toolang
A language and runtime for agents and humans.
# no installation needed
uvx toolang https://toolang.ai/dev.too chatuv tool install toolang
# or: pip install toolangPrompts and loops leave much of how work gets done to the model—too coarse when finer control is needed. SDKs provide that control, but bury intent in workflow code and create barriers for non-developers.
Toolang lets you express know-how in a small subset of natural language—familiar to humans and agents, precise enough for a runtime to execute.
Agic — The basic unit of agentic programs
An agic is Toolang’s counterpart to a function in traditional programming. It formalizes an agent loop as a distinct unit of work, together with its inputs, output, context, instructions, and permitted resources.
Defining and naming these units further enables us to reuse and compose them to tackle complex tasks.
agic search_web(_: Text):
tools = web/*
recall = none
Research question:
{{ _ }}
Search for reliable sources.
Summarize the key findings.
Include a URL for each source.Flow — Shareable, executable know-how
A flow organizes agics into an explicit method for complex, long-running work. Written in a language humans and agents can read and write, it gives the runtime a precise structure to follow while leaving room for each agic to reason and act.
These methods can be shared, adapted, and refined by humans and agents, making practical know-how reusable across tasks and teams.
flow research(_: Text):
let topic =
{{_}}
## Expand the research question into diverse search queries
scatter 6 using expand_queries
## Search the web for each query
map using search_web in 4 lanes
## Keep evidence bundles that answer the research task
keep if is_relevant
## Prioritize the strongest evidence
sort descending by relevance
keep first 8
## Extract source-backed findings from each evidence bundle
map using extract_findings in 4 lanes
## Synthesize the final research brief
gather using synthesize_reportCaps — Composable agent primitives
Caps are the reusable pieces an agent is built from. Beyond skills, there are psyches to shape behavior, services to reach external tools through MCP, and prompts for reusable templates.
They can be defined inline or referenced from a remote source in a .too file—which can then be cloned into an agent or run directly from a URL.
with psyche briceyan/concise
psyche source_backed:
Support factual claims with sources.
Be clear about what is uncertain.
agic:
psyches = concise, source_backed
tools = web/*
Help the user research a question.
Compare sources and explain findings.Meet your Toolang agent
Toolang works out of the box. Create your agent in seconds with toolang new <NAME>.