Skip to main content
CatalyX Daml Coding Assistant

AI coding assistant that actually knows Daml

Write Daml in VS Code with completions from a model trained on Daml. Fewer round trips to the compiler, less time in the reference docs.

Payment.daml
1
2
3
4
5
6
7
8
9
10
11
12
13
template Iou
  with
    issuer : Party
    owner  : Party
    amount : Decimal
  where
    signatory issuer
    observer owner
 
    choice Transfer : ContractId Iou
      with newOwner : Party
      controller owner
      do create this with owner = newOwner
Tab to accept
Free to install
Free to self-host
Funded by the Canton Development Fund
10+ years of Daml experience

Daml smart contracts with less guesswork

There is very little public Daml code, so general-purpose models guess at it and you spend your time correcting the guesses.

CatalyX uses its own model, trained on real Daml. You get completions that already know how templates, choices and signatories work, running on your own hardware so your code stays private and you are not paying a third party per token.

Available now
Daml autocompletion

Inline completions for templates, choices and signatories, from a model trained on Daml rather than inferred from Haskell.

Coming soon
Code generation

Describe a contract in plain language, or hand over tests you already have, and get implementation files. Called over MCP by the agent you already use.

Coming soon
Test generation

Daml tests from an implementation, scored on whether they catch real bugs rather than whether they run.

Runs on your own hardware

The model behind autocompletion today is 7B with open weights, small enough to run inside your own network instead of calling someone else's API. Free for internal use, deployment docs included.

Your code stays private

Nothing is collected by default. Proprietary Daml stays on your machines unless you opt in.

Benchmarked in public

Evaluation code, test data and results on GitHub under Apache 2.0. Check the numbers rather than trusting them.

What gets delivered, and when

The roadmap under our Canton Development Fund grant. Each milestone ships its benchmark before the tool that has to beat it.

01
Daml autocompletion: VS Code extension and open-weight model
Delivered, Aug 2026

Fill-in-the-middle completions in VS Code from a 7B model trained on real Daml. Beats GitHub Copilot on exact-match accuracy across Daml 2 and Daml 3.

VS Code extensionOpen-weight model, self-hostableThe FIM benchmark, Apache 2.0
02
Daml code generation
Sep 2026

Your agent keeps the reasoning and hands the Daml to a fine-tuned model over MCP, both to write implementation files from scratch and to correct the ones that hit compilation errors or fail their tests. Less time and token spend on getting Daml details right.

MCP server for your existing agentTrained with execution feedback (RLEF)Code-generation benchmark, published first
03
Daml test generation
Oct 2026

Generates and refines Daml tests, either from an implementation file or from a description of how the contract should behave. Served over the same MCP connection as milestone 2, so there is nothing new to set up.

Same MCP connection as milestone 2Mutation-style scoring, not just compile-and-runTest-generation benchmark, published first
04
Federated learning system: on-premise client and aggregation server
Q1 2027

Train locally, share weight updates instead of code. The shared model improves from repositories no crawler reaches, and raw Daml never leaves your premises.

On-premise federated client, opt-inSecure aggregation, differential privacy under reviewDiscounted access for participants
Fill-in-the-middle autocompletion

Benchmarks

Exact-match accuracy, higher is better. A span is masked out of a real Daml file and the model reproduces it. Scored on exact match, plus Daml syntax validity.

Daml 3
daml-fim-7b
32.7%
+4.3 pp
GitHub Copilot
28.4%
Daml 2
daml-fim-7b
20.9%
+3.1 pp
GitHub Copilot
17.8%
Combined
daml-fim-7b
25.0%
+3.5 pp
GitHub Copilot
21.4%

Full test split, 7,493 examples, no sampling. The fine-tune is cross-validated: each repo is scored by a fold model that never trained on it, so nothing is measured in-domain. Deltas carry 95% confidence intervals and paired McNemar significance.

Frequently Asked Questions

Common questions about Daml and AI coding assistants.

Daml is a smart contract language for multi-party applications, used on the Canton Network and other ledgers. It describes who can see and do what with a contract, so privacy and authorization are part of the contract rather than added afterwards.

Not reliably on its own. General-purpose models have very little Daml in their training data, so they produce code that looks correct and often fails to compile. On our published autocompletion benchmark, off-the-shelf completers perform worse on Daml, and our Daml-trained model is measurably better. A separate benchmark for whole-file writing is in progress.

Yes. The CatalyX Daml Coding Assistant gives you Daml autocompletion in VS Code from a model trained specifically on Daml. Code generation and test generation follow, both as MCP tools.

On autocompletion it beats GitHub Copilot on our published benchmark. That is one task and one baseline, and the claim stays there. Frontier models stay far ahead on general reasoning, so these tools are built for those models to call. Worth knowing: Nobody uses a frontier model for autocompletion, GitHub Copilot included. Completion has to return in milliseconds at low cost, so the job calls for a small fast model. CatalyX’s model is small, fast, and trained on Daml.

A 7B model trained on Daml beats a general-purpose assistant at Daml completion, and it fits on hardware you control. Frontier models keep the advantage on general reasoning, which is why code and test generation ship as MCP tools those models call rather than as replacements for them.

Training data. There is orders of magnitude more public Python and Solidity than public Daml, and most production Daml sits in private repositories. Models fall back on the nearest language they know well and inherit the wrong assumptions.

Yes, free for internal use. Weights are published under BUSL-1.1, which permits self-hosting and prohibits reselling the model as a competing hosted service. Deployment documentation is included.

The benchmark is under Apache 2.0. The model is open-weight under BUSL-1.1, which publishes the weights for self-hosting with limits on commercial resale.

Only if you opt in. Nothing is collected by default.

Training and evaluation data includes contracts upgraded in an SCU-compatible way. SCU compatibility is easy to get subtly wrong in generated Daml, so review output here with care.

Part of the CatalyX product suite

CatalyX Daml Coding Assistant is one part of a suite for complete blockchain operations orchestration: validator management and deployment, application discovery and distribution, tokenization and the infrastructure underneath.

Become a Canton (Super) ValidatorBuild DAML ApplicationsJoin the Canton FoundationTokenize Real-World Assets on CantonEarn Network Rewards in Canton Coin