What changed
FACT: Remote MCP tool connections are now natively supported in major agent stacks (Gemini managed agents announcement; MCP is Claude Code's native tool protocol), and Anthropic-API-compatible backends (Ollama) show the Claude Code client layer is open enough to point at third-party infrastructure. FACT: A concrete Ask HN post documents two cofounders whose per-machine Claude Code agent context is drifting with no shared/private boundary and only hacky workarounds (similarity 0.837 to this convergence).
Why now
Claude Code team adoption is growing faster than Anthropic's team-memory features; MCP means a third party can inject a memory tool into the agent loop with one config line, no forking or plugins. The window is now because this gap is obvious and will be filled — by Anthropic, an open-source project, or a fast solo builder — within quarters, not years.
Converging signals
(1) Documented team pain: shared agent context across machines with private/shared boundary (HN 48848840). (2) Remote MCP as a provider-supported integration surface (Google blog). (3) Anthropic-compatible local backends proving the client is pluggable (Ollama blog). Together: the exact users complaining can install the fix in one line.
Customer pain
FACT (from HN source): two cofounders each accumulate strategy, customer notes, and decisions in separate local Claude Code agents; context is trapped per-machine, the two 'company brains' drift, one founder re-derives what the other already worked out, and there is no shared/private separation, sync-conflict handling, or history. Current workaround cited: a shared Google Drive folder.
Who pays
HYPOTHESIS: 2-15 person AI-native startups where multiple people run Claude Code daily — the same demographic already paying $100-200/user/month for Claude Max/Team plans, so a $30-100/month team memory add-on is small relative to existing spend. No direct evidence of payment for this specific product exists in the input.
Solved today
FACT (source): shared Google Drive/notes folders and manually pasted CLAUDE.md files. INFERENCE: also git-committed CLAUDE.md conventions, and open-source memory layers (mem0/OpenMemory, Zep, Letta) that exist but require self-hosting and aren't team-boundary-aware out of the box.
Why current solutions are bad
File-sync approaches have no merge semantics (drift and conflicts, as the HN poster states), no private-vs-shared boundary, no versioned history, and no retrieval — the agent can't query them semantically. Self-hosted OSS memory servers demand infra work the complaining users explicitly don't want.
Proposed product
Hosted multi-tenant MCP memory server: team namespace of shared memories (decisions, conventions, customer notes) + per-user private overlay, append-only versioned history with attribution ('who taught the agent this'), semantic retrieval, and conflict-surfacing when two agents write contradictory facts. Install = one `claude mcp add` line per teammate. SOC2-lite posture: encryption, export, delete.
MVP version
Single-region FastAPI/Postgres+pgvector service exposing MCP tools (remember/recall/list/forget) with team API keys, shared vs private scopes, and a plain-text audit log. No UI beyond a minimal web console. 2-4 weeks of AI-assisted build — this is squarely in the founder's demonstrated stack (Python, FastAPI, Postgres, MCP-adjacent bridge work in his own Convergence Engine).
30-day build
Ship MVP; answer the HN thread and every similar thread with a working demo; publish to MCP server directories/awesome lists; recruit 10 design-partner teams free.
60-day build
Add versioned history, conflict surfacing, and org SSO-lite (Google login); convert design partners to $30-50/mo team plans; write 'how two cofounders share one company brain' content targeting Claude Code searches.
90-day revenue plan
Target 20-40 paying teams at $30-100/mo = $1-3k MRR by day 90-120. Modest but real; expansion via per-seat pricing as teams grow.
Distribution path
HN/Show HN (the pain thread is literally there), MCP server directories, GitHub (open-core: OSS single-user server, paid hosted team tier), Claude Code subreddit/Discord, SEO on 'share Claude Code context with team'. Demonstrated-value selling — fits founder's style; no relationship sales needed.
Pricing hypothesis
Free solo tier (OSS/self-host), $30/mo per team up to 5 seats, $10/additional seat. Per-team flat beats per-GB for trust; usage costs are trivial (text + embeddings).
Technical difficulty
Low-moderate for the founder: MCP server + Postgres/pgvector + auth. Hard parts are merge/conflict semantics and retrieval quality, both iterable post-launch. He has already built an MCP-adjacent Claude bridge and a Postgres+embeddings pipeline in production.
Legal / regulatory risk
Low but non-trivial: the product stores customers' strategy and customer data — needs a real DPA, encryption at rest, and deletion guarantees. No regulated data required. Not a blocker.
Platform dependency
HIGH and this is the core risk: Anthropic could ship native team/cloud memory for Claude Code and erase the wedge overnight. MCP portability (works with Gemini managed agents, Cursor, etc.) is the hedge — position as cross-agent team memory, not Claude-only.
Founder fit
Mixed. Fits his micro-SaaS/API/agent preference, his exact tech stack, and demonstrated-value distribution. Does NOT fit his strongest proven edge (government-mandate forced-buyer filing tools — lesson conf 0.80 says those score 8-9 fit); this is a competitive dev-tools space where he'd be one of several fast movers, selling to discretionary buyers, not forced ones.
Breakout potential
Moderate: if cross-agent team memory becomes a category, an early neutral player could be the 'shared brain' layer for all agent runtimes. But the same dynamic attracts VC-funded competitors (mem0 et al. already funded).
Final recommendation
CONDITIONAL NO as a primary bet, YES as a fast side wedge. The pain is real and well-specified, the build is 2-4 weeks in the founder's exact stack, and distribution is cheap — but demand evidence is a single thread, the competitive gap is thin, and platform risk (Anthropic native team memory) is the highest of any idea shape. Worth a 30-day timeboxed build ONLY if the founder wants a dev-tools property; it should not displace a government-mandate forced-buyer opportunity, which remains his structurally better shape (lesson, conf 0.80). Note: the engine's known demand-blindness (lesson, conf 0.85) means team-agent-memory demand may be under-represented here, but I will not invent spend that isn't in evidence.
Next action
Spend one day, not thirty: reply to HN 48848840 and 3-5 similar threads offering a hosted beta, and put up a landing page with a waitlist + $30/mo preorder. If <15 signups or 0 preorders in 14 days, kill; if strong, do the 2-4 week MVP.