What changed
Remote MCP connections are now natively supported in major agent stacks (FACT: Gemini API added managed background agents + remote MCP; Anthropic-API-compatible local backends like Ollama let Claude Code run against local models), so a network-attached memory tool can plug into both cloud and fully-local Claude Code setups without custom plumbing.
Why now
The Claude Code / coding-agent user base is growing fast and multi-person teams are hitting the 'two drifting company brains' problem now (FACT: Ask HN post describes exactly this). Remote MCP support means the integration surface finally exists. HYPOTHESIS: a 6-12 month window before Anthropic or an OSS memory project ships team-shared memory natively.
Converging signals
(1) Anthropic-API compatibility lets local open models run Claude Code workflows (ollama.com/blog/claude); (2) provider-hosted agent orchestration with remote MCP lowers solo shipping cost (blog.google); (3) a concrete, well-specified team pain: per-machine agent context drifting apart with only a Google Drive workaround (news.ycombinator.com/item?id=48848840).
Customer pain
FACT (from the HN post): two cofounders each run local Claude Code agents; accumulated strategy/customer/decision context is trapped per machine, duplicated work results, the two 'brains' drift, and there is no shared/private boundary or history. Workarounds are hacky (shared folder).
Who pays
HYPOTHESIS: 2-10 person AI-native startups where cofounders live in Claude Code daily; the buyer is a technical founder with a company card. No evidence in the input that anyone currently pays money for this specific thing.
Solved today
FACT (source): hacky workarounds like a Google Drive folder of markdown notes; manually re-explaining context. INFERENCE: some teams commit CLAUDE.md files to git, which handles conventions but not evolving decisions/customer notes or shared/private separation.
Why current solutions are bad
No sync-conflict handling, no history, no shared-vs-private boundary, no structure an agent can query selectively — stated directly in the pain post.
Proposed product
An MCP memory server with a team namespace: agents read/write tagged memory items (decision, customer, convention, strategy); shared items sync via a hosted or self-hosted server, private items stay local; append-only history with attribution and conflict-free merge (CRDT or last-writer-wins per item). Works identically for cloud Claude Code and local-model setups because both speak MCP.
MVP version
Single-binary or Docker MCP server + Postgres/SQLite: store/search/list memory tools, team API keys, shared/private flag, item history. No UI beyond a read-only web viewer. 2-3 weeks of AI-assisted build for this founder.
30-day build
Build MVP; recruit 10 design partners directly from the HN thread and Claude Code Discord/Reddit-adjacent communities; run free while instrumenting what memory operations teams actually use.
60-day build
Add the hosted version (Stripe, $29-49/team/mo), publish to MCP server directories, Show HN launch, open-source the client/local layer to seed adoption (open-core).
90-day revenue plan
Convert design partners to paid hosted plans; target 15-30 paying teams (~$500-1.5k MRR) — realistic given dev-tool willingness-to-pay, not a fast-cash outcome.
Distribution path
MCP server directories/registries, Show HN, Claude Code community channels, the original HN thread itself. Solo-reachable channels, but every competitor uses the same ones.
Pricing hypothesis
Open-core: self-host free; hosted sync $29/team/mo (up to 5 seats), $99/mo for 15 seats. Per-seat pricing is the expansion path.
Technical difficulty
Low-moderate: MCP server scaffolding is well documented; the hard 10% is merge semantics and making retrieval good enough that agents actually use it. Well within founder's AI-workflow strength.
Legal / regulatory risk
Low. Customer data (strategy/customer notes) means standard confidentiality/security expectations, but no regulated data by design.
Platform dependency
HIGH and this is the core problem: the product's substrate is Anthropic's product surface. Anthropic is actively building memory into Claude Code; if they ship team-shared memory natively, the paid wedge evaporates overnight. Mitigation (multi-agent-stack support) dilutes focus.
Founder fit
Mediocre. This is a developer-tools product sold to developers — a crowded arena full of competent builders scratching their own itch, where OSS-for-free is the default expectation. It uses his AI-workflow and fast-prototyping strengths but none of his structural edges (government portals, forced buyers, industrial ops, per-filing monetisation). A lesson (confidence 0.80) says gov-portal mandate shapes fit him best — this is the opposite shape: discretionary purchase, no deadline, no forced buyer.
Breakout potential
Moderate ceiling if it became the standard team-memory layer across agent stacks, but that outcome favors an OSS project or an incumbent, not a solo paid product.
Final recommendation
PASS for this founder (revisit later). The pain is real and the build is easy, but it is a discretionary dev-tool purchase in a space with heavy platform-copy risk, free OSS substitutes, and zero evidence of existing spend — the inverse of his proven forced-buyer/government-portal edge. Only pursue as a fast open-source credibility play (2-3 weeks, seed reputation in the agent ecosystem), not as a primary revenue bet.
Next action
If pursuing despite the pass: reply in the HN thread today offering the tool, and pre-validate by asking 10 Claude-Code-using teams whether they would pay $29/mo — build nothing until at least 3 say yes with a card.