AI tools
Lumen Agent Platform
A local-first AI workspace that brings model chat, systematic testing, persistent agents, deep research, evidence pipelines, local tools, and safety controls into one inspectable system.
More than a chatbot interface
Lumen is a local-first AI desktop workspace for using models, testing them systematically, and letting agents carry out longer tasks with visible evidence and clear controls. It began at v1.0.0 as a small GPT OSS 120B chat client called Local Intelligence Console. By v4.14.2, after 587 recorded DEVLOG updates, it has grown into a broader AI Agent and research runtime.
The difficult part is not forwarding a prompt to a model API. Lumen focuses on the systems around the model: deciding which capabilities a task needs, connecting tools, preserving task state, checking evidence, exposing failures, and keeping the user in control.
Three connected workspaces
Chat is the everyday workspace. Before answering, it can determine whether a task needs web research, a document, an image, or code execution. Its evidence pipelines handle source selection and citation checks, page-aware PDF and office-document retrieval, visual evidence extraction, and sandboxed analysis that returns generated files as artifacts.
Test provides isolated batch evaluation for models and toolchains. It can compare models, thinking settings, speed, and tool permissions across text, mathematics, files, vision, safety, and practical tasks—while recording whether the expected tools were actually used and whether a failure came from the model or the surrounding infrastructure.
Persistent agents and deep research
Agent runs a plan–act–observe loop across an embedded browser, Terminal, local files, web search, memory, and subagents. Tasks are backed by Temporal rather than a single long HTTP request, so their state can survive pauses, retries, temporary failures, network interruptions, and application restarts. The Tasks Center makes planning, execution, approvals, pauses, completion, and failures visible and controllable.
Deep Research turns a question into structured research units and tracks claims against an evidence registry. It checks whether important conclusions are actually supported, identifies conflicts or coverage gaps, and performs targeted follow-up research before final writing. Its controller is also persisted through Temporal so evidence state, budgets, cancellation, and recovery remain consistent.
Safety around real tools
Lumen does not give an agent unrestricted access to the computer. High-risk file changes require previews and approval; allowed folders and protected paths are enforced against path traversal and symbolic-link bypasses. External webpages, documents, emails, and tool results are isolated as untrusted content and checked for prompt injection before they re-enter the agent context.
Memory follows the same principle: the agent proposes a candidate memory, but permanent storage requires user approval. Together with bounded code execution, structured tool results, failure reporting, and regression testing, these controls are intended to make long-running work more transparent and recoverable.
Current status
Lumen already has a substantial end-to-end system, but it remains an active project. The next work is less about adding another feature list and more about improving reliability, evaluation coverage, observability, and the boundary between unpredictable models and real local systems.
AI is used extensively as a development and review tool, while product decisions, architecture, testing standards, debugging, version direction, and final integration remain my responsibility. The platform is local-first rather than always offline: its interface, storage, local tools, and task runtime can run locally, while each model can be local or accessed through an external provider.