Agents and tools
Run bounded tool loops with typed inputs, callbacks, and recoverable tool errors.
Lisp·LLM primitives·Rust runtime
A Lisp with LLM primitives.
Sema is a Scheme-like Lisp with built-in agents, typed tools, budgets, replay, and tracing. It runs on a Rust bytecode VM and builds to one standalone executable.
$curl -fsSL https://sema-lang.com/install.sh | sh
Windows, Homebrew, Cargo, and other install methods MIT · macOS, Linux, and Windows
The defining difference
The whole agent fits on one screen.
Model calls, tools, conversations, and agents are language values. The runtime owns the repetitive work around them, so the source stays focused on the program you meant to write.
- Tools have schemas.
deftoolconnects a description and input schema to an ordinary Sema function. - The agent loop is built in.
defagentdeclares the prompt, tools, and turn limit;agent/runhandles calls and tool results. - Limits are scoped. Token caps and known-price cost limits apply to every model call inside
llm/with-budget. - Runs can be replayed. A cassette records provider traffic once, then reuses the exact responses in local development and tests.
(deftool get-weather
"Get weather for a city"
{:city {:type :string}}
(lambda (city)
(http/get f"https://weather.test/${city}")))
(defagent guide
{:system "Give a short forecast."
:tools [get-weather]
:max-turns 3})
(llm/with-cassette
"weather.jsonl" {:mode :auto}
(lambda ()
(llm/with-budget
{:max-cost-usd 0.10}
(lambda ()
(agent/run guide
"Weather in Oslo?")))))The language
Small enough to learn. Complete enough to use.
Sema keeps Scheme's lexical scope, functions, macros, and tail calls, then adds the data forms people expect in current programs.
(define tickets
[{:id 101 :priority :high}
{:id 102 :priority :low}])
(define urgent-ids
(->> tickets
(filter
#(equal? (:priority %) :high))
(map :id)))
(match urgent-ids
([] "nothing urgent")
([id & _] f"start with ticket ${id}"))- Scheme core
- Lexical scope, first-class functions, tail-call optimization, quasiquote macros, and structured exceptions.
- Clojure-flavored data
- Keywords, maps, vectors, short lambdas, threading macros, destructuring, pattern matching, and f-strings.
- One runtime
- REPL, formatter, LSP, debugger, notebook, MCP server, compiler, and executable builder ship in the
semabinary. - Designed for generated code
- A uniform syntax and a concise agent guide give coding agents a small, stable target. Read the guide →
The runtime, in one screen
Control is part of the program.
These forms wrap ordinary functions. They compose, nest, and clean up when the scope exits.
(llm/with-budget limits f)cap tokens or known-price cost for a block(llm/with-cache {:ttl 3600} f)reuse responses during development(llm/with-cassette path opts f)record and replay provider traffic(llm/with-fallback providers f)try an ordered provider chain(with-span "ingest" attrs body...)emit OpenTelemetry spans around a run(checkpoint :result value)journal work and resume completed phasesWhat it is for
From one model call to a complete workflow.
Start with a script. Add tools, structure, persistence, or an interactive notebook without changing the language or runtime.
Agents
Structured extraction
Turn unstructured model output into validated maps, lists, and scalar values.
Journaled workflows
Split long work into phases, record checkpoints, bound fan-out, and resume a run.
Notebooks
Develop in shared-state cells, inspect rich output, then run the same code as a script.
Then ship it
A script in. One executable out.
sema build traces imports, bundles declared assets, and writes a standalone executable. The target machine does not need Sema, Cargo, or another language runtime.
- Cross-platform builds. Target macOS, Linux, and Windows from the same source.
- Capability sandbox. Restrict filesystem, network, shell, process, environment, and LLM access at the CLI boundary.
- Embeddable when needed. Run Sema inside Rust, JavaScript, or a browser through the same VM.
$ sema build weather.sema -o weather
traced imports and bundled assets
built ./weather
$ scp weather server:
$ ssh server ./weather Oslo
runs without a Sema installationProvider choice
Switch models without rewriting the program.
Sema auto-configures providers from environment variables. Your tools, agents, budgets, and traces stay the same whether a run uses a hosted model, a local model, or an ordered fallback chain.
See all providers- Built-in chat
- Compatible chat
- Embed & rerank
(llm/with-fallback
[:anthropic :openai :ollama]
(lambda ()
(agent/run guide question)))Start with one file
Build your next LLM program in Sema.
Try it in the browser, follow the quickstart, or give the concise language guide to your coding agent.
curl$curl -fsSL https://sema-lang.com/install.sh | sh
brew$brew install helgesverre/tap/sema-lang
agent$curl -fsSL https://sema-lang.com/docs/for-agents.md >> AGENTS.md