New startup ideas · collection
collection
AI for people who run the AI themselves
Startup ideas for the early adopters and pro users who want to wield AI tools and skills directly, while most buyers still want it done for them as a service.
Most people do not want to operate AI. 61% of Americans have used it but 3% pay; half of US workers never touch it at work; 95% of enterprise pilots show no return, while more than 40% of knowledge workers quietly run personal AI tools that work better than the sanctioned ones. The market is splitting. One side wants the outcome delivered: agents and services that do the work, with a human accountable. The other side, a minority that is growing fast, wants the controls: Cursor went from $100 million to about $4 billion of annualized revenue in seventeen months, 80% of Lovable's builders are non-technical, Claude Max sells $200-a-month seats to heavy users, and Agent Skills and MCP became open standards anyone can install into their own assistant. This collection is for the second side: software, skills, tools and infrastructure that a person who runs AI themselves pays for, self-serve, with no services layer in between. United States only. No done-for-you agent services, no marketplaces.
What changed in the last twelve months
Dated facts the collection is built on; each links to where it was read.
61% of Americans use AI, 3% pay
Menlo Ventures' 2025 State of Consumer AI (a survey of more than 5,000 US adults in April 2025, published June 26, 2025) finds 61% of Americans have used AI and nearly one in five use it daily, yet only 3% pay for it; consumer AI spend is about $12 billion against 1.8 billion global users.
Half of US workers never use AI at work
Gallup's fourth-quarter 2025 workplace data: 49% of US employees never use AI in their role, 26% use it at least a few times a week and 12% daily. Frequent use keeps rising, but from a base where one worker in two has not started.
Gallup, Frequent Use of AI in the Workplace Continued to Rise in Q4
The GenAI Divide: 95% of pilots return nothing, workers run their own tools
MIT NANDA's State of AI in Business 2025 (research January-June 2025, published in August): despite $30-40 billion of enterprise spending, 95% of organizations see no measurable return and 5% of integrated pilots capture the value. Meanwhile more than 40% of knowledge workers use personal AI tools at work, a shadow AI economy whose users call the same models reliable in their own hands and unreliable inside enterprise systems.
Fortune, MIT report: 95% of generative AI pilots at companies are failing
Gartner: over 40% of agentic AI projects cancelled by 2027
On June 25, 2025 Gartner predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027 for cost, unclear value or weak risk controls; of thousands of agentic vendors it counted only about 130 as real and named the rest 'agent washing'. The done-for-you side of the market is where the failures cluster.
Augmentation edges ahead of automation on Claude.ai
The Anthropic Economic Index (November 2025 data, reported January 2026) measures 52% augmentation against 45% automation on Claude.ai, where people iterate with the model, while enterprise API traffic stays 75% automated; the February 2026 report has augmentation rising again on both. The people who run the AI themselves and the systems that run it for them are two different populations with two different curves.
Cursor: $100 million to about $4 billion annualized in seventeen months
Cursor's annualized revenue went from $100 million in January 2025 to $1 billion in November 2025 (Series D), $2 billion by February 2026 per Bloomberg and about $4 billion by May 2026; on June 16, 2026 SpaceX agreed to acquire it for $60 billion in stock, per press reports, the largest sale of a venture-backed startup. A tool the user operates, sold self-serve by the seat.
Lovable: $500 million ARR, 80% non-technical builders
Lovable reached $200 million ARR in November 2025 and about $500 million by June 2026 with 146 staff; 80% of the people building on it are not developers. The buyer builds the software themselves and pays by the month.
TNW, Lovable: $500M ARR, 146 staff, 80% non-technical builders
'Vibe coding' is Collins' Word of the Year 2025
On November 6, 2025 Collins Dictionary named 'vibe coding', turning natural language into software with AI, its Word of the Year, nine months after Andrej Karpathy coined it. The practice of running the AI yourself became a household word before it became a category.
CNN, 'Vibe coding' named Collins Dictionary's Word of the Year
Agent Skills: launched October 16, 2025, an open standard since December 18
Anthropic's Agent Skills package procedural knowledge as folders with a SKILL.md, scripts and resources that an assistant loads when the task calls for it; a skill-creator skill writes them. Released as an open standard on December 18, 2025 (agentskills.io), and by June 2026 about 40 products supported it, including Claude, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Goose. Skills are something a user can build, install, share and sell.
VentureBeat, Anthropic launches enterprise Agent Skills and opens the standard
MCP donated to the Agentic AI Foundation
On December 9, 2025 Anthropic donated the Model Context Protocol to the new Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI (goose and AGENTS.md are the other founding projects), with Google, Microsoft, AWS, Cloudflare and Bloomberg behind it. MCP had 97 million monthly SDK downloads and 10,000 active servers, with first-class support in ChatGPT, Claude, Cursor, Gemini, Copilot and VS Code: one plug any user's assistant can take.
Model Context Protocol blog, MCP joins the Agentic AI Foundation
$200-a-month seats for heavy users
Anthropic launched Claude Max on April 9, 2025 at $100 and $200 a month (5x and 20x the Pro limits) for people who 'collaborate with Claude extensively', with priority access to Claude Code and new models; OpenAI's $200 ChatGPT Pro had opened the tier in December 2024. A price point exists now for the individual who runs the AI all day.
IntuitionLabs, Claude Max Plan: $100 vs $200 Pricing and Usage Limits
n8n: $180 million at $2.5 billion for self-run automation
In October 2025 n8n raised a $180 million Series C at a $2.5 billion valuation led by Accel, with Nvidia's NVentures and Deutsche Telekom in the round, on revenue past $40 million growing tenfold in a year. Operators wiring AI steps into their own workflows, without an agency, is a venture-scale business.
n8n blog, n8n raises $180m to get AI closer to value with orchestration
What the directory shows
Companies across 15 accelerators whose pitch mentions each term: last two years / all-time / alive now.
- workflow 314 / 957 / 874
- automation 165 / 782 / 676
- prompt 63 / 144 / 126
- creator 47 / 340 / 285
- copilot 30 / 86 / 76
- no-code 12 / 118 / 101
- personal ai 10 / 24 / 22
- self-serve 8 / 31 / 24
- power user 7 / 17 / 16
- knowledge worker 6 / 14 / 13
- vibe cod 6 / 13 / 11
- agent builder 3 / 6 / 6
- developer tool 2 / 23 / 20
- prosumer 2 / 6 / 6
- low-code 1 / 22 / 18
- solopreneur 0 / 3 / 3
- self serve 0 / 1 / 1
Fresh concepts
Written for this collection, one call per sub-theme, from the dated developments above and what the directory shows from 2025 on; each names the development it rides and the 2025-26 signals it is grounded in. Each concept has its own page.
Set of 59, generated 2026-08-26 by claude-fable-5.
Skills, not services
Packaged Agent Skills, MCP servers and skill packs a pro user installs into their own assistant: building, testing, versioning, selling, auditing and updating skills; the SKILL.md economy and what it lacks.
Rides 'Agent Skills
Skillproof
Regression testing for the Agent Skills you actually depend on
A test harness for SKILL.md folders: a builder writes assertions against a skill's outputs, then Skillproof replays the suite across Claude, Codex, Copilot, Cursor and Gemini CLI and across model updates, flagging when a skill silently breaks.
grounded in 'Anthropic's Agent Skills...
Rides 'MCP donated to the Agentic AI Foundation' on December 9
Provenary
Scan third-party skills and MCP servers before you let them touch your data
A pre-install auditor: it statically and dynamically inspects a skill folder or MCP server for prompt injection payloads, silent network calls, credential access and data exfiltration paths, then issues a signed report and a personal allowlist.
grounded in 'MCP had 97 million monthly SDK downloads and 10,000 active servers' (December 9, 2025 donation to the Agentic AI Foundation).
Rides 'Agent Skills
Ledgerkit
Versioned skill packs that make a solo CPA's assistant work like a tax practice
A subscription library of Agent Skills for solo CPAs and small firms: entity-selection analysis, IRS notice responses, workpaper review checklists and state nexus rules, packaged as SKILL.md folders with the citations and review steps a licensed preparer needs, updated as rules change.
grounded in Emerging term 'cpa firms': first seen 2025, 3 companies in 2025-2026 vs 0 before; 'accounting firm': first seen 2025, 3 companies vs 0 before.
Rides 'Agent Skills
Vendfold
Licensing, signing and auto-update infrastructure for people who sell Agent Skills
An API and dashboard that lets a skill author sell direct from their own site: signed versioned packages, license keys checked at load time, staged rollouts and an update channel that pushes fixes into every buyer's assistant.
grounded in 'Skills are something a user can build, install, share and sell' (Agent Skills, open standard December 18, 2025; about 40 supporting products by June 2026).
Rides the YC Fall 2026 RFS 'Multiplayer AI'
Packroom
One shared skill library for a team where everyone runs their own agent
A private skill repository for 3 to 30 person teams: members propose skills, review diffs like pull requests, and approved versions propagate automatically into each person's Claude, Cursor or Copilot setup.
grounded in YC Requests for Startups, Fall 2026: 'Multiplayer AI'.
Rides '$200-a-month seats for heavy users'
Tokentab
Per-skill cost, routing and drift telemetry for the person who runs AI all day
A local telemetry layer for heavy individual users: it meters which installed skills and MCP servers burn tokens, routes each skill to the cheapest model that passes the user's own quality bar, and alerts when a model update changes a skill's behavior.
grounded in 'Anthropic launched Claude Max on April 9, 2025 at $100 and $200 a month (5x and 20x the Pro limits)'; OpenAI's $200 ChatGPT Pro opened the tier in December 2024, but the Max launch and the June 2026 40-product skill ecosystem are the 2025-2026 base.
Workbenches for people who run many agents
Orchestration surfaces, context and memory a person owns across assistants, session and task management, handoffs between Claude, Codex, Cursor and browser agents, for one person running a dozen agents.
'MCP donated to the Agentic AI Foundation'
Threadkeep
A memory vault you own that every assistant you run can read
Threadkeep is a local-first store of one person's working context, clients, projects, decisions, preferences, exposed as a single MCP server that Claude, Codex, Cursor and browser agents all read and write.
grounded in MCP donated December 9, 2025 with '97 million monthly SDK downloads and 10,000 active servers'; Claude Max launched April 9, 2025 at $100 and $200 a month for people who 'collaborate with Claude extensively'; jo (yc X26, 2026) shows the personal-agent demand this memory layer sits under.
'Agent Skills
Relayfile
Hand a running task from Claude Code to Codex without losing state
Relayfile snapshots a live agent session, the plan, constraints, files touched, decisions made and open questions, into a portable handoff bundle and injects it into whichever assistant picks the task up next.
grounded in Agent Skills became an open standard December 18, 2025 with about 40 supporting products by June 2026; Cursor grew from $100 million annualized in January 2025 to about $4 billion by May 2026, so heavy users now hold multiple paid coding assistants; 'claude code' is an emerging term with 5 companies in 2025-2026.
'Gartner
Meterhouse
One budget, meter and kill switch for every agent you run
Meterhouse connects to a user's Anthropic and OpenAI API keys, Claude Max usage and Cursor seat, shows live spend per agent and per task, routes background jobs to the cheapest model that passes the user's own quality bar, and hard-stops runaway loops.
grounded in Gartner prediction dated June 25, 2025 citing cost, unclear value and weak risk controls; Claude Max launched April 9, 2025 at $100 and $200 a month; Woz (yc W25, 2025) reducing Claude Code token cost by 50% shows individuals already pay for spend reduction.
'The GenAI Divide
Attestly
Audit trail and approval inbox for the agents you run at work
Attestly routes every action from an employee's personally run agents through a logged proxy, lands outputs in one review queue with checks the user defines, and exports a clean activity log they can hand to IT or compliance.
grounded in MIT NANDA's State of AI in Business 2025 (published August 2025): 95% of organizations see no measurable return while more than 40% of knowledge workers use personal AI tools at work; Anthropic Economic Index (November 2025 data): 52% augmentation on Claude.ai, meaning these users already review what their AI produces.
'Agent Skills
Skillvane
Version control and regression tests for the skills your agents load
Skillvane is a registry and CI service for SKILL.md folders and MCP server configs: it versions them, runs them against Claude, Codex, Copilot and Cursor on every model release, flags behavioral drift, and pins or rolls back with one call.
grounded in Agent Skills open standard at agentskills.io since December 18, 2025, with about 40 supporting products by June 2026; MCP at 10,000 active servers as of the December 9, 2025 donation; a skill-creator skill already writes skills, so the corpus of untested skills grows daily.
'[Fall 2026] Multiplayer AI'
Crewline
A shared board where each teammate's agents pick up each other's work
Crewline is a multiplayer workbench for small teams where every person runs their own Claude, Codex or Cursor agents: a shared task graph where one person's agent output becomes another person's agent input, with per-person ownership, budgets and review gates.
grounded in YC Requests for Startups, Fall 2026 edition: 'Multiplayer AI'; n8n's October 2025 Series C, $180 million at a $2.5 billion valuation on revenue past $40 million growing tenfold, proving self-run orchestration is a venture-scale purchase; Sim (yc X25, 2025) validating the single-player agent workspace.
What vibe coders need after the demo
Hosting, auth, data, payments, monitoring, backups and maintenance for software built by non-developers on Lovable-class tools; keeping it alive and secure at month six without hiring an engineer.
Lovable
Wardkey
Security scanner that finds and fixes exposed keys in vibe-coded apps
Wardkey connects to a Lovable or Bolt-built app, scans for exposed API keys, missing row-level security, open database policies and unauthenticated endpoints, then generates the exact prompt to paste back into the builder to fix each hole.
grounded in Lovable reached '$200 million ARR in November 2025 and about $500 million by June 2026 with 146 staff; 80% of the people building on it are not developers'; Lovable raised another $400M on 2026-08-12; 'vibe coding' was Collins' Word of the Year on November 6, 2025; YC's Fall 2026 RFS includes 'A Cloud for Small Software'.
YC Fall 2026 RFS 'Self-Maintaining APIs'
Stillup
Uptime and error monitoring that answers in fix prompts, not stack traces
Stillup watches a vibe-coded app in production: uptime, crashes, failed API calls, broken flows.
grounded in YC's current Fall 2026 RFS lists 'Self-Maintaining APIs' and 'A Cloud for Small Software'; Lovable at 'about $500 million by June 2026' ARR with '80% non-technical builders'; the Anthropic Economic Index (November 2025 data, reported January 2026) shows 52% augmentation on Claude.ai, people who iterate with the model rather than delegate to it.
A Cloud for Small Software
Copystone
Automatic backups and one-click restore for apps built without engineers
Copystone connects to the Supabase or Firebase project behind a vibe-coded app and takes versioned snapshots of the database, storage and auth config, with point-in-time restore and a portable export when the builder outgrows the platform.
grounded in YC Fall 2026 RFS 'A Cloud for Small Software'; Lovable's '80% of the people building on it are not developers' at about $500 million ARR by June 2026; Bloom (yc X25) 'build and share mobile apps in seconds' and bitrig (yc S25) 'vibe code, test, and deploy Swift apps' show 2025 cohorts shipping app volume with no ops layer behind it.
Self-Maintaining APIs
Groundskeep
Monthly maintenance for shipped vibe-coded apps, applied as reviewable patches
Groundskeep watches the dependencies, API versions and platform deprecations behind a non-developer's shipped app, then opens plain-English patch proposals: what broke or will break, what the fix changes, one button to apply and one to roll back.
grounded in YC Fall 2026 RFS 'Self-Maintaining APIs'; Agent Skills 'released as an open standard on December 18, 2025' with 'about 40 products' supporting it by June 2026, so maintenance procedures can ship as versioned SKILL.md folders the user's own assistant executes; MIT NANDA (published August 2025) found 'more than 40% of knowledge workers use personal AI tools at work'.
A Cloud for Small Software
Tillhouse
Payments, sales tax and refunds as one drop-in for non-developer founders
Tillhouse is a payments layer purpose-built for vibe-coded apps: a snippet the builder pastes via a prompt that handles Stripe checkout, US sales tax registration thresholds, subscription lifecycle, dunning and refunds, with a dashboard written for someone who has never seen a webhook.
grounded in YC Fall 2026 RFS 'A Cloud for Small Software'; Lovable at about $500 million ARR by June 2026 with 80% non-technical builders; Menlo Ventures (published June 26, 2025) found only 3% of Americans pay for consumer AI, so builders monetizing their apps need conversion-grade payment flows, not duct tape; MCP with '97 million monthly SDK downloads and 10,000 active servers' after its December 9, 2025 donation makes the integration installable from the builder's own assistant.
$200-a-month seats for heavy users
Spendgate
Meter, cap and route the AI spend inside apps vibe coders shipped
Spendgate is a proxy key for the LLM calls inside a shipped vibe-coded app: per-user metering, hard monthly caps, automatic fallback to cheaper models when a budget nears, and alerts before the bill surprises the owner.
grounded in Claude Max launched 'on April 9, 2025 at $100 and $200 a month'; Woz (yc W25) is a 'Claude Code plugin that reduces token consumption and cost by 50%', proof the 2025 cohort pays for token cost control; Cursor's run from '$100 million in January 2025' to 'about $4 billion by May 2026' annualized shows usage-based AI exposure compounding; the collection brief names cost control as one of the unbuilt layers.
Self-run automation for small-business operators
n8n and Zapier-class workflows with AI steps that the owner builds and maintains: templates, testing, error handling, data hygiene and cost control for an operator with no agency and no developer.
'n8n
Dryloop
Rehearsal mode for the automations a small business owner builds alone
Dryloop connects to an operator's n8n or Zapier account and runs every workflow against synthetic records generated from their real data before it touches a live customer.
grounded in n8n raised a $180 million Series C at $2.5 billion in October 2025 on revenue past $40 million growing tenfold in a year; Gartner (June 25, 2025) predicts over 40% of agentic AI projects cancelled by end of 2027 for cost, unclear value or weak risk controls; Confident AI (yc W25) proves eval tooling sells, but to AI teams, not operators.
'$200-a-month seats for heavy users'
Meterly
One metered key with spend caps for every AI step you run
Meterly issues a proxy API key an operator pastes into the AI nodes of their Zapier, Make or n8n workflows.
grounded in Claude Max launched April 9, 2025 at $100 and $200 a month for people who run AI all day; Woz (yc W25) sells a 'Claude Code plugin that reduces token consumption and cost by 50%', showing cost control alone is a purchasable product; Menlo Ventures (June 26, 2025) counts only 3% of Americans paying for AI, the exact self-selecting minority Meterly serves.
'Self-Maintaining APIs' (YC RFS
Flowmedic
Watches your automations, explains failures in plain English, proposes the fix
Flowmedic monitors an operator's live workflows and catches silent failures: a webhook that stopped firing, a third-party API that changed a field name, an AI step returning garbage.
grounded in YC's current Fall 2026 Requests for Startups includes 'Self-Maintaining APIs'; Kestra raised a $25M Series A on 2026-03-31 for workflow automation and orchestration tools; Libretto (yc X25) turns website workflows into reliable APIs, evidence that brittleness of small-operator integrations is a fundable problem.
'MCP donated to the Agentic AI Foundation'
Scrubdeck
A data-cleaning step any workflow can call, with rules the owner keeps
Scrubdeck is an MCP server and native n8n and Zapier step that deduplicates, normalizes and validates customer and order records as they flow between an operator's apps.
grounded in On December 9, 2025 MCP was donated to the Agentic AI Foundation with 97 million monthly SDK downloads and 10,000 active servers, with first-class support in ChatGPT, Claude, Cursor, Gemini, Copilot and VS Code; Hanji (yc X25) turns 'your hardest documents into reliable data', showing data quality sells standalone.
'Agent Skills
Opshand
Turns your written SOPs into versioned Agent Skills with tests included
Opshand takes an operator's existing SOP document or checklist and compiles it into an Agent Skill folder plus a matching n8n workflow, with test cases derived from the SOP's own steps.
grounded in Agent Skills launched October 16, 2025 and became an open standard on December 18, 2025 (agentskills.io); by June 2026 about 40 products supported it, including Claude, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Goose; Altrina (yc W25) pitches 'The SOP Automation Platform', confirming SOP automation is a live wedge.
'Multiplayer AI' (YC RFS
Crewtrace
Shared visibility when five people at one business each run their own automations
Crewtrace is a registry for a small team where the owner, the office manager and two techs each built their own workflows and skills.
grounded in YC's Fall 2026 Requests for Startups includes 'Multiplayer AI'; MIT NANDA (research January to June 2025, published August 2025) finds more than 40% of knowledge workers use personal AI tools that outperform sanctioned ones; the Anthropic Economic Index (January 2026 report) measures 52% augmentation on Claude.ai, people iterating with the model themselves.
Solo professionals who wield AI themselves
Lawyers, accountants, physicians, consultants, real-estate and insurance agents using AI directly under their own license and liability: verification, citations, records of what the AI did, and the profession's knowledge packaged for the practitioner, not sold as a service.
Agent Skills
Veracite
Citation verification and AI work records for solo attorneys who draft with Claude
A desktop and browser tool where a solo attorney runs their AI drafting, and every cited case is checked against primary law before it reaches a filing.
grounded in 'ainative law' is an emerging term with 3 companies in 2025-2026 vs 1 before; Agent Skills became an open standard on December 18, 2025 with about 40 products supporting it by June 2026; MIT NANDA (published August 2025) found more than 40% of knowledge workers using personal AI tools at work; Claude Max launched April 9, 2025 at $100 and $200 a month for people who collaborate with AI extensively.
MCP donated to the Agentic AI Foundation on December 9
Tickstone
Turns a solo CPA's AI sessions into reviewable workpapers with tickmarks and source trails
A workspace where a solo CPA runs AI over client books and every number the model touches is tied back to a source document with a tickmark, producing workpapers that survive peer review and an IRS exam.
grounded in 'cpa firms' first seen 2025, 3 companies in 2025-2026 vs 0 before; 'accounting firm' first seen 2025, 3 companies vs 0 before; Rillet raised $100M growth on 2026-08-21 and reached unicorn status as an AI accounting startup; MCP had 97 million monthly SDK downloads and 10,000 active servers when donated on December 9, 2025.
The GenAI Divide
Chartproof
A verification layer for physicians who use AI on clinical notes under their own license
A tool for independent physicians who already draft notes and letters with AI: it checks model output against the chart facts the doctor loaded, flags unsupported clinical statements, and keeps a record of what the AI suggested versus what the physician signed.
grounded in MIT NANDA's State of AI in Business 2025 (published August 2025): more than 40% of knowledge workers use personal AI tools at work and call the same models reliable in their own hands, unreliable inside enterprise systems; 'health assistant' first seen 2025, 3 companies in 2025-2026 vs 0 before; Galen AI (yc X25) built a personal AI healthcare agent and its site is down.
MCP donated to the Agentic AI Foundation on December 9
Coverlens
Policy-form verification for independent insurance agents who quote with AI
Software for the independent P&C agent who uses AI to compare coverage: it reads the actual policy forms, verifies every AI-generated statement about limits, exclusions and endorsements against form language, and archives the comparison the agent showed the client.
grounded in 'ainative insurance' first seen 2025, 4 companies in 2025-2026 vs 0 before; Harper (yc W25) is an AI-native commercial insurance brokerage; the crowded-phrase list shows 6 companies on 'commercial insurance' and 5 on 'insurance claims' in 2025-2026; MCP reached 10,000 active servers and 97 million monthly SDK downloads by its December 9, 2025 donation.
Agent Skills
Methodkit
Solo consultants package their methodology as versioned Agent Skills they own and resell
A build-test-version environment where an independent consultant turns their frameworks, diagnostic questionnaires and deliverable templates into Agent Skills, runs them against past-engagement test cases, and installs them into Claude, Copilot or Cursor.
grounded in Agent Skills released as an open standard on December 18, 2025 (agentskills.io) with about 40 supporting products by June 2026, including Claude, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Goose; Claude Max launched April 9, 2025 at $100 and $200 a month; 'knowledge work' first seen 2025, 4 companies in 2025-2026 vs 0 before.
MCP donated to the Agentic AI Foundation on December 9
Attestrail
Tamper-evident logs of every AI action, built for licensed professionals' liability files
An API and local proxy that sits between a professional's assistant and their MCP servers, writing a hash-chained record of every prompt, tool call, source document and model version, then rendering it as a human-readable file for a malpractice carrier, board, or court.
grounded in MCP had 97 million monthly SDK downloads and 10,000 active servers at its December 9, 2025 donation, co-founded with Block and OpenAI with Google, Microsoft, AWS, Cloudflare and Bloomberg behind it; Gartner predicted on June 25, 2025 that more than 40% of agentic AI projects will be cancelled by end of 2027, with weak risk controls a named cause; the Anthropic Economic Index (November 2025 data) shows 52% augmentation on Claude.ai, people iterating with the model.
Shadow AI made sanctioned
The 40%-plus running personal AI tools inside companies: data boundaries, policy, receipts and reimbursement for personal subscriptions, bring-your-own-AI programs, and the tooling that lets a power user stay inside the rules without giving up their tools.
'The GenAI Divide
Scrubline
Local redaction proxy that makes your personal AI accounts safe for work data
A browser extension and local proxy that masks customer names, financials and code identifiers before a prompt leaves for the user's personal ChatGPT or Claude account, then reverses the mapping in the response so the answer reads naturally.
grounded in MIT NANDA's State of AI in Business 2025 (research January-June 2025, published August): 'more than 40% of knowledge workers use personal AI tools at work' and users call the same models 'reliable in their own hands and unreliable inside enterprise systems'.
'$200-a-month seats for heavy users'
Stipendly
Turn personal Claude Max and ChatGPT Pro seats into managed employer stipends
Self-serve stipend software that lets a manager reimburse employees' personal AI subscriptions with receipts, policy acknowledgment and a monthly work-use attestation generated from the employee's own usage export.
grounded in Claude Max launch (April 9, 2025) at $100 and $200 a month for people who 'collaborate with Claude extensively'; MIT NANDA (published August 2025): $30-40 billion of enterprise spending with 95% of organizations seeing no measurable return while over 40% of knowledge workers use personal tools; Gallup Q4 2025: 12% of US employees use AI daily.
'MCP donated to the Agentic AI Foundation' (December 9
Tollgate
A policy gateway between your assistant and every MCP server it touches
A lightweight gateway a power user runs on their own machine that sits between their assistant and its MCP servers, enforcing an allowlist, blocking defined data categories from egress, and writing a signed local audit log.
grounded in MCP donation to the Agentic AI Foundation under the Linux Foundation on December 9, 2025, with 97 million monthly SDK downloads, 10,000 active servers and first-class support in ChatGPT, Claude, Cursor, Gemini, Copilot and VS Code; MIT NANDA (August 2025) on the 40%-plus shadow AI economy inside companies.
'Agent Skills
Skillvet
Scan, pin and approve Agent Skills before they touch company data
A vetting service for the Agent Skills and MCP servers employees install themselves: it scans SKILL.md folders and server code for prompt injection, undisclosed network calls and data exfiltration paths, pins approved versions, and produces an approval record a power user can attach to an IT request.
grounded in Agent Skills launched October 16, 2025 and became an open standard on December 18, 2025 (agentskills.io); by June 2026 about 40 products supported it including Claude, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Goose.
'The GenAI Divide
Daylight
Self-serve shadow AI registry and policy for companies with no security team
An ops lead at a 20-200 person company sends one link; each employee self-declares the AI tools they actually use and optionally installs a browser extension for usage counts.
grounded in MIT NANDA (published August 2025): more than 40% of knowledge workers use personal AI tools at work; Gallup Q4 2025: 26% of US employees use AI at least a few times a week and 12% daily, while 49% never do, meaning usage is concentrated in a visible minority a survey can actually capture.
The pro user's back office
Usage metering, cost control across many subscriptions and APIs, model routing and fallback for one person or a small team, the $200-seat decision, sharing capacity inside a household or a studio.
'$200-a-month seats for heavy users'
Ledgerline
Rightsizing dashboard for everyone paying for AI out of their own pocket
Ledgerline pulls billing exports and usage logs from a person's AI subscriptions and API keys into one ledger, then answers the recurring question: is the $200 Max seat, the $20 Pro seat, or pay-per-token API the cheapest way to get the usage you actually have.
grounded in Menlo Ventures' 2025 State of Consumer AI (published June 26, 2025): 61% of Americans use AI, only 3% pay, consumer AI spend about $12 billion.
'MCP donated to the Agentic AI Foundation'
Switchyard
One metered endpoint with routing, fallback and per-person caps for tiny teams
Switchyard is a drop-in API proxy for a two-to-ten person studio: every member gets a key, every request is routed to the cheapest model that meets a declared quality floor, and rate-limit errors fail over to a second provider automatically.
grounded in MCP donated to the Linux Foundation's Agentic AI Foundation on December 9, 2025 with 97 million monthly SDK downloads.
'$200-a-month seats for heavy users'
Hearthmeter
Usage budgets and one bill for the household that shares AI plans
Hearthmeter gives a household one view of every family member's AI accounts: who is on which plan, how much of each plan's limit they actually use, and which combination of tiers covers the family for the least money.
grounded in Menlo Ventures' survey of 5,000+ US adults (April 2025, published June 26, 2025): 61% of Americans use AI and nearly one in five use it daily, but only 3% pay.
'The GenAI Divide
Seatcase
Measures who on your team earns a Max seat and who wastes one
Seatcase is a self-serve tool for an engineering or ops manager who owns ten to fifty AI seats across Cursor, Claude and Copilot.
grounded in MIT NANDA's State of AI in Business 2025 (published August 2025): 95% of organizations see no measurable return while 40%+ of knowledge workers run personal tools.
'Agent Skills
Tokencairn
Profiler that shows what each installed skill and MCP server really costs
Tokencairn instruments a user's assistant sessions and attributes token spend to each installed Agent Skill and MCP server, the way a browser profiler attributes CPU to tabs.
grounded in Agent Skills launched October 16, 2025, released as an open standard on December 18, 2025, with about 40 supporting products by June 2026 including Claude, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Goose.
'n8n
Fusegate
Budget caps, fallback and kill switches for automations you run yourself
Fusegate is a credential proxy for people who run their own automations in n8n and similar tools: each workflow gets its own budget, a fallback chain of models for rate-limit and outage events, and a kill switch that halts a loop the moment spend deviates from that workflow's baseline.
grounded in n8n raised $180 million in October 2025 at a $2.5 billion valuation on revenue past $40 million growing tenfold in a year.
Evaluation and trust the user controls
Checking outputs, regression tests for prompts and skills, provenance and citations, guardrails and review gates the user sets rather than the vendor, for people whose name goes on the result.
Agent Skills
Skillbench
Regression testing for Agent Skills before every model and skill update
A hosted test harness for Agent Skills.
grounded in 'Agent Skills package procedural knowledge as folders with a SKILL.md' launched October 16, 2025, 'released as an open standard on December 18, 2025 (agentskills.io), and by June 2026 about 40 products supported it, including Claude, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Goose'.
The GenAI Divide
Citelock
Verifies every citation in AI-drafted work before a licensed professional signs it
A checker for solo lawyers, CPAs and consultants who draft with AI under their own license.
grounded in MIT NANDA's State of AI in Business 2025 (published August 2025): 'more than 40% of knowledge workers use personal AI tools at work'.
MCP donated to the Agentic AI Foundation
Middlegate
A local gateway where you set the rules for what your MCP servers can do
An MCP proxy an individual installs between their assistant and their servers.
grounded in 'On December 9, 2025 Anthropic donated the Model Context Protocol to the new Agentic AI Foundation', with 'MCP had 97 million monthly SDK downloads and 10,000 active servers, with first-class support in ChatGPT, Claude, Cursor, Gemini, Copilot and VS Code'.
n8n
Driftwatch
Catches output drift in the automations small operators wired themselves
Snapshot-based regression testing for AI steps inside self-built workflows.
grounded in 'In October 2025 n8n raised a $180 million Series C at a $2.5 billion valuation led by Accel...
'Vibe coding' is Collins' Word of the Year 2025
Shipcheck
Pre-launch review gates non-technical builders run on their own vibe-coded apps
A self-serve gate a Lovable or Bloom builder runs before shipping: it scans the generated app for exposed keys, open database rules, missing auth on routes and broken payment flows, explains each finding in plain language, and blocks the builder's own publish step until they accept or fix each one.
grounded in 'On November 6, 2025 Collins Dictionary named vibe coding...
Augmentation edges ahead of automation on Claude
Traceline
A claim-level provenance trail for every number in an AI-assisted report
A workspace layer for analysts and researchers who draft with AI but answer for the numbers.
grounded in The Anthropic Economic Index (November 2025 data, reported January 2026): '52% augmentation against 45% automation on Claude.ai', with 'the February 2026 report has augmentation rising again on both'.
Learning to wield AI
Practice environments, drills, certifications employers accept, coaching and curricula for people crossing from 'done for me' to 'I run it': measured skill, not courses, sold to the individual.
Rides 'Cursor
Drillyard
Scored practice repos where you learn to drive coding agents well
Drillyard is a bank of sandboxed, deliberately broken codebases with timed drills: fix the bug, ship the feature, refactor safely, but you must do it by directing Cursor or Claude Code, not by hand.
grounded in Cursor went from $100 million annualized in January 2025 to about $1 billion in November 2025 and about $4 billion by May 2026; Agent Skills became an open standard on December 18, 2025 with about 40 products supporting it by June 2026; YC's Fall 2026 RFS includes 'The Primer'; MangoDesk (YC S25) and Halluminate (YC S25) sell RL environments, proving graded environments are buildable, but both train models, not people.
Rides 'Half of US workers never use AI at work' from Gallup's Q4 2025 data
Passrate
A proctored AI operation exam scored from your real agent transcripts
Passrate is a timed assessment where the candidate drives their own AI assistant through realistic work tasks: research, analysis, document production, automation.
grounded in Gallup Q4 2025: 49% of US employees never use AI in their role, 12% daily; MIT NANDA's State of AI in Business 2025 (published August 2025): more than 40% of knowledge workers use personal AI tools at work; Guideless raised €1M on 2026-08-20 to streamline software training, and Medly AI raised $8M on 2026-08-19 for AI tutoring, showing money moving into measured AI learning; YC's Fall 2026 RFS lists 'The Primer'.
Rides ''Vibe coding' is Collins' Word of the Year 2025' (November 6
Patchcraft
Debugging drills that teach non-technical builders to maintain what they vibe coded
Patchcraft takes the app a non-technical founder shipped with Lovable or Bloom, generates sandboxed copies with realistic injected failures, broken auth, a bad migration, a silent API change, and coaches them through fixing each one with their AI assistant.
grounded in Collins named 'vibe coding' Word of the Year on November 6, 2025; Lovable hit $200 million ARR in November 2025 and about $500 million by June 2026 with 80% non-technical builders, then raised $400M on 2026-08-12; bitrig (YC S25) extends vibe coding to Swift apps from an iPhone; YC's Fall 2026 RFS includes 'A Cloud for Small Software'.
Rides 'Agent Skills
Skillsmith
A workshop for writing, testing and versioning Agent Skills that actually hold up
Skillsmith is a practice environment for authoring SKILL.md skills: you write a skill, it runs your skill against a battery of held-out tasks across Claude, Codex and Cursor, and shows where the instructions fail, with graded exercises that build from a one-file skill to a versioned pack with scripts and evals.
grounded in Agent Skills launched October 16, 2025 and became an open standard at agentskills.io on December 18, 2025, supported by about 40 products including Claude, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Goose by June 2026; MCP's donation to the Agentic AI Foundation on December 9, 2025 with 10,000 active servers shows the parallel curve; the keyword 'prompt' appears in 62 companies from the last two years, but skill authoring craft is untracked.
Rides 'Augmentation edges ahead of automation on Claude
Tickmark
Synthetic client caseloads where CPAs drill AI-assisted work before trying it on real clients
Tickmark gives solo accountants and small-firm partners fictional but realistic client files, messy ledgers, ambiguous 1099s, a sales tax nexus problem, and drills them on running AI through the engagement while catching every hallucinated figure, with scoring built around review standards a licensee is personally accountable for.
grounded in Anthropic Economic Index, November 2025 data reported January 2026: 52% augmentation vs 45% automation on Claude.ai; emerging terms 'cpa firms' and 'accounting firm' both first seen 2025 with 3 companies each in 2025-2026; Cifrato (YC W25), Combinely (YC X25) and Moby Analytics (YC X25) all sell AI that does accounting work; Rillet raised $100M on 2026-08-21 for AI accounting.
Rides 'MCP donated to the Agentic AI Foundation' (December 9
Postgame
An MCP server that scores your own agent sessions and drills your weakest habits
Postgame plugs into whatever assistant you already run via MCP, keeps your transcripts in a store you own, and grades each session on cost, verification, delegation and rework, then generates personalized drills from your actual failure patterns, the prompt you re-ran four times, the output you accepted unverified.
grounded in MCP donated December 9, 2025 with 97 million monthly SDK downloads and 10,000 active servers; Claude Max launched April 9, 2025 at $100 and $200 a month for people who 'collaborate with Claude extensively'; Woz (YC W25) cuts Claude Code token cost 50% via plugin, proving session-level waste is real and measurable; MIT NANDA (August 2025): users call the same models reliable in their own hands and unreliable in enterprise systems.
Analysts, researchers and creators running their own stack
Personal data pipelines, source management, note and knowledge systems, publishing and monetization tooling for individuals who use AI as their instrument and sell the output themselves.
Rides 'Augmentation edges ahead of automation on Claude
Citegrid
Every number in your published research links to a source snapshot you verified
A verification layer for independent analysts who sell their own research.
grounded in Anthropic Economic Index (November 2025 data, reported January 2026): 52% augmentation vs 45% automation on Claude.ai, with the February 2026 report showing augmentation rising again.
Rides 'MCP donated to the Agentic AI Foundation'
Mnemos
Your research corpus as a private MCP server every assistant can query
A hosted personal memory for researchers: it ingests notes, PDFs, highlights, interview transcripts and web clips, builds a citable index, and exposes it as a private MCP server.
grounded in On December 9, 2025 Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, with 97 million monthly SDK downloads, 10,000 active servers, and first-class support in ChatGPT, Claude, Cursor, Gemini, Copilot and VS Code.
Rides '$200-a-month seats for heavy users'
Meterline
Model routing and cost accounting for one person's AI research pipeline
A routing and metering layer for solo analysts who run heavy multi-model pipelines.
grounded in Claude Max launched April 9, 2025 at $100 and $200 a month for people who 'collaborate with Claude extensively'; OpenAI's $200 ChatGPT Pro opened the tier in December 2024.
Rides 'Agent Skills
Skillcask
Version, test, and sell your expertise as licensed Agent Skills
Authoring and licensing infrastructure for experts who package their methodology as Agent Skills.
grounded in Agent Skills launched October 16, 2025 and became an open standard on December 18, 2025 (agentskills.io); by June 2026 about 40 products supported it, including Claude, OpenAI Codex, GitHub Copilot, VS Code, Cursor, Gemini CLI and Goose.
Rides the Fall 2026 YC Request for Startups 'Self-Maintaining APIs'
Stackfeed
A personal data pipeline that repairs itself when sources change
A self-maintaining pipeline for analysts who keep proprietary datasets current by hand: point it at recurring sources such as SEC filings, agency releases and earnings transcripts, and it builds extraction schemas, detects when a source's format shifts, patches its own extractor, and delivers dated diffs to the analyst's dataset.
grounded in 'Self-Maintaining APIs' is a current Fall 2026 YC RFS theme.
Rides the Fall 2026 YC Request for Startups 'Multiplayer AI'
Claimboard
A shared evidence ledger for small teams where everyone runs their own agent
A workspace for two-to-five person research shops in which each member runs their own assistant and pipelines.
grounded in 'Multiplayer AI' is a current Fall 2026 YC RFS theme.
From the catalog
8 ideas from Horizontal AI assistants, Agent infrastructure, Developer tools, Vertical AI agents, B2B SaaS, Consumer that pass the collection's filter, 2 of them venture-grade, sorted by Idea Score. A deck focused on this collection deals exactly these.
Consumer and commerce · Consumer
Phonika
Speech practice for kids at one tenth the cost of a therapist.
Phonika is a subscription app that gives children with articulation delays daily speech practice, using on-device speech models trained to score child phoneme production - a problem general ASR handles badly - and adapt drills the way a speech-language pathologist would between sessions.
AI and software · Agent infrastructure
Certloop
Accredited hardware-in-the-loop cloud that certifies agent policies before they touch real machines.
Certloop builds racks of real PLCs, drives, sensors, and robot actuators wired into a cloud API; enterprise robotics and industrial teams upload agent policies and run them against physical hardware and vendor-published digital twins before deployment.
AI and software · Developer tools
Palmforge
Build private software for your own life that never leaves your phone.
Palmforge is a consumer subscription that lets one person describe a tool they need (a budget rule, a medication tracker, a school-schedule agent) and compiles it into a small program that runs entirely on their own device against their own data.
AI and software · Agent infrastructure
Mnemora
A private memory layer that follows you across every AI app.
Mnemora is a consumer subscription that builds one persistent, user-owned memory from everything you tell any AI assistant, then serves it back to ChatGPT, Claude, and every agent app through a standard connector.
AI and software · Developer tools
Miniplex
The cloud built for a billion small apps.
Miniplex is purpose-built cloud infrastructure for agent-generated software: sub-second cold-start microVMs, per-app metering in fractions of a cent, and a deploy API that coding agents and app builders call directly.
B2B, security and compliance · B2B SaaS
Stocklight
A demand forecasting model any small merchant can self-serve, any app can embed.
Stocklight is a subscription forecasting layer for small merchants who make, move or sell physical goods: connect a store, a POS or a wholesale ledger and get SKU-level demand, reorder points and stockout risk that account for promotions, seasonality and local events.
AI and software · Developer tools
Portway
A machine-readable catalogue of every internal enterprise app, so agents can use them.
Portway is a subscription platform where a large company describes its internal applications once, in a no-code editor, as capability manifests: what each app does, which actions it exposes, who may call them, and what the fields mean.
Consumer and commerce · Consumer
Wildframe
A subscription social world where friends generate and inhabit persistent AI game worlds.
Wildframe is a consumer subscription: pay monthly, and you and your friends describe worlds that render into persistent, playable multiplayer games in seconds.