Santanu Dasgupta's second idea failed on an ordinary question. The MVP received positive reactions, but mostly to the idea itself. Then people asked what it could automate. His answer: anything.
He recounted this in a later Indie Hackers post (Journey so far building an AI startup: Mistakes and Learnings[1]). That product aimed to replace Zapier for AI workflow automation. Saying “anything” answered nothing. From a procurement perspective, it cannot be matched to a budget or process, or used to explain boundaries to a specific person. That is an interpretation of his account, not a customer's words. He wrote that financial and mental pressure became severe and that quitting briefly seemed the only option.
Over 8 months, he went through two failed ideas. The third did not initially resemble today's website either. In the same post, Meerkats AI was a no-code LLM app builder for solopreneurs. Explaining the idea, he said AI was good at generating content but that higher-level work such as style, tone and SEO still needed people. His concrete example was an app built with the tool that could deeply research any topic online and rewrite the findings into multiple forms of content. The build used Langchain, Langgraph, OpenAI APIs, Node.js and other components (same source). With the MVP just finished, he asked how to acquire his first users, saying Twitter and LinkedIn had not produced much.
Early distribution targeted agencies already running client campaigns but not yet using AI agents. According to Startup Founder Stories' account of an Indie Hackers interview, the team cold-started and conducted targeted outreach to agencies in its network, ran online and in-person educational events, and acquired customers through LinkedIn posts (Startup Founder Stories[2]). The materials give no agency count, conversion rate or CAC.
Today's website sells another layer. As read on November 17, 2026, Meerkats describes itself as a context layer for AI systems: business data, definitions, state, rules and permitted actions connected into managed context that AI agents can read and act on (Meerkats: the context layer for AI systems[3]).
In practical terms, Claude Code, Codex, custom agents, n8n workflows and internal apps can read the same context through MCP and APIs. The model side connects to Claude, GPT, Gemini or proprietary models; changing models or rules does not require rebuilding everything. Data syncs on a schedule, but is read live again before any action occurs.
Image: official page[3], an architecture diagram captured in a browser on 2026-11-18; not a test performed in this research.
Compared with the old “anything,” the semantic layer makes a real distinction. Each entity, metric and relationship is defined once; platform labels map to the customer's own definitions; unrecognised names are rejected rather than guessed. The website gives a permissions example too: a read-only partner attempting to change an advertising budget is blocked.
Delivery is designed to be inspectable. A person writes a playbook in natural language; the system turns it into typed automation with triggers, windows, conditions and limits. It runs on a schedule, retains state between runs, and applies rule changes on the next run. Every agent run produces a receipt: what it read, which rule version it followed, which permission authorised it and what it changed, plus read-back verification, tokens and duration. Work that can be checked off and a statement that “it has run for you” are different businesses.
In his Indie Hackers interview, Dasgupta said he had spent 20 years in SaaS GTM, moving from development to growth and consulting across the US, Europe and India, repeatedly seeing GTM depend on manual work, fragmented tools and agency execution (Growing an AI orchestration platform to $3k MRR in 4 weeks[4]). That background explains his description of Meerkats as a “digital growth agency in software”: replacing much of the repetitive work of SDRs, marketers and agencies—capturing and enriching relationship data, generating leads, running campaigns and following up, all through chat (same source).
The financial account has a specific reference point. In the same interview, whose page shows no publication date, he said the platform had launched four weeks earlier and MRR exceeded $3k. That is an interview-time figure, not today's. Other financial support mentioned included Alumni New Venture Challenge funding from the University of Chicago Polsky Center, funding operations through agency services delivered using the platform, and model-cost credits from Azure, OpenAI and Anthropic. No amounts were given, leaving profit unclear.
The pricing page, read on 2026-11-17, lists four monthly tiers starting at $29. Enterprise is a custom annual contract running on the customer's own Snowflake, BigQuery or Databricks warehouse, without copying data out. For someone facing procurement, that statement can matter more than an accuracy number (Pricing[5]).
Why not build it yourself? His interview answer was practical. Startup GTM teams chose Meerkats over Claude Managed Agents and OpenAI Agent Harness because they wanted the cheapest suitable model for each task, and might not have anyone capable of building an Agent Harness. They outsourced the whole job. He called this “Service as a Software,” even an outcome-oriented service. The pricing page sells connections, data volume, AI credits and concurrent automation allowances; the interview describes taking over a stretch of growth execution. Those are not automatically the same deliverable.
On the same path of giving AI business context, DataHub's Shirshanka Das told the SourceForge Podcast that its current focus was analytical agents: retrieve context, then write corrections and observations after an answer back into the graph, creating memory across tasks. He described providing context versus providing the agent itself as a boundary still under consideration, currently preferring to let companies build their own agents on top (The Context Platform for AI Agents: DataHub | SourceForge Podcast, episode #126[6]). Meerkats' website instead puts playbooks, permitted actions and per-run receipts in the same layer (Meerkats: the context layer for AI systems[3]). Both sell the idea of agents understanding the business. DataHub's central current use case is querying data; Meerkats turns playbooks into rule-based automation with inspectable records.
The difference between today's website and the interview deserves attention. The website says context layer, selling to anyone connecting agents to a business, from $29 a month. The interview says digital growth agency in software, selling the takeover of repetitive growth execution. One sells foundations, the other results. The materials provide neither a launch or migration date for the current context layer nor a complete change history from the early content tool. It is unclear whether this is a pivot or two descriptions of the same thing.
One thread can still be read through it. Today's product offers a narrower answer to the question he once could not answer. The semantic layer rejects unknown metric names, read-only partners cannot change budgets, and every run leaves a receipt. These actions replace “automate anything” with “this task, under these rules, within these permissions, completed and inspectable.” The materials do not yet connect those boundaries to willingness to pay. But at least the answer is no longer “anything.”