What changed
OpenAI now ships multi-hour deliverable agents (FACT: openai.com announcement) and Google now hosts background agent tasks with remote MCP tool connections natively in the Gemini API (FACT: blog.google announcement). Real work now executes unattended against remote tool servers, and remote MCP creates a standard, universal insertion point where a third-party proxy can sit between agent and tools.
Why now
Before provider-hosted background agents, oversight tooling had to be embedded in each team's bespoke agent loop. Remote MCP standardizes the wire where tool calls flow, so one proxy works across providers. The window is now because the primitives are weeks old; the risk is that OpenAI/Google/Anthropic absorb approval-gating natively within 6-12 months (HYPOTHESIS).
Converging signals
(1) Gemini API managed agents with background tasks + remote MCP; (2) ChatGPT multi-hour cross-app deliverable agents; (3) HN user explicitly asking for frameworks/tools to regain oversight of agents (PAIN signal). Capability supply and articulated anxiety are converging; paid demand is NOT yet evidenced in the input.
Customer pain
Fear of unattended agents taking irreversible actions (sending, deleting, spending, writing to production) with no approval step, no audit trail, and no way to reconstruct what happened. Supported by exactly one HN complaint in the signals (FACT for that user; generalization is HYPOTHESIS).
Who pays
HYPOTHESIS: solo devs and 2-20-person teams shipping agent features on OpenAI/Gemini who face a customer or internal requirement for 'human-in-the-loop on risky actions' and an audit log. Secondary: agencies deploying agents into client businesses that demand accountability. demand_evidence array is EMPTY — no hiring/spend or forced-buyer evidence was provided, so willingness to pay is unproven.
Solved today
Teams hand-roll confirmation steps in their agent code, use framework-level interrupts (LangGraph human-in-the-loop), or adopt existing tools: HumanLayer (approval-gating for agents), Langfuse/LangSmith (tracing/observability), or simply run agents in low-stakes sandboxes. (Competitor existence is FACT from general knowledge; their exact coverage of remote-MCP background agents is HYPOTHESIS.)
Why current solutions are bad
Hand-rolled gates don't survive the shift to provider-hosted execution where the loop isn't your code anymore; observability tools trace but don't BLOCK; framework interrupts are framework-locked. A protocol-level MCP proxy is framework- and provider-agnostic. That's a real architectural wedge, but it is thin and copyable.
Proposed product
A hosted (or self-hostable) MCP proxy: point your agent's remote MCP URL at the proxy instead of the tool server. Policy file marks tools/arguments as auto-allow, notify, or require-approval; approvals arrive via Slack/email/web with full context; every request/response is logged immutably; any run can be replayed step-by-step or re-executed against recorded responses for debugging. Per-seat + per-1k-gated-calls pricing.
MVP version
A single-binary/hosted MCP passthrough proxy supporting: (1) allow/deny/approve policy per tool, (2) Slack-webhook approval with timeout behavior, (3) JSONL append-only log of all tool traffic, (4) basic web viewer with replay of a recorded run. Skip SSO, RBAC, and fancy policy language. ~3-4 weeks solo with AI-assisted build (HYPOTHESIS).
30-day build
Build MVP against Gemini remote MCP + one OpenAI agent flow; open-source the proxy core (distribution engine); publish 'I put an approval gate on a background agent' demo posts to HN/r/LocalLLaMA/X; list in MCP server directories; get 20 design-partner installs.
60-day build
Ship hosted version with team approvals + retention; instrument which tools users gate most (pricing signal); convert 5 design partners to $29-99/mo paid pilots; publish replay-a-failed-run case study.
90-day revenue plan
Target $500-2k MRR from 10-30 small teams. Honest assessment: dev-tool monetization against free/OSS alternatives usually ramps slower than 90 days (HYPOTHESIS); 120-180 days to meaningful revenue is more realistic and within founder's stated runway tolerance.
Distribution path
Open-source core + MCP registry listings + HN/show-and-tell content + being the cited answer on 'how do I approve agent actions' threads. No enterprise sales motion. This matches founder's demonstrated-value selling style.
Pricing hypothesis
Free OSS self-host; hosted $29/seat/mo starter, $99-299/mo team tier with retention/replay/policy audit export; usage overage per 1k gated calls. Anchors to what teams already pay Langfuse/LangSmith-class tools (HYPOTHESIS on exact willingness).
Technical difficulty
Moderate and squarely in founder's strengths: an HTTP/SSE proxy with policy checks, a queue, webhooks, and a log viewer. Hardest parts are MCP protocol edge cases (streaming, auth passthrough, OAuth to downstream servers) and keeping up with fast-moving specs. No ML required.
Legal / regulatory risk
Low. The proxy carries customer tool traffic, so data-handling/security posture matters (SOC2 questions will come from bigger buyers — decline that segment initially). No regulated-industry exposure unless customers bring it.
Platform dependency
HIGH — the product lives inside the MCP spec and OpenAI/Google agent platforms. If providers ship native approval-gates and run-replay (plausible; it's an obvious roadmap item), the standalone wedge collapses to 'cross-provider neutrality + better UX', which may not sustain pricing. This is the central strategic risk.
Founder fit
Mixed. Fits: AI workflows, fast prototyping, systems thinking, demonstrated-value selling, micro-SaaS shape. Doesn't fit his proven edge: there is NO forced buyer here — nobody is legally compelled to gate agents (the gov-portal lesson, confidence 0.8, applies and this idea fails it). Buyers are developers, a notoriously payment-averse audience, and he has no standing distribution in the AI-dev community. Fit is moderate, below his mandate-filing archetype.
Breakout potential
If 'agent governance' becomes a compliance line-item (e.g., insurers or the EU AI Act effectively requiring audit trails for autonomous actions), this flips into a forced-buyer product and expands into policy/attestation — genuine breakout. That trigger is speculative today (HYPOTHESIS).
Final recommendation
HOLD / REVISIT — do not build yet. The convergence is real and the build is cheap, but this scores poorly on the founder's own decision rules: zero demand evidence in the input, no forced buyer, crowded adjacent tooling, and high platform-absorption risk. Revisit immediately if (a) a regulation or insurer/procurement standard starts requiring agent action audit trails (converts to his forced-buyer archetype), or (b) demand ingestion surfaces hiring/spend evidence for 'agent oversight/HITL' roles. Note: the system's own lesson (confidence 0.85) says demand sources are under-ingested, so the empty demand_evidence may be a collection gap rather than true absence — worth a targeted demand scan before final kill.
Next action
Run a demand probe, not a build: search job postings and vendor case studies for 'human-in-the-loop agent approval', 'agent audit trail', 'AI action governance' spend; interview 5 teams shipping on Gemini managed agents about what they'd pay to gate. Timebox 2 weeks; build only if ≥3 teams commit to a paid pilot.