I’m building Glance, a platform for creating coding agents, document Q&A agents, and custom agent workflows.

I have seen many teams build their own agent harnesses to turn AI capabilities into practical value. Glance brings that foundation into one platform, helping individuals and enterprises build, evaluate, and monitor agents faster. It runs within the customer’s own environment, giving them control over their data and infrastructure.

At its core is a persistent agent harness with asynchronous tool calls, subagent management, and agent-to-agent communication. Agents can use Code Mode, terminals, and browsers.

Glance brings four parts of the agent lifecycle together:

Build: Configure models, instructions, tools, and data sources.

Connect: Integrate your agents with third-party platforms such as Slack, so teams can use them within their existing workflows.

Evaluate: Manage eval datasets, define scoring rubrics, and run online and offline evaluations.

Monitor: Inspect agent runs and runtime behavior.

Data connections include live local directories, native connectors, and MCP servers. Native connectors cover document uploads, Confluence, Jira, OneDrive, SharePoint, Google Drive, SQLite, and DuckDB, with source permission enforcement.

Other capabilities include guest mode, encryption with keychain integration, multi-tenancy, and quotas. Model providers include OpenAI, Anthropic, Gemini and OpenAI Compatible Endpoint (Baseten, Fireworks etc). GPU-enabled Linux servers support document processing with OCR and vision-language models.

We’re starting with a Mac app and Linux server deployments. A self-hosted Kubernetes service is planned.

[Mac App] https://github.com/vpanghal/glance-releases/releases/downloa... [Interactive Demo] https://app.supademo.com/demo/cmuntgnla0136n820l9tqly2c

I’d appreciate feedback on the agent harness, evaluation workflow, and what you need to see when inspecting a run.