fongap-labs releases Delta, a local-first AI agent

What changed

fongap-labs has introduced Delta, a local-first personal AI agent designed to assist with productivity, research and analysis, and content creation. Unlike agents that focus on accumulating models, Delta is built around completing real work with an emphasis on making the execution process controllable, recoverable, verifiable, and traceable. The agent's architecture makes boundaries explicit, defining a flow from Goal to Workspace/Session, then to Run, Sources/Citations, Capabilities, and finally Artifacts/Validation. This structure ensures that while the AI model can decide how to proceed, Delta retains authority over the work state, approvals, recovery, provenance, and validation.

Core principles guiding Delta include:

• Local first: Core work state remains local unless an external service is explicitly used by a capability.
• Human control: High-consequence actions require passing through policy and approval boundaries.
• Recoverable runs: Work can be paused, canceled, resumed, and inspected.
• Verifiable artifacts: Sources, citations, versions, provenance, and validation are integral to the work record.
• Bounded extension: New domains are integrated through Skills and Capabilities, avoiding parallel runtimes or authority models.
• Narrow model boundary: Delta provides OpenAI-compatible and Anthropic-compatible interfaces, with provider pooling, key rotation, and other external service management handled outside Delta Core.

The extension model centers on Skills, which are reusable work methods defining instructions, workflows, required capabilities, permissions, validation, templates, and optional scripts. Capabilities can be implemented as native tools, controlled Workers, MCP servers, Connectors, or external adapters. Extensions depend on Delta's public extension contracts and do not own core Runtime, Trust, Work, or state authority.

The repository follows a "small core, large framework" architecture, utilizing Rust for the runtime and TypeScript/React for the user experience. Development involves syncing dependencies, navigating to the desktop application directory, installing npm packages, and running the development server. Release intent is managed via .github/release.manifest.json, with an Action Worker handling builds, provenance generation, validation, and publishing to an external vault, ensuring release authority is not stored within the repository.

Why it matters for builders

Delta's design offers a structured approach to building AI agents that handle complex, long-running tasks. By enforcing explicit boundaries for state, permissions, and deliverables, it provides a more predictable and manageable environment for developers. The emphasis on human control and verifiable artifacts is particularly relevant for applications requiring auditability and reliability, potentially reducing common failure points in agentic workflows.

Practical impact

Developers can explore Delta's repository to understand its architecture and extension model. The project's focus on explicit boundaries and recoverable runs suggests potential for building more robust AI applications for tasks ranging from document analysis to complex research projects. Experimenting with Delta's Skill and Capability extension mechanisms could provide insights into creating modular and auditable AI workflows.

Caveats and source limits

The provided source is a GitHub repository description and README. Specific details regarding performance benchmarks, pricing, or a definitive release date are not available. The project is described as local-first, but the extent of its integration with external services and the specifics of its "external-vault" for releases are not fully detailed. The excerpt mentions "4 AI signals, 5 developer signals" but does not elaborate on what these signals represent or their implications.