Zatanna
Turns legacy software backends into APIs so AI agents can use them directly.
On the record
Checkpoint: Anthropic, OpenAI, and Google's next computer-use reliability and cost benchmark update
The call
Zatanna is racing to make legacy software agent-ready before the labs whose agents it depends on learn to do it themselves.
Zatanna takes old software systems like hospital records programs and payment systems, then finds hidden doorways that let AI assistants use them directly. Right now, AI from companies like Anthropic and OpenAI can only see screens and click buttons like humans do, which is slow and unreliable, so Zatanna's approach is genuinely faster. However, Anthropic and OpenAI are building the same ability into their AI models themselves, and if they get it working well within the next 12 to 24 months, companies may not need Zatanna anymore. The company has no paying customers yet and no official partnerships, so if the big tech companies close the gap in speed and cost before Zatanna can prove its value to real businesses, the entire reason to use it disappears.
Why
01
AI pressure: Anthropic Claude Computer Use, OpenAI CUA, Google Gemini computer-use. Frontier labs already ship native computer-use agents that could make Zatanna's reverse-engineered APIs redundant. Could AI do the core job? Yes, likely.
02
Bigger players: Composio, Arcade.dev, Nango, Merge, Browserbase, Steel.dev. Established agent-integration and browser-automation platforms already sell the same connective tissue Zatanna is building. Could a giant copy it? Yes, easily.
03
What protects it. If its reverse-engineered APIs are faster and cheaper than screen-based computer-use, agents may route through them instead of clicking.
What would change this call
Zatanna announces its first named enterprise or ERP/EHR customer within 12 months. Would show the wedge is converting into real, defensible revenue rather than a pitch.
Google's Gemini native computer-use (expected June 2026) ships with reliability matching or beating Zatanna's stated gap. Would show frontier labs are closing the reliability gap Zatanna's pitch depends on.
Composio or Arcade.dev publicly launches a legacy-backend reverse-engineering feature. Would show structurally unblocked rivals are copying Zatanna's core wedge.
Google's Gemini native computer-use (expected June 2026) ships with reliability matching or beating Zatanna's stated gap. Would show frontier labs are closing the reliability gap Zatanna's pitch depends on.
Composio or Arcade.dev publicly launches a legacy-backend reverse-engineering feature. Would show structurally unblocked rivals are copying Zatanna's core wedge.
Key numbers
- 12 to 17 point reliability gap vs native computer-use
- 12 to 24 month window before that gap likely closes
- Zero named customers or partnerships disclosed
Named risks
- Entire value prop depends on a reliability/cost gap frontier labs are actively closing.
- No named customers or independently verified traction as of July 2026.
- Core method relies on reverse-engineering third-party portals, raising ToS and legal exposure.
- Team of roughly 3-4 people with only its first outside hire posted in July 2026.
- Conflicting company identity in data providers, e.g. Pitchbook describes a different healthcare business.
Rivals, as read in July 2026
| Rival | Direction | Evidence |
|---|---|---|
| Anthropic (Claude Computer Use) | losing to | Claude Computer Use shipped Oct 2024, directly narrowing the gap Zatanna sells against |
| OpenAI (Computer-Using Agent) | losing to | OpenAI's CUA launched Mar 2025 with 38.1% OSWorld success, competing approach |
| Google (Gemini native computer-use) | losing to | Gemini 3.5 Flash added native computer-use Jun 2026, replacing Project Mariner |
| Composio | contesting | Composio offers 1,000+ toolkits and 3,000+ tools as of Dec 2025, crowded field |
| Arcade.dev | contesting | Arcade.dev pitches agent-native action runtime, published rival content Jul 2026 |
Questions people ask
- Is Zatanna profitable in 2026?
- No revenue, ARR, or profit figures have been independently reported for Zatanna as of July 2026. The company is a pre-seed, YC W26 startup with no confirmed round beyond unconfirmed aggregator estimates, and its own site only cites a self-modeled request-volume figure, not financial results.
- Could AI replace Zatanna's product?
- Yes, that is the core risk. Zatanna sells network-layer APIs as faster and cheaper than AI 'computer-use' agents, but Anthropic (Oct 2024), OpenAI (Mar 2025), and Google's Gemini 3.5 Flash (Jun 2026) are shipping native computer-use features that could close the reliability gap Zatanna depends on.
- Who are Zatanna's biggest competitors?
- Composio and Arcade.dev compete directly in agent-integration tooling, Nango and Merge extend into agent use cases, and Browserbase/Steel.dev compete on browser automation. The bigger threat is frontier labs Anthropic, OpenAI, and Google building native computer-use directly into their models.
- Who owns Zatanna?
- Zatanna is a private company founded by Rithvik Vanga (CEO), Tarun Vedula, and Alex Blackwell (CTO), incubated in Y Combinator's Winter 2026 batch. YC is the named investor; no independent press confirmation of a priced funding round exists as of July 2026.
- What does Zatanna actually build?
- Zatanna reverse-engineers the backend network requests of software lacking public APIs, such as ERPs, EHR portals, and PMS/POS systems, and turns them into stable callable APIs for AI agents, delivered via a hosted OAuth-protected MCP server at api.zatanna.ai as of July 21, 2026.
- Has Zatanna raised venture funding?
- Funding aggregators report conflicting, unconfirmed figures, one citing a $500K pre-seed led by Y Combinator and another showing a $1M headline total with a $500K body figure. No primary-source press release or filing confirming a priced round has been found as of July 2026.
Timeline
- 2026-07-21Hosted OAuth-protected MCP server goes live at api.zatanna.ai
- 2026-07-06Posts first outside hire, Founding Engineer role in San Francisco
- 2026-06-24Google ships native computer-use in Gemini 3.5 Flash, rival approach
- 2026-05-07Co-founders discuss product philosophy on Fondo's START podcast
- 2026-03-10Publishes YC Launch post titled APIs for AI agents
- 2026-02-28Site claims ~3.9 billion modeled requests processed since Mar 1 2026
- 2025-03-11OpenAI announces Computer-Using Agent, a rival approach to Zatanna
- 2024-10-22Anthropic ships Claude Computer Use beta, early rival capability
Revision history
27 July 2026
First assessed. Losing its protection. The text above is fixed as of this date and is not edited afterwards.
In the same band
Sources