Developers
Review responsibilities, compose agents and test the actions that must also be refused.
Why Touba
An agent that works is not enough. Know what it can do.
A convincing result does not reveal which access was granted, who could trigger an effect or where human approval was required. When these rules live only in prompts and conventions, enforcement depends on each integration path.
Touba connects rules expressible in its contracts to execution controls: declared capabilities, limits, human validation and traces. Model intent does not replace authorization.
Read Why ToubaOrganizations
Turn applicable rules into verifiable technical constraints with explicit responsibilities.
Users
Make information, human validation and traceability requirements of the system.
Touba Explorer
Keep your project in view
Open your project map to find its agents, tools, workflows and permissions.
Discover our three favoritestouba explore
Research mission
Understanding agentic AI and building verifiable responses
PAP and Touba produce resources that can be evaluated and reused in industrial contexts. Their primary purpose remains research: understanding the technical, organizational and societal challenges raised by agentic AI, then developing programming models, tools and evidence that respond to them.
Explore the research programPeople, teams and institutions working toward responsible, safe, secure, transparent and inclusive AI are invited to contribute to this effort.
Benchmarks
Evaluating outcomes and controls
The evaluation examines business outcomes, model-facing interfaces and enforcement of declared responsibilities. Results and measurement conditions are gathered on the Benchmarks page.
View benchmarksDownloads
Touba distributions by platform
Download the distribution for your platform, then verify the archive before installing it. Every published architecture stays available from the Hub.
Choose your platform
macOS
Processor
138 MB
sha2568f4eda4c774554347e66127be7d7ef5b6a3cb8ff2e48aa7554a6307a85e96361
Other platformsAll archives and checksums
Public resources
Public resources in the Touba ecosystem
Touba connects distinct public resources: research, language, documentation, developer experience, a capability library, and a distribution registry.
Research program
Probabilistic Agentic Programming
Public overview of Probabilistic Agentic Programming and governed agentic systems.
Open resourceLanguage and runtime
Touba
Language, compiler, runtime, CLI and SDKs. The language repository will soon be available on GitHub under Apache 2.0.
Open resourcePublic documentation
Touba Docs
Official documentation for installation, concepts, references, guides and the Hub.
Open resourceVS Code extension
Touba Language Toolkit
Editor support, syntax, diagnostics and commands for working with Touba files.
Open resourceReusable capabilities
Touba Library
Library of Touba packages following the common language contract.
Open resourceFoundation
A stable story: paradigm, contract, execution
Touba starts from a paradigm, writes authority as code, compiles toward an observable contract, then executes that contract with visible bounds.
PAP
The explicit-responsibility paradigm
A paradigm for treating agentic autonomy as a governed programming responsibility, not just as a sequence of prompts.
AIR
Agentic Intermediate Representation
The compiled contract linking .tb source, runtime, traces, CLI and Touba surfaces without letting every tool redefine the semantics.
Touba
Thinking Orchestration Using Bounded Agency
A reference language and runtime for making limits, tools, validations, refusals and execution evidence visible.
Positioning
Execution semantics for agentic authority
Touba connects declared responsibilities to execution controls. JavaScript and Python handlers implement your API calls and business logic; traces let you follow decisions and actions.
Understand ToubaQuick read
What, why, how, who
What?
Touba brings Agents as Code to governed agentic systems: .tb language, AIR, runtime, traces, manifests, locks, CLI, project server, and VS Code extension.
Why?
Current agent systems scatter responsibility, authority, memory, tools, validation and traces across prompts, middleware, logs and team discipline.
How?
Touba code is compiled into AIR, executed by a bounded runtime, connected to handlers through an explicit manifest, then observed through usable traces.
Who?
Khadim Lo, Jean-Sébastien Dessureault and Soukaina Alaoui Ismaili research agentic AI and develop Touba’s foundations, language and tools.
Available capabilities
An executable base for governed agentic workflows
Compose agents, skills and sequential or parallel workflows. Connect resources with the JavaScript and Python SDKs. Declare access, budgets and approvals, then inspect traces from each execution.