ProjectTelescope is a local-first observability tool for AI coding agents. It records tool calls, file changes, token usage and decisions in one dashboard and CLI, so developers can see what an agent actually did.
ProjectTelescope is a local-first observability layer that makes the work of AI coding agents visible instead of opaque. Agents spawn sub-agents, call tools, read and write files, and consume tokens behind the scenes; this tool captures those actions as structured events and stores them locally. The workflow has three parts: collectors watch an agent, either by scanning its session logs or by intercepting MCP traffic, a background service ingests the events into a local SQLite store, and a desktop dashboard plus the tele command-line tool present the results. Everything runs on your own machine, so no session data, prompts or responses are sent to a third party.
Coverage is not limited to a single assistant. Built-in collectors read GitHub Copilot and Claude Code session logs, and an MCP proxy sits between an agent and its MCP servers to capture JSON-RPC traffic in real time, which means tool calls, resource reads and prompts are recorded even when the agent writes no log file. Setup is deliberately quiet: one command rewrites existing MCP configuration files to route through the proxy, and the same command can undo the change. Teams that prefer a small, self-contained Windows utility will find the install pattern familiar; Project My Screen App for Windows Phone follows a similar single-purpose desktop model, while ProjectTelescope focuses entirely on agent activity that stays on the local machine.
The main benefit of ProjectTelescope is turning guesswork into evidence. When an agent reports that a task is finished, the dashboard shows the turns that led there, the tools it invoked, the files it touched and the approvals it requested, so a surprising edit or a failed command can be traced to the exact step that produced it. Token and model reporting make cost visible as well, which helps when comparing agents, prompts or workflows over time. Because the data lives in local SQLite databases and requires no account, there is little friction between wanting answers and getting them — nothing has to be uploaded, and sensitive prompts never leave the workstation. The open Collector SDK extends the same visibility to agents that ship no logs, using one consistent interface rather than a privileged internal path, so a custom collector sees exactly what a built-in one does. Finally, the CLI makes the same records scriptable, which suits developers who want agent behavior checked in automation rather than only in a window.
Key capabilities and highlights.
Every session opens as a timeline with an execution graph that links related tool calls to one another. You can follow a long run from the first user message through each tool invocation to the final response, instead of piecing the sequence together from a chat transcript that hides everything in between.
A bundled proxy wraps stdio-based MCP servers and records the JSON-RPC traffic that passes between them and the agent, capturing tool calls, resource reads and prompts in real time. That matters for agents that keep no useful local log, because the proxy becomes the record of what was actually requested.
Log-based collectors ingest GitHub Copilot and Claude Code session exports automatically, so both assistants appear in one place without manual file wrangling. You get a combined view of sessions, turns and tool calls across the two, which is far easier to compare than reading separate log files by hand.
ProjectTelescope ships a CLI that lists agents, sessions, turns and loaded collectors, and every command accepts a flag for machine-readable output. That makes agent activity queryable from a terminal and easy to wire into scripts or a CI job, where a result can be checked automatically rather than inspected by eye.
A graphical dashboard presents the same records as execution graphs, side-effect details and token usage charts. It filters duplicate noise, merges tool start and completion events into single readable entries, and keeps the full event payload in a detail pane, which suits readers who would rather browse agent activity than type commands.
Tool calls, file reads and writes, shell commands, sub-agent spawns, approvals, errors, rate limits and token usage are normalised into a shared set of event types. Because every collector emits that same schema, activity from different agents can be filtered, compared and reported on the same terms.
A documented SDK with a Rust API and a C-ABI lets teams write collectors for agents that ship no support. Custom collectors use the identical interface and capabilities as the shipped ones, with no privileged internal path, so extending ProjectTelescope means implementing four methods and letting the SDK handle transport.
ProjectTelescope keeps sessions in user-scoped SQLite databases and requires no cloud account or API key. Prompts, results, file side effects and token counts therefore stay on the workstation, which keeps sensitive code out of third-party services and means there is nothing extra to configure before you start collecting.
Browse historical releases and previous builds.
Common questions about ProjectTelescope.
It is a local-first observability tool for AI coding agents. It captures what an agent did during a session — prompts, tool calls, file changes, shell commands, sub-agent spawns, errors and token usage — and presents that record through a desktop dashboard and the tele CLI, with the underlying data stored on your own machine.
No. The design is local-first: events are written to user-scoped SQLite databases, and the tool requires no API keys, cloud accounts or third-party services. Telescope itself sends nothing off the workstation, although the agents you monitor may still make their own network requests.
Built-in collectors cover GitHub Copilot session logs, Claude Code session exports, and any MCP-compatible agent routed through the proxy. Other agents can be added by writing a collector against the open SDK.
It sits between an agent and an MCP server so JSON-RPC traffic passes through it, letting Telescope record every tool call, resource read and prompt in real time. The setup command rewrites known configuration files for Copilot, Claude Desktop, Cursor and VS Code to use the proxy, and the same command with the undo flag restores the originals.
Yes. You can build a custom collector for it. Collectors are standalone binaries that connect to the local ProjectTelescope service over a named pipe or Unix socket, and the SDK handles registration, batching, backpressure and reconnection, so you only define what data to return and how often to collect it.
No. The Collector SDK exposes both a Rust API and a C-ABI, so collectors can be written in C, Python, Go or any language able to link the shared library and call its init, submit and shutdown functions.
It is published as an experimental release in active development. Core functionality works, but event schemas, database formats and CLI commands may change between releases, and no semantic versioning guarantees are offered while the project remains experimental. It suits developers who are comfortable with occasional breaking changes.
Yes. Cost and model events record token usage and rate-limit hits, and the dashboard charts consumption over time with breakdowns by model and agent, which makes it easier to compare workflows and spot sessions that burn through budget.
ProjectTelescope is a local-first observability tool for AI coding agents. It records tool calls, file changes, token usage and decisions in one dashboard and CLI, so developers can see what an agent actually did.
KDE Connect links a phone and a computer into one working environment: files, links, notifications, and the clipboard move between them, and the phone doubles as a trackpad, keyboard, or presentation remote. It is free, open source, and pairs over the local network.
Microsoft Edge Dev is a preview web browser for testing upcoming Edge capabilities, inspecting websites with built-in developer tools, and evaluating browser changes before they reach stable releases.
Microsoft Edge Beta is a preview browser for evaluating upcoming Edge features, testing websites and extensions, and identifying compatibility problems before browser changes reach the stable channel.
Ekiga is a free, open-source softphone and video conferencing client for Linux and Windows that places audio and video calls over SIP or H.323, with contacts, presence, chat and call forwarding built into one desktop app.
Comments