AniSpark

What They Build

AniSpark was founded in Dublin, Ireland in August 2025 by former ByteDance technical leadership, and is registered there as AniSpark Limited (CRO no. 795506) [2][1]. It closed its Seed round in December 2025 at a $5M valuation with investors across the UAE and Hong Kong [2]. The company ships three product lines over shared self-hosted infrastructure [1]:

• InstClaw Systems — the infrastructure line: a deep-customized OpenClaw agent harness with an advisor architecture for task distribution and supervision, a proprietary sandbox engine for tool execution, and elastic harness deployment; completed February 2026 at $10K MRR [1][2]. As a service, InstClaw provisions managed OpenClaw instances in seconds over messaging channels (Telegram, Lark, WeChat) at $1–10/day plan tiers, with sandboxed browser automation, scraping and coding-agent tooling pre-installed, and full support for air-gapped, corporation-level deployment on the customer's private resources and models [3].
• Persona — a real-time avatar stack: a low-cost 2D generative avatar SDK first released in October 2025, and Persona 2.5 (April 2026) with fully 3D-rendered avatars, real-time voice, transparent-layer rendering and a sub-$0.01-per-minute cost point [2][1].
• Square — a marketplace for AI best practices ("hire" researcher, engineer, and growth agents) [1].

The company runs self-hosted servers with content-auditing capability for enterprise visibility [1]. The timeline that matters to our thesis ends in February 2026: InstClaw complete, at revenue, before the category had a name [2].

Since then, clearly dated: in March 2026 AniSpark formed a strategic partnership with Ghisha AI, integrating its sandbox infrastructure into Ghisha's Launchpad platform for unified model management [2] — the transition from selling a tool to being a component another vendor's platform depends on. In April 2026 it launched Persona 2.5 [2]. Its stated next milestones are fully autonomous multi-agent collaboration, real-time avatar-driven workflows, and AI-native company infrastructure [2].

Why We Invested

The firm's own position. Our February 2026 research documented two things at once: LLM agents crossing into production quant workflows (factor-mining agents, benchmarked trading agents), and the now-concrete operational risk of building those agents on hosted AI — models retired on the provider's schedule, incentives changed by monetization, regulators moving at product speed. The investable conclusion was that the durable layer is the one that contains and executes agents on infrastructure the customer controls. A sandbox is the primitive that conclusion reduces to: an execution boundary the agent cannot argue its way past, supervised by an advisor architecture rather than trusted objectives. AniSpark had built exactly that — and reached paying revenue with it in the same month our note went to press — which is why the round is disclosed at the bottom of that report rather than as a standalone announcement.

Two further points weighed in. First, the cap table: AniSpark's Seed investors sit across the UAE and Hong Kong [2] — SteadyHash's own two centres of presence. Second, the founding team's ByteDance systems pedigree matched the infrastructure-not-demo character of what they shipped: a harness at revenue, not a framework in a repo.

How It Maps to Our Research

Disclosed at the bottom of our February 2026 note, Foundation Models Enter the Quant Stack — the report whose findings explain the round:

• The month's quant sections documented FactorMiner's self-evolving LLM agents for alpha discovery and the first serious **benchmarking and behavioral-consistency validation of LLM trading agents** — agents moving from demos toward deployment, which is precisely when execution containment starts to matter.
• The month's AI sections documented **GPT-4o's retirement and ads arriving in ChatGPT** — the platform-risk surface (lifecycle, incentives, regulation) of hosted AI made concrete. AniSpark's answer is structural: self-hosted harness, proprietary sandbox, air-gapped deployment support [1][3].
• The report's cross-cutting signal argued for reproducible, version-pinned, self-controlled model infrastructure. InstClaw is that argument productized [1][3].

Risks We Accept

The firm's own judgement. AniSpark's avatar and marketplace lines (Persona, Square) are consumer-adjacent, outside our thesis, and compete for the same engineering attention — we priced the persona business at zero and accept the focus risk. The harness category is young and OpenClaw-adjacent tooling is open source; the defensibility we underwrote is the proprietary sandbox engine and the operational track record, not the harness shell. And since March 2026 the company's infrastructure runs inside Ghisha's control plane, which concentrates our exposure:

Disclosures

Sources

1. AniSpark — Platform (homepage — Persona 2.5 capabilities and cost point, InstClaw Systems advisor architecture / sandbox engine / elastic deployment, Square marketplace, Trust & Governance self-hosted servers and content auditing, AniSpark Limited Dublin registration CRO 795506) — https://www.anispark.ai/ (accessed 2026-07-31)
2. AniSpark — Roadmap (Dublin founding Aug 2025 by former ByteDance technical leadership; first 2D avatar SDK Oct 2025; Seed at $5M valuation Dec 2025 with UAE and Hong Kong investors; InstClaw completed Feb 2026 at $10K MRR; Ghisha AI partnership Mar 2026 integrating the sandbox into Launchpad; Persona 2.5 Apr 2026; stated next milestones) — https://www.anispark.ai/roadmap (accessed 2026-07-31)
3. InstClaw — Professional OpenClaw Instances (managed OpenClaw hosting over Telegram/Lark/WeChat, $1–10/day plan tiers, sandboxed Chromium/scraping/agent-browser tooling "in SandBox, by InstBox", Claude Code integration, air-gapped enterprise deployment on private resources and models) — https://instclaw.com/ (accessed 2026-07-31)