run a local 4–7B model for offline-mode autocomplete and quick-question intents. Investigating which routes degrade gracefully when no network is present.

• Tahoe-100M ingest pipeline for healthcare DiscoveryOS

  Researching

  /healthcare/discoveryos and the 6 sister sub-OS surfaces ship with 9 empty Prisma tables (tahoe_genes, tahoe_drugs, tahoe_cell_lines, sample_metadata, expression_stats, drug_response_signatures, biomarker_candidates, resistance_signatures, combination_candidates). APIs return typed-503 nextAction:ingest_tahoe until the Tahoe-100M ingest job lands and Tahoe-x1 inference is wired.

  Schema is live; the gap is the data pipeline + the inference adapter.

• Activate /admin/revenue action endpoints (GitHub-issue, email, helpdesk, Stripe)

  Researching

  The Command Center alerts panel and customer-row actions ship with real Prisma audit trail (RevenueAlert), but the outbound side returns typed-503 until GITHUB_TOKEN, RESEND_API_KEY, HELPDESK_API_*, and Stripe upgrade-event credentials are wired. cost-optimizer.ts is the back-end the AI-margin tab calls — also operator-gated until the keys land.

• SDLC worker activation: GitHub App + GCP Workload Identity Federation

  Researching

  8 GitHub workflows + /admin/founder dashboard + 3 backing APIs are in the repo, but the loop only fires once a dedicated GitHub App with least-privilege scopes (no administration:*, no workflows:write) is provisioned, WIF is configured for GCP (no JSON keys), and the new Prisma models (MergeAuthorization, CanaryPromotion) are pushed.

  Deliberately operator-gated — the GitHub App + cloud creds are sensitive enough that the build couldn't auto-create them.

• Activate Luma Dream Machine end-to-end for paying users

  Researching

  Full Luma integration is in code (image / video / reframe / modify / character_ref / async lifecycle, 6 routes, 6 agent tools, MediaGenerationJob ledger), but generation returns typed-503 until LUMA_API_KEY is set and prisma migrate dev --name luma_media_generation_jobs is run.

• Install ECC + Superpowers as Claude Code plug-ins (not just vendored files)

  Researching

  48 ECC agents, 139 skills, 65 commands, 37 rule docs, and the Superpowers skill pack are vendored on disk under packages/agent-skills/skills/ and .claude/, but the harness loads them as documentation only until the operator runs /plugin install ecc@ecc and /plugin install superpowers@superpowers. Optional context7, github, and exa MCP keys would unlock the rest.

Things we want to build eventually. No timeline, no commitment, but worth keeping visible so the team and customers can shape them.

• White-label tenant theming (logo, palette, copy)

  Wishlist

  Re-skin the workspace shell for resellers and agencies. Lower priority because the current pure-white theme is intentional brand equity — but customer demand is non-zero.

• iPad pencil surface for the design + intelligence panels

  Wishlist

  Sketch a wireframe → agent generates the React component; mark up a chart → agent writes the SQL. Wishlist because it requires a dedicated touch-first redesign of the canvas layer.

• Public agent benchmarks page

  Wishlist

  Side-by-side scores of the neww master dispatcher vs. raw frontier models on the same task suite. Wishlist because we need internal eval discipline (more trials, clean separation of train/test) before we publish.

Asks we deliberately turned down. We keep these visible so people can challenge the reasoning instead of re-asking in private.

• Lifetime-deal pricing (one-time payment, unlimited usage)

  Declined — with reason

  We get this ask weekly. Declining because per-call AI inference is a recurring variable cost — an unlimited-usage lifetime SKU would either bankrupt the product or be quietly throttled into uselessness.

  We'd rather hold prices low than promise economics we can't sustain.

• Browser extension that scrapes any site as a 'real' integration

  Declined — with reason

  A real integration means a maintained adapter with auth, rate limits, and an SLA. A scraper extension fails silently when sites change markup, and we won't ship a feature whose first failure mode is invisible.

  Bright Data fabric (apps/web/src/lib/web-data) is the supported path for web data.

• Anonymous agent runs (no workspace, no session)

  Declined — with reason

  We tried this in an earlier prototype. Without identity there's no rate-limit ceiling, no abuse boundary, and no way to honour deletion requests. The demo experience needs a session even if it's a throwaway one.