Developer tools and coding agents

Started small, grew after funding · sold 11 months after launch

Its founders wrote some of the methods that the industry already used to run models cheaply, such as AWQ for 4-bit weights. They turned that research into a paid serving API, proved the speed on a public benchmark, and partnered with a cloud company that needed exactly that speed. Eleven months after its first public post, the partner agreed to buy it.

Solo founder · 2 months after launch, revenue flat

AI coding agents can build a working page or deck, but in the founder's words their visual output often looks generic. Skillry sells finished visual styles, one task per Skill, shows real output before the sale, and adds new Skills almost every day under one subscription. When sales began on 28 September 2026, they held at about US$430 a day for 17 days.

Solo founder · 12 months after launch, revenue falling

It launched in October 2025, when Anthropic's Claude Code had made skills, slash commands and subagents a normal way to steer a coding agent, and when most developers were still writing their own. A founder who presents themself as a CTO with 18 years of experience sold a tested set as a one-time purchase with a money-back promise. Verified revenue reached US$16,040 in March 2026 and US$82,195 in total, but it fell to under US$2,000 a month by September 2026.

Solo founder · 4 months after launch, revenue falling

Said over US$36K in a monthRecords US$39.6K by end of JulyHeld

It turned a new working habit, running several AI coding agents at the same time, into a native Mac app with voice control and a strong visual style, and it charged once instead of every month. The founder sold it to his own audience through daily live streams, X and a Discord community. Stripe recorded more than three quarters of the revenue in the first seven weeks.

Solo founder · 6 months after launch, revenue growing

The compression engine already existed as a free, open-source Python tool, but it was hard to install. Garm Lucassen packaged it into a desktop app that installs with a double click, connects the coding agents by itself and removes its changes when it quits. The engine's creator suggested that he build it and keep the Headroom name. Developers who hit the usage limits of their AI plans pay a small monthly price to make those limits last longer.

Solo founder · 8 months after launch, revenue flat

AI agents that run for hours or days need an address that people can write back to, and a normal personal mailbox needs a browser login and OAuth consent screens that a program cannot click through. Robotomail sells a mailbox that a program can create and read without a person. It started in March 2026 and connects to Claude, Codex, Cursor and other agents through MCP, an API or a command-line program.

Solo founder · 21 months after launch, revenue falling

It puts an AI coding agent where a team already tracks work: the GitHub issue. Nobody has to open a new editor or tool; assigning the issue starts the work, and the output is a normal pull request that the team reviews as usual. The verified revenue shows that this idea alone did not bring many paying customers.

Started small, grew after funding · 30 months after launch

Every team that builds an agent must connect it to other software, and the hard part is not the API call but the login: running the OAuth flow, storing tokens, refreshing them and scoping permissions for each user. Composio took over that chore for all apps at once, gave the SDK away as open source, and charges for usage above a large free tier.

Solo founder · 14 months after launch, revenue growing

The founder builds the product live on YouTube, X and Twitch, four to six hours at a time on several days a week, and the title of each stream shows the current annual run rate. The viewers are the target customers: people who write software by giving instructions to AI agents. A reviewer who watched one stream says that every paying customer came from streaming and that nothing was spent on ads.

Started small, grew after funding · 14 months after launch

Running a large model fast needs performance engineers, and the founders write that only a few hundred people in the world can deeply optimize accelerator hardware. Wafer tries to do that work with AI agents, on AMD as well as NVIDIA GPUs, and to keep doing it after a deployment goes live instead of tuning it once.

Team of 2 to 10 · 16 months after launch, revenue growing

Coding agents built on large models are slow and make mistakes at one small step: writing an edit into an existing file. Morph trained a small model that does only that step and wrote its own GPU code so the model runs at more than 10,000 tokens a second. Agent builders add one tool call and their edits become faster and fail less often. The founder posted in August 2026 that revenue was US$6 million a year, as the Lean AI Leaderboard reported.

Team of 2 to 10 · sold 9 months after launch

The two founders had spent their careers building data systems at Databricks and ThoughtSpot, and they said that half of their engineering time went to debugging. They built an agent that does the search across logs and code that an on-call engineer does by hand, and they named DoorDash and Foursquare as production users when they came out of stealth in November 2025. AI Chat Daily reported in June 2026 that Elastic agreed to buy the company; Elastic announced the deal in July and completed it in August 2026.

Team of 2 to 10 · 3 months after launch

Its users were mostly beginners and part-time developers who build with AI coding tools, which in the founder's words "hallucinate" and "assume things". He had written such planning documents for his agency's clients, automated his own prompts, and showed a demo to an X audience that had read his posts on AI prompting since 2022. More than 1,800 people joined the waitlist before the product existed, and in its first months there was no free trial.

Team of 2 to 10 · 9 months after launch

The three founders were competitive hackers and professional penetration testers. In mid-2025 they bet that models that are good at code would also be good at finding security bugs. They showed it in public: a pentest report of Gumroad at launch, then zero-day bugs in widely used software, and in September 2026 a chain of bugs into OpenAI employee accounts that the press covered widely.

Team of 2 to 10 · 2 months after launch

Every AI agent that must answer from private data needs a connector for each app and a search index that stays current. Airweave sells that layer once, for many agents. The founders said in their seed announcement that it began as a fix for their own problem: agents could not search their apps. Made With Models reported in July 2025 that the founder said the company was making US$17,000 a month.

Solo founder · 12 months after launch, since closed

In mid-2024 the models could read a large codebase, but no tool moved the right files into the prompt and the edits back out, so developers copied and pasted by hand. The founder built the tool for his own Vision Pro game, gave it away through TestFlight to people who found a Reddit post, and charged only after a community of power users had formed. According to the Indie Hackers interview summarised by Startup Founder Stories, that free start turned into a 5-figure monthly recurring revenue within a year of the first tester build.

Solo founder · 19 months after launch

Developers dislike writing documentation, and it goes out of date as soon as the code changes. In April 2023 the home page sold one-click documentation of single source files at US$29 a month, according to an Internet Archive copy. By 2026 the product documents whole repositories and keeps the pages current, and the founder sold it to engineering teams, not only to single developers. Starter Story reported US$7,200 a month in November 2024; no later dated figure was found.

Team of 2 to 10 · 6 months after launch, revenue growing

Two childhood friends started it at Entrepreneur First in Bangalore in 2025, when many companies had put chat and voice agents in front of customers but had no tool that read the conversations themselves. An aggregator of startup lists reported in August 2026 that it had passed US$50,000 in annual recurring revenue by its April 2026 demo day. It then entered Y Combinator and added custom models trained on customer traces.

Solo founder · 1 months after launch

The founder had an audience of builders before launch, and the product is a thin layer over agents that the user already pays for, so it has no model costs. Its message is aimed at people who do not code: say what you want in a sentence, and the agents build it.

Team of 2 to 10 · 4 months after launch

The four founders worked on GPU orchestration and observability at Amazon and AWS. Their about page says they saw the same problem at every large company: some teams wait for GPUs while the reserved GPUs of other teams stand idle. Their launch post said that 30 to 60% of reserved GPU capacity sits idle. Chamber sells the internal platform that large companies build for themselves to companies that cannot build it.

Team of 2 to 10 · 6 months after launch

A weekend experiment that stored text embeddings inside video frames went viral on GitHub in 2025, and the founders say thousands of developers started building on it. They rewrote it as a file format in Rust, kept the core free and open source, and charge for larger memory files and for enterprise deployment.

Solo founder · 2 months after launch

In early 2026 Codex ran on the developer's own Mac, so a developer who left the desk could not answer the agent's questions or correct it. The founder built the first version in one night in March 2026, on OpenAI's official open-source Codex App Server, before OpenAI had an iPhone app of its own. He published the Mac part as free code on GitHub, so developers could check the security, and sold the convenience of a hosted connection in the App Store. MRR Story reported in May 2026 that this brought about US$10,000 in the first month.

Solo founder · 24 months after launch

CSS keyframe animations are short pieces of code that most web developers write by trial and error. A tool that turns one sentence into working code saves that time. The founder added a directory of 2,635 ready animations and prompts, which gives search engines many pages to index, and later added paid features built on Claude.

Solo founder · 4 months after launch

Coding agents such as Claude Code became able to run long tasks in 2026, but they run on one person's laptop and stop when the laptop closes. CRHQ puts them on an always-on server with a database, a scheduler and a shared skill library, and handles the server work. The founder runs his own SaaS products on two of these servers, and the company publishes what those agents did in a week.

Team of 2 to 10 · 7 months after launch

A coding agent that runs a long task stops at every risky step and waits for a yes, often in a terminal window nobody is watching. Pushary moves that question to the phone. The founder built it for his own agents, after a first version as a general push notification service.

Team of 2 to 10 · 5 months after launch

The two founders, childhood friends who worked in AI infrastructure at Microsoft, ThirdAI and Meta AI, wrote that they built the first product because Claude Code kept them waiting at the desk for its questions. After two launches for coding-agent users, they changed the product in August 2026 to the infrastructure that companies build for their own agents, and positioned it against Anthropic's own hosted service, Claude Managed Agents.

Team of 2 to 10 · 0 months after launch

Many product teams record every user session but nobody has time to watch the recordings, and most users who hit a bug do not report it; they leave. The founder's launch post says this directly. Lucent sells the watching: a model reviews each recording and only the problems reach the team.

Team of 2 to 10 · 13 months after launch

A voice agent must answer within a fraction of a second, or the conversation feels broken. The founder writes that the slow step is almost always retrieval: each lookup crosses the network to a cloud database. Moss moves that lookup into the agent's own process.

Team of 2 to 10 · 6 months after launch, since closed

The two founders had built and scaled CI systems at Docker and Dagger for more than ten years. They wrote that coding agents make teams produce more code, so more builds run and fail, and that the people who write the code with AI tools are not the ones who debug the failed builds. Mendral was the specialist agent for that work. In July 2026 the team joined Anthropic and the hosted product was wound down.

Team of 2 to 10 · 13 months after launch

MCP became the common way for AI agents to call outside tools, but running MCP servers in production meant Docker setups, OAuth sign-in flows and no logs. The founders say they first built that infrastructure for themselves, then turned it into a hosted platform with hundreds of ready servers. In 2026 the pitch moved from developers to companies that want to control which agent may do what.

Team of 2 to 10 · 0 months after launch

Chip design teams spend much of their time writing and debugging hardware code and testbenches by hand, with tools that the founders say have changed little since the 1990s. The two founders had worked on custom circuits and processors and on fine-tuning language models, and built a coding assistant for this one field.

Started small, grew after funding · 8 months after launch

Developers dislike writing and repairing test scripts, and many companies test by hand. The founders argued in 2023 that AI coding tools would produce more code, and so more code to test. In November 2026 the company repeated the argument: AI assistants multiply the changes a team ships, and fewer of them are read line by line.

Solo founder · 6 months after launch

AI coding tools let engineering teams ship more changes than sales, support and product managers can follow. The founder had spent years at legal and government technology startups explaining code to non-technical colleagues, and built a tool that does this translation automatically, from the code rather than from tickets or old documents.

Team of 2 to 10 · 10 months after launch

An agent can be allowed to read a customer record and allowed to send an email, and still leak data by doing the two in that order. Kyvvu's founders argue that identity systems and content filters check one step at a time and so cannot see this. Banks, insurers and hospitals in the Netherlands also need records of what their agents did for the EU AI Act.

Team of 2 to 10 · 21 months after launch

Inspection models fail on the defects that matter most because those defects are rare, so the training set has only a few examples of them. Synthgen generates more examples of exactly that class from a dozen real ones, writes the labels at the moment the defect is placed, and promises to judge the result only on real images that the generator never saw.

Solo founder · 10 months after launch

Apple rejects many submissions, and a rejection costs a developer days of waiting for a second review. A general chat model can discuss the guidelines, but AcceptMyApp retrieves the exact guideline sections, scores each risk and keeps a history per app. The founder runs several iOS apps himself, so he builds for a problem he meets in his own work.

Team of 2 to 10 · 6 months after launch

After an enterprise buys software, a services team often needs months to set it up for that customer. Lab0's own page says a typical implementation takes about six months with about nine people, and that its agents bring it to a few weeks with about three people. The founders say they saw this work from inside Oracle and in earlier startups.

Team of 2 to 10 · 12 months after launch

Coding agents such as Claude Code, Cursor and Codex can write an app, but the founders say they struggle to run, deploy and maintain it, because hosting tools were made for people who click through dashboards. Specific gives the agent a command-line tool, documentation inside the binary and an account the agent can create without a human.

Solo founder · 13 months after launch

Coding agents can now write front-end code, but they usually produce one answer to one prompt. Rivet's own story page says designers 'have to leave great ideas unexplored' because there is not enough time to prototype every direction. Rivet uses the coding agent to try many directions at once, so the cost of one more idea is machine time instead of a designer's day.

Solo founder · 23 months after launch

In many teams a finished feature waits days for translations from an agency, or ships in one language first. The founder, a developer for many years, saw this pattern again and again. Localhero.ai moves the translation into the normal code review step, so the feature and all its languages are merged together.

Team of 2 to 10 · 18 months after launch

The founder, a software engineer, saw releases wait for missing translations at his own employer. Translation management platforms keep their own copy of the texts, an editing interface and an asynchronous translator workflow; doloc keeps the repository as the only copy and translates inside the build, using the existing translations as context for terminology.

Started small, grew after funding · 28 months after launch

SAP has set 2030 as the deadline for customers to move to its newer S/4HANA platform, and decades of custom code make that move slow and expensive. The co-founders started from research on AI that writes code and moved to SAP because, in the company's words, every dollar spent on SAP software drives about seven dollars of services. A co-founder who co-invented SAP HANA gave the AI team the domain knowledge.

Solo founder · 9 months after launch

In 2026 many developers used several AI agents at once, and each agent needed its own sign-in to every service. The launch post names three problems: each agent wants its own OAuth sign-in, most connectors allow one account per service, and an agent that holds a token can do anything the token allows. Executor moves the accounts and the permission rules out of the agents into one place.

Team of 2 to 10 · 6 months after launch

The founder wrote that the idea came from the failure of the company's first product, an AI agent that wrote database code: the code looked right, but there was no safe place to test it, so customers would not run it in production. The launch post says that at many later-stage startups the founder talked to, most engineers now use Claude, Codex or Cursor for migrations and backfills, and a shared staging database cannot give each agent its own copy.

Team of 2 to 10 · sold 14 months after launch

It trained its own small models for one specific task, ranking documents by relevance, and published benchmarks against larger providers. Notion used its zerank-2 reranker in Notion AI's search, said that the reranking step became 85% faster, and then bought the company to get the team.

Team of 2 to 10 · sold 20 months after launch

It was live by December 2024 with one command that installed a server into Claude Desktop, and in the buyer's words it was built on the idea that finding and running an MCP server should take one click. Arcade.dev, which sells security and governance for the tool calls of enterprise agents, bought it for that developer reach: its post says enterprises want both the developer speed of Smithery and the confidence to run those tools in production.

Started small, grew after funding · sold 44 months after launch

Two founders started Chroma in 2022 on tools for curating training data, and turned to inference when ChatGPT came out. They released the open-source database in early 2023, the moment developers wanted "ChatGPT for my data", and it became one of the most downloaded tools for vector search. A hosted cloud version with usage-based prices followed in 2025. Two of its early investors lead Replit, which agreed to take in the company in October 2026.

Team of 2 to 10 · sold 16 months after launch

Several agents that share one checkout of a repository edit the same files on disk. Conductor gave every agent its own working folder and branch through Git worktrees and put all the agents in one window, so one developer could run several tasks in parallel and review them one by one. The founder wrote that at the launch in July 2025, 'Writing most of your code with agents (let alone a team of coding agents) felt crazy'.

Team of 2 to 10 · sold 6 months after launch

The founders had built and run large production systems at Cloudflare, Segment and Twitch, and they aimed at a gap that coding agents made wider: code was written faster than teams could check it in production. Cursor, which makes an AI code editor and coding agents, took the team to add production checks to its own agents and said Change Monitors would come to Cursor.

Solo founder · 17 months after launch

Clients describe website problems in vague emails and long screen videos, and the founder's studio says it spent more than two hours per sprint turning such messages into developer tickets. GiveFeedback records what the client said and what the client did at the same moment, so nobody has to rewatch a video to guess which button was meant. It hands that record to a coding agent that already knows the codebase, instead of writing the tasks itself.

Team of 2 to 10 · sold 37 months after launch

Developers who put language models into products needed a way to stop bad answers before users saw them, and Guardrails gave them a free library for it. The company then sold a managed version to large companies and a testing service on top. Harvey bought it because its own legal AI agents need the same kind of testing before law firms trust them with client work.

Team of 2 to 10 · sold 24 months after launch

Its founder was building an AI agent meant to drop unused database tables and wanted a human to approve that step, so he built approvals for agents as a product. His essay "12-Factor Agents" in April 2025 and later talks on context engineering for coding agents drew a large audience of engineers, and the company followed that audience to tools for coding agents. A buyer that builds coding agents for enterprises wanted its team features.

Team of 2 to 10 · 8 months after launch

Agents that read web pages, install packages and write their own code can leak any key they hold, through prompt injection, a malicious tool or a log line. The founders, who worked in network security before, argued that the fix is to keep the key out of the agent entirely, as zero-trust network access does for people. NanoClaw, an open-source agent framework, made the gateway its default credential layer in March 2026, and users then asked for the agent itself.

Team of 2 to 10 · sold 22 months after launch

One of the founders had done this checking by hand every week while she led AI products at Instacart, so the two founders built the tool that she had needed. Rootly, a company that sells incident-management software and ships its own AI agents, bought it in July 2026 to check those agents and to sell agent reliability to its customers.

Team of 2 to 10 · sold 9 months after launch

AI coding tools of 2024 worked well on small projects but needed the developer to pick the right files by hand on large ones. Jolt picked them itself, and its pricing page said it was used on codebases of more than ten million lines. Ramp, a finance company with about 1,200 employees, wanted exactly this skill for its own engineers and hired the team.

Team of 2 to 10 · sold 35 months after launch

The founders wrote in their 2023 launch post that AI tools for developers were "cumbersome black boxes" and that developers were "stuck copy/pasting from ChatGPT". Continue was open, configurable and worked with any model, so a developer whose employer did not allow code to leave the company could run it with a local model. Cursor, which makes an AI code editor, bought the company in June 2026; no price and no reason for the purchase were found in the sources we could read.

Team of 2 to 10 · sold 28 months after launch

The founder showed in 2024 how batch tests over real customer questions find the places where an agent fails before customers see them. Mitratech Legal, a legal-software company with decades of customer data, said the purchase brings agents to its customers significantly faster than its own plan.

Team of 2 to 10 · 16 months after launch

The founders wrote mcp-use in early 2025 because, in their words, MCP worked only inside a few closed apps such as Cursor and Claude Desktop. Developers adopted the free library quickly, and then asked how to run their servers in production: deployment, sign-in, monitoring and store review. Manufact is the paid cloud for that step, and in 2026 its pitch moved to companies that want their product listed inside ChatGPT and Claude.

Team of 2 to 10 · 15 months after launch

An agent that signs up for services, buys things or talks to people needs an email address of its own. The founders' launch post said that Gmail and similar providers lack an API for creating inboxes, charge a subscription per inbox and limit sending. AgentMail sells the inbox as one API call, the way a cloud provider sells a server.

Team of 2 to 10 · sold 22 months after launch

Before a new device ships, engineers must prove by measurement that it meets its specification, and Glue's launch release says that setting up these tests can take from a few weeks to 18 months. Glue started with the customer contacts of Tektronix, an instrument maker, and the staff of a venture studio, and its product fitted the AI plans of Emerson, a large instrument group, which bought it 22 months after the launch.

Started small, grew after funding · sold 50 months after launch

Large companies wanted to put language models and agents into production, but their security teams, boards and lawyers needed proof that the systems could not be tricked into harmful answers or leak data. Enkrypt sold that proof as software, and in August 2026 Anaconda, the company behind a Python and data-science platform, bought it to add a security layer to its platform, after it had bought the coding-agent company Kilo Code in July 2026.

Team of 2 to 10 · sold 38 months after launch

Teams that ship chatbots get thumbs-down ratings and complaints every day and no tool that says why the answer failed. The founder had met the problem herself when she delivered AI projects. ActiveCampaign's own AI engine holds thousands of agent conversations a day and needed the same measurement, so it bought the company and its team.

Team of 2 to 10 · 31 months after launch

Many AI companies in health care sell agents that must do work inside payer portals and EHRs, and those systems are built for people who click, not for programs. Each company would otherwise write and maintain its own browser scripts. CloudCruise sells that layer to them: the founders' launch post of April 2024 says zero-shot web agents were unreliable for long workflows, so they model every workflow as a graph of steps instead.

Started small, grew after funding · 8 months after launch

Ordinary test scripts name the exact page elements to click, so they break each time a page changes, and many online shops check their launches by hand. Spur's tests describe what the user wants to do, so they can keep working when the page changes, by the founders' launch posts. Its customer stories for DocuSign, Abercrombie & Fitch and Bombas report manual checking replaced or shortened from days or hours to a review of flagged results.

Team of 2 to 10 · 47 months after launch

Its founder ran a software agency for eight years and says that poor requirements are where many software problems start. He released a structured requirements tool in late 2022 and offered AI generation in every plan by February 2023; he says that GPT-3.5 and GPT-4 then significantly improved what Userdoc could offer its customers. Customers then asked for documentation of the old systems they already had, which became Code-to-docs.