Jaseci / Jac

A programming language (Jac) + full-stack framework for building AI-integrated applications. Broader than an agent harness — it's a language — but its distinctive features target agentic AI, so it belongs in this wiki's framework neighbourhood.

What's distinctive (for agents)

• Meaning Typed Programming (MTP) — you declare typed, annotated function signatures and the compiler constructs the LLM prompt from names/types/semantics, instead of hand-writing prompt strings. A different stance on the "capability as code vs. prompt" question than agent-skills (markdown procedures): here the type system is the interface to the model. Shipped as the by language construct, implemented in the byllm module and measured against DSPy and LMQL in the peer-reviewed meaning-typed-programming (OOPSLA2 2025) — one of this spoke's few evaluated, not just claimed, framework comparisons.
• Native graph constructs — nodes / edges / walkers for agentic workflows where an agent traverses a structured state space and calls tools — a language-level take on agent-orchestration and on the state machines behind durable-agents.
• Automatic persistence (nodes on root persist with no DB code) and scale-native deploy (identical locally on SQLite, as HTTP server, or on Kubernetes). Compiles to Python bytecode, JS, and native C-ABI.

Why it belongs here

A framework for building AI/agentic apps — the language-level end of the agent-tooling spectrum (vs. the markdown-skills + CLI-harness products elsewhere here). Its MTP and graph-walker model are a notably different design point worth tracking. (Caveat: it's a general full-stack language, not exclusively an agent tool; placed here for its agentic constructs.)

Related

agent-orchestration · durable-agents · agent-skills · agentic-coding-harness · byllm · meaning-typed-programming

Re-verified 2026-08-04 — the framing changed

No version is published on the quick-guide page (jac --version is the stated check), so this pass cannot record one.

The positioning has shifted, which is the useful finding. The guide now leads with two named properties — synechic (one continuous compiler-checked medium) and topokinetic (computation moves to the data) — and the pitch "one language, one compiler, the whole stack. No glue." That is a language-and-runtime claim aimed at eliminating context-switching between frontend, backend and deployment, addressed to founders and AI engineers.

It reframes what this page files under. The spoke recorded Jaseci as an agent framework; the project is presenting itself as a full-stack language whose agent capabilities are one use case among CLI tools and web apps. Whether that is a repositioning or always was the intent, one doc read cannot say.

Re-verified 2026-09-04 (staleness queue)

The 60-day volatile window came due. The repository this page could not cite before is jaseci-labs/jac — read from the GitHub search API today: 697★, not archived, and pushed the same day (2026-09-04T15:38Z). So the project is actively developed, which the previous check could not establish either way.

The star count is the more interesting half. A language-and-full-stack-framework pitch aimed at founders, live for years, sits under 700 stars — an order of magnitude below the harnesses this spoke tracks. Recorded as a fact about reach, not about quality: adoption and merit are separate axes here, and the spoke has no evidence on the second.

Nothing in the positioning has changed since August: docs.jaseci.org still leads with synechic and topokinetic and the one-language-one-compiler claim.