ProjectTelescope is a local-first observability tool for AI coding agents, providing a desktop dashboard, command-line queries, collectors, and MCP traffic capture for examining agent activity on a developer workstation.
ProjectTelescope is a local-first observability layer that helps developers inspect what AI coding agents do during their sessions. Instead of treating agent execution as an opaque process, it records structured activity that can be explored through a desktop dashboard or queried with the tele command-line interface. The software can represent sessions, turns, tool calls, file operations, shell commands, network requests, token usage, errors, sub-agents, and other agent events through a shared event model. Built-in collectors cover GitHub Copilot JSONL data, Claude Code JSONL exports, and MCP traffic, giving developers several practical ways to bring agent activity into one local inspection workflow.
ProjectTelescope keeps its observability data user-scoped on the local machine, with SQLite databases stored under the Telescope data directory and no cloud account required for its core workflow. Its MCP proxy can wrap stdio-based MCP servers and transparently forward JSON-RPC traffic while recording tool calls, resource reads, and prompts for later inspection. Developers working with desktop automation may also compare these local inspection workflows with Power Automate for desktop when evaluating related Windows automation tooling. The project additionally provides a Collector SDK for custom integrations, while its experimental status means event formats, APIs, and storage structures can change before a stable 1.0 release.
ProjectTelescope gives developers a clearer way to investigate agent behavior without requiring telemetry to leave their workstation. Centralizing sessions, turns, tool activity, file changes, shell operations, errors, token information, and other events can reduce the effort required to reconstruct what an agent actually did during a run. The dashboard supports visual exploration, while the tele CLI makes the same environment useful for terminal-oriented investigation and machine-readable queries. Built-in collectors lower the setup effort for supported agent data sources, and the MCP proxy can expose otherwise hidden interactions between agents and MCP servers. A common event schema also makes different collectors easier to reason about together. For teams building specialized integrations, the Collector SDK provides a defined route for feeding additional telemetry into the service. Because ProjectTelescope is still experimental, it is best suited to development, debugging, research, and evaluation workflows where users can accommodate possible API and data-format changes.
ProjectTelescope includes a desktop dashboard for visually exploring recorded agent activity instead of relying entirely on raw logs. Developers can use the interface to investigate agents, sessions, turns, and captured side effects, making complex execution histories easier to examine when debugging unexpected behavior or reviewing how an agent completed a task.
The tele CLI provides terminal access to the local observability service, including commands for checking status and listing agents, sessions, turns, and collectors. Commands can also produce machine-readable JSON, which makes the CLI useful for developers who prefer scripted inspection, terminal workflows, or integration with other local diagnostic utilities.
ProjectTelescope can place its MCP proxy between an AI agent and stdio-based MCP servers while continuing to forward their JSON-RPC traffic. This allows the system to capture tool calls, resource reads, and prompts that would otherwise be difficult to inspect, providing useful evidence when diagnosing agent-tool interactions and unexpected execution paths.
Built-in collectors provide ready-made ingestion paths for GitHub Copilot JSONL sessions, Claude Code JSONL exports, and live MCP proxy traffic. These collectors convert different sources into the project's shared telemetry model, reducing the amount of custom parsing required before developers can examine activity from supported AI coding-agent workflows.
A Collector SDK lets developers create out-of-process collectors that submit structured events to ProjectTelescope. The SDK supports a Rust interface and a C ABI for integration from other languages, while handling connection management, registration, collection loops, batching, backpressure, reconnection, and graceful shutdown around the custom collection logic.
Captured activity is normalized into a common event schema covering sessions, turns, tools, files, shell commands, sub-agents, planning, context, approvals, errors, network activity, token usage, Git operations, and more. This consistent representation helps developers analyze different agent sources without treating every collector's incoming data as an unrelated logging format.
Observability records are designed to remain user-scoped on the workstation, with local SQLite storage under the Telescope data directory. The core system does not require API keys, cloud accounts, or a third-party telemetry service, which is useful when developers want agent inspection data available locally rather than automatically transmitted to a hosted backend.
The CLI, service, dashboard, collectors, and local IPC design have builds for Windows, macOS, and Linux, although the project identifies Windows as its officially supported platform and notes more limited testing elsewhere. Windows uses named pipes for IPC, while macOS and Linux use Unix domain sockets for local collector communication.
ProjectTelescope is used to observe and investigate AI coding-agent activity locally. It can collect structured information about sessions, turns, tool calls, file operations, shell commands, network activity, token usage, errors, sub-agents, and other events, then expose that information through a desktop dashboard and command-line interface.
Its core design is local-first. The official project documentation states that its data is user-scoped and stored in local SQLite databases, and that the core workflow does not require API keys, cloud accounts, or third-party services.
The project ships with built-in collectors for GitHub Copilot JSONL session logs, Claude Code JSONL exports, and MCP proxy traffic. Developers can also build additional collectors through the included Collector SDK.
The MCP proxy wraps stdio-based MCP server commands and transparently forwards JSON-RPC traffic while capturing interactions such as tool calls, resource reads, and prompts. This gives developers visibility into communication occurring between an agent and its configured MCP servers.
The project is explicitly described as an experimental release under active development. Its documentation warns that event schemas, database formats, APIs, and other behavior can change, including breaking changes before version 1.0, so users should account for that instability when adopting it.
Comments