Invest / Investor brief
Arcanist
Haskell tooling that makes AI-written code cheap to check, and records each check so nobody, human or model, has to repeat it.
Models now write code faster than people can review it, and studies keep finding that a large share of that code is wrong or insecure. Arcanist ships hx, one fast Haskell toolchain with reproducible builds and an interface for coding agents, and bhc, a second Haskell compiler measured in public against GHC. Its signed registry already records which package versions bhc has built. The next step is a commons where a function checked once can be cited by every human or model that needs it.
The problem
Generation became cheap. Checking did not. Every agent session re-checks the same functions and throws the result away, and every team pays for review again.
- Teams running Haskell They juggle ghcup, cabal or Stack, formatters, linters, and coverage tools. In the 2025 State of Haskell survey, 27 percent said a compiler upgrade broke their code in the last year.
- Anyone shipping AI-written code Output arrives faster than anyone can review it, and “almost right” is the most common complaint about AI tools.
- Coding agents themselves They need tools with structured errors and exit codes, not output written for a human at a terminal.
The product
- hx The whole Haskell workflow in one Rust binary: toolchain install, build, test, format, lint, coverage, and publish. Release 0.11, the 42nd since January 2026.
- Reproducible builds A lockfile pins every dependency by SHA-256 and Hackage revision, and CI fails on drift.
- Built for agents hx serves build, test, lock, and doctor tools to coding agents over MCP, with structured errors.
- bhc The Basel Haskell Compiler, a clean-slate compiler and runtime in Rust. All 221 Pandoc modules type-check and compile, and 219 test programs agree with GHC.
- The registry registry.arcanist.sh is signed, readable by plain cabal, and stores a signed bhc build result for every version. Preview, publishing by invitation.
- Everywhere Homebrew, cargo, Nix, aqua, winget, a container image, and binaries for macOS, Linux, and Windows.
In a self-reported benchmark against cabal on a small project, hx starts 6 times faster, builds cold 5.6 times faster, and rebuilds 4.8 times faster. bhc publishes its gaps as openly as its numbers: compiled programs do not yet collect garbage, so long-running services wait for that milestone.
The market
No one has sized a market for verifying AI-written code yet, and this page does not invent one. The markets around it are large and moving fast.
- USD 8.1 bn AI code assistants in 2025, projected at USD 127 bn by 2032. MarketsandMarkets
- USD 1 bn Run-rate revenue of Claude Code six months after launch. Developers pay for AI that writes code. Anthropic
- USD 6.4 bn Software development tools in 2025, growing to USD 15.7 bn by 2031. Mordor Intelligence
Why now
- 45 % Share of test cases where AI-generated code introduced a security flaw, with no improvement as models grew. Veracode
- 1.7 x More major issues in AI-coauthored pull requests than in human-only ones. CodeRabbit
- 46 % Developers who distrust the accuracy of AI output, up from 31 percent a year earlier. Stack Overflow survey via TechRepublic
- USD 1.45 bn Valuation of Harmonic, a formal-verification company, in 2025. Checking is attracting capital. Crowdfund Insider
AI labs are also buying developer tools: OpenAI announced the acquisition of Astral, the company behind uv and Ruff, and Anthropic acquired Bun. Fast, coherent toolchains have become strategic for agent platforms.
What sets it apart
- One binary for the whole workflow Six separate tools become hx commands, installable through eight channels.
- Reproducible by default Hash-pinned lockfiles, offline sync, and a CI check that fails on drift.
- Built for agents in the loop An MCP server with structured errors, plus llms.txt and AGENTS.md shipped with the code.
- A second compiler measured in public Dated, published numbers against GHC as ground truth, including the gaps.
- Verification that accumulates The registry keeps signed build results per version from a scheduled verifier, so a check done once stays on record.
Competition
| Alternative | What it does | How Arcanist differs |
|---|---|---|
| cabal-install | The standard Haskell build tool | hx drives cabal and adds toolchain install, a hash lockfile, format, lint, coverage, and MCP in one binary |
| Stack | Curated snapshots and reproducible builds | hx reads .cabal and Hackage directly, and imports Stack projects |
| GHCup | Installs GHC, Cabal, and HLS | hx reads the same metadata and installs per project, without a separate tool |
| Nix, haskell.nix | Hermetic builds | Heavier to adopt; hx ships a flake and keeps a simpler default |
| GHC | The reference compiler | bhc is an independent second compiler with a WebAssembly backend; GHC stays the ground truth |
| Hackage | The central package index | Signed, cabal-compatible, and records verification results per version |
| Lean, Dafny | Proof assistants | Complementary: Arcanist makes checks reusable records in a production language |
Business model
hx, bhc, and the registry are free and stay free. Today they are funded through GitHub Sponsors and Ko-fi, and companies can fund specific features or security reviews. The paid layers sit on top, the way Astral built adoption with free tools first.
- Private registry Verification records, policy controls, and audit logs for companies running Haskell or AI-written code.
- Verification as a service Paid verifier capacity with signed results.
- Support Contracts for teams adopting hx in CI.
- Strategic fit Agent platforms need output they can check, as the Astral and Bun acquisitions showed.
Where it stands
- 68 Releases of hx and bhc in about nine months: 42 and 26.
- 3,000+ Release downloads from GitHub and crates.io across hx and bhc.
- 221 / 221 Pandoc modules that bhc type-checks and compiles.
The registry is live in preview with its first verified package, and the scheduled verifier signs its results. hx was presented at ZuriHac 2026. There are no paying customers yet: this is an early ticket into tooling that already ships every few weeks.
Risks
- A small language Haskell has a devoted but small community. hx serves existing Haskell teams now, and the verification thesis aims at the much larger market for AI-written code.
- Compiler maturity bhc does not yet run long-lived services. hx works with GHC today, so the business does not depend on bhc.
- Open-source revenue Donations rarely fund compilers. A hosted registry and paid verification are the natural paid layers.
- One maintainer Arcanist is built by one person with heavy use of AI agents. Funding would add engineers.
What funding would do
- Garbage collection A live collector on the compiled-code path, so bhc can run services.
- Pandoc end to end The pandoc command running fully under bhc.
- An open registry From preview to open publishing, with verifier capacity beyond one CI job.
- The first commons milestone Content-addressed definitions and attestations for reproducible artifacts.
- Real GPU validation The CUDA and ROCm backends tested on hardware.
- A team and design partners Two more engineers, and a private registry pilot with companies.