What are the Substreams agent skills?
Updated August 2026
The Substreams agent skills are five open-source knowledge packages that give AI coding assistants real expertise in building Substreams. They cover manifests, Rust module patterns, protobuf schema design, SQL sink schemas, production deployment, and testing.
They are published by StreamingFast under Apache 2.0 at github.com/streamingfast/substreams-skills, and work in Claude Code, Cursor, and VS Code.
Why do assistants need a skill at all?
A general-purpose coding assistant knows Rust. It does not know Substreams conventions — how a substreams.yaml manifest is structured, when to use a map module versus a store module, how protobuf schemas map to module outputs, or which crate versions actually work together on the current toolchain.
Without that knowledge it produces plausible code that does not compile. The skills supply the missing domain expertise, including the specific gotchas that waste the most time.
What does each one cover?
| Skill | Scope |
|---|---|
substreams-dev | Manifests, map/store/index modules, protobuf schema design, performance optimisation, debugging |
substreams-sql | Database changes (CDC), relational mappings, PostgreSQL indexing strategies, ClickHouse materialised views and time-series patterns |
substreams-sink | Consuming Substreams output in Go, JavaScript, Python, or Rust — cursor management, reorg handling, gRPC streaming |
substreams-sink-deploy | Running sink binaries in production — sink selection, DSN formats, backfill plus live tailing, monitoring, restart safety |
substreams-testing | Unit and integration testing with real blockchain data, performance benchmarking, CI/CD pipelines |
Installing substreams-dev pulls all five.
Are they any good?
They have been evaluated publicly rather than just described. In a published test, an agent using these skills built 14 Substreams projects from plain-English prompts across Ethereum and Solana: 100% built, 100% ran, and 12 of 14 matched a golden reference byte for byte.
The same document records the failure modes — vague prompts, non-standard ERC-20 tokens, inconsistent proto field naming when no schema is specified. The repository also ships 16 worked examples, including two deliberate cautionary tales.
How do I install them?
For Claude Code, add the marketplace and install the plugin — claude plugin marketplace add streamingfast/substreams-skills, then claude plugin install substreams-dev@streamingfast-substreams. Run /skills to confirm all five appear. For Cursor, clone the repository and add the skill directory path in settings. VS Code 1.107+ supports them as an experimental feature.
Full setup instructions for each editor live in the repository's guides/ directory.
Can I contribute to them?
Yes. The repository is open source and includes a SKILL_DEVELOPMENT.md guide for adding new skills, plus a validation script you can run against your own.
Frequently asked questions
What licence are the skills under? Apache 2.0, the same as Firehose and Substreams themselves.
Do they work with assistants other than Claude? They are documented for Claude Code, Cursor, and VS Code. Because they are open-source markdown knowledge packages, they can be adapted to other assistants.
Do I need the skills to use Substreams? No. They help you build packages faster, but you can reuse packages from the registry or write modules by hand without them.
Install from github.com/streamingfast/substreams-skills, and get an API key at thegraph.market — no personal information required.
Related: Can I build a Substreams package without knowing Rust? · How do I build a blockchain indexer with an AI coding assistant? · How reliable is AI-generated blockchain indexing code?