Project Telescope: Local Observability to See What Your AI Agents Are Actually Doing

Telescope is a local-first observability tool for AI Agents and MCPs, giving teams clear visibility into what agents did, why they did it, and where they got stuck.

AI agents have become a part of our daily workflows. They refactor code, orchestrate sub-agents, call tools, read and write files, and burn through tokens — all while you're sipping your coffee. But here's the problem: its hard to see what they're actually doing.

Until now.

Project Telescope is a local-first observability tool for AI agents. Think of it as DevTools for your AI pair programmer — a live, structured view of everything happening beneath the surface of tools like GitHub Copilot, Claude, and any MCP-compatible agent.

😰 The Problem

	Challenge	What's Happening
🔮	Black Box Agents	AI agents work in isolation. You can't see what they're doing, what tools they call, or why they make decisions.
🤷	No Unified View	Multiple agents, multiple logs, multiple formats. No single place to understand the full picture.
🐛	Hard to Debug	When things go wrong, tracing the issue across agent sessions, MCP calls, and tool responses is painful.

When you run a traditional CI pipeline, you get logs. When your web app throws an error, you get a stack trace. But when your AI agent decides to spawn three sub-agents, rewrite a config file, and make a network request — what do you get? A chat message that says "Done."

That gap between what agents do and what developers see is growing fast. As agents become more autonomous the need for transparency grows with them.

---

✨ The Solution

Telescope captures, structures, and surfaces what your AI agents do — tool calls, decisions, conversations, and reasoning — in one beautiful dashboard and CLI.

Copilot CLI    Claude CLI    OpenClaw    Any Agent
      |              |            |           |
      +--------------+------------+-----------+
                     |
                     v
                Collect --> Local Store --> Visualize

No cloud accounts. No API keys. No config files. Everything runs on your machine, and everything stays on your machine.

---

🔧 How It Works

Telescope has three layers:

1. Collectors (open-source) — shared libraries that watch AI agents and emit structured events. They hook into agent activity via stdin/stdout interception, JSONL log scanning, SDK hooks, and MCP proxy interception.

2. The Service (installed locally) — ingests collector events, stores them in local SQLite databases, and surfaces insights through a CLI and desktop dashboard.

3. CLI & Dashboard — a powerful command-line interface (tele) and a desktop dashboard for exploring sessions, watching live events, and drilling into agent behavior.

The architecture is deliberately flat: every collector, whether built-in or community-authored, uses the exact same C-ABI interface. A collector you write has identical capabilities to one that ships in the box.

Agent (Copilot, Claude, your agent, etc.)
                |
                v
    Collector                  <-- open-source, extensible
                |
                v
    Telescope Service          <-- local, no network egress
                |
          +-----+-----+
          v           v
      Dashboard      CLI

---

🚀 Key Features

• 📊 Timeline view of all agent activity
• 🔗 Execution graph showing tool call relationships
• 🔄 Chat turn grouping (user → agent → response)
• 🧹 Automatic noise and duplicate filtering
• 🔀 Tool start/complete event merging
• 🏷️ Agent-specific parsing (Copilot, Claude, etc.)
• 📋 Full event data in detail pane

---

👀 What You See

	Category	Details
💬	Conversations	Every user message and agent response, with full context and reasoning.
🛠️	Tool Calls	Which tools were called, with what arguments, and what they returned.
🎯	Decisions	Agent intents, reasoning, and choices.
⚡	Performance	Tool durations, session lengths, and execution patterns.

This isn't just logging. It's a structured, queryable record of agent behavior that lets you debug failures, audit decisions, spot patterns, and understand where agents get stuck.

---

⚡ Quick Start

1️⃣ Install

Windows

Download the MSI from the releases page

msiexec /i telescope-windows-x64.msi /quiet

This downloads and runs the MSI installer, which sets up the background service, dashboard, tele CLI, and all built-in collectors.

macOS & Linux — coming next week! Stay tuned for installation instructions.

2️⃣ Verify it's running

tele doctor

3️⃣ See what your agents are up to

tele watch        # Live stream of agent activity
tele sessions     # Browse recent sessions
tele insights     # Surface patterns and anomalies
tele dashboard    # Launches the dashboard

4️⃣ (Optional) Build a custom collector

git clone https://github.com/microsoft/project-telescope.git
cd project-telescope/examples
cargo build --release
tele collector install ./target/release/
tele collector enable my-custom-collector

That's it. Your collector is now feeding events into the same pipeline as the built-in ones.

---

What's Next

Project Telescope is available now on Windows, with macOS and Linux support coming next week. The collector plugin system is open-source under the MIT license.

If you've ever stared at an AI agent's output and wondered how it got there, give Telescope a look. Your agents are doing more than you think. Now you can finally see it.

---

🔭 See Everything. Debug Agents. Trust Your Agents.

---

Project Telescope is open-source preview project at github.com/microsoft/project-telescope. The collector SDK is MIT-licensed. The service, dashboard, and CLI are distributed as closed-source binaries under MIT license terms.

Built by Cassie Breviu and Daniel Pagan