Bello Agent: The AI workspace that shows its work.

Name: Bello Agent
Author: BelloWare

Conversations, coding tools and side discussions in a native Mac app, connected to your own LiteLLM gateway, with every request, token and cost one click away.

Download Bello Agent 0.1.135

16.51 MiB · Apple Silicon · macOS 14 or later

Developer ID signed and notarized
Updates through Sparkle
Open source under the MIT License
Source on GitHub

Requires your own LiteLLM gateway.

Fix flaky retry test

Projects

Filter chats and topics

payments-api

Fix flaky retry test$0.0412 Latest 91 tok/s

Jitter strategies$0.0036 Latest 104 tok/s

Update the API docsWorking $0.0021

Webhook signatures$0.0118 Latest 77 tok/s

docs-site

The retry test in payments-api fails about one run in five. Can you find out why and fix it?

4 tool calls

Readsrc/retry.ts

SearchedsetTimeout

Editedsrc/retry.ts+12 −3

Rannpm test2.1s

The flake came from real timers: with jitter, the third attempt could land after the test’s 500 ms timeout. Two changes:

retryWithBackoff now takes an injected clock, so tests advance time instead of sleeping.
Jitter is capped at maxDelayMs, so no wait outlasts the retry budget.

All 48 tests pass.

Turn · Completedauto → coder-large$0.02316

Duration18.642s

AI 16.459s · Tools 2.183s

Input61,840

Cached 52,904 · 85.55%Uncached 8,936 · 14.45%

Output1,912

Reasoning 704 · 36.82%Other 1,208 · 63.18%

Now add a regression test for the jitter cap.

⌘Team gateway · ResponsesautoEffort · high

3 turns 9 steps · 86 tok/s108.4K tok · Cache hit 83% · $0.04127%

Illustration of a finished turn: its tool calls one line each, the answer in Markdown, and a compact report with model, duration, token split and reported cost.

Native: SwiftUI and AppKit throughout, with no web view
Your gateway: LiteLLM Responses API with tools, streaming and routed models
Transparent: Model, tokens, cost and timing on every turn
Open source: MIT License, with the full source on GitHub

Features

Calm on the surface. Detailed underneath.

Each turn reads as a few quiet lines. Open any of them to see how the answer was made and what it cost.

A native Mac workspace

Projects gather chats around trusted folders, with saved drafts and side conversations that branch off without derailing the main thread. Queue a follow-up or steer a task while it runs.

Tools with clear outcomes

Open files with ⌘P, read them beside the chat, and review Changes in a tab. Files, shell commands, skills and MCP tools each read as one line. A call stopped mid-run says “Stopped · outcome unknown”, a skipped one says “Skipped”, and nothing is replayed automatically.

Inspect every request

Open any reply to see its model, cost, timing and headers, the retained request and response bodies, and the streaming events as they arrived.

Usage you can trust

Each turn report splits input into cached and uncached and output into reasoning and the rest, next to the cost the gateway reported. The Usage Report adds it up across sessions.

Activity in the menu bar

Click the menu bar icon for running work, output speed, tokens, cost and model distribution. Speed is output tokens after the first, hidden reasoning included, over the time from the first generated token to the last; replies under 250 ms of generation are left out.

A model for every chat

Choose the model and reasoning effort per chat from the included catalog or your own. Context use is calculated locally, and long chats can be compacted.

Smooth by design. Scroll back while a reply streams and the transcript holds your reading position.

Get started

Your gateway. Your project. Four steps.

Bello Agent requires your own LiteLLM gateway. Model availability, routing, reported telemetry and usage charges depend on that gateway.

Install
Open the DMG and drag Bello Agent into Applications.
Connect your gateway
Enter your LiteLLM Responses endpoint, API key and model alias. Optionally, replace the included model catalog with your own catalog URL.
Choose and trust your project
Test & Start checks the selected model with a small request before your first chat opens.
Inspect turns and usage
Click Info on a turn report to search its captured requests, responses and headers, or open the Usage Report for more.

Privacy and local storage

What Bello Agent keeps, and how.

Captured traffic is what makes every request inspectable. Adjust capture and retention in Settings.

Kept locally

Credentials and configuration are kept in the native Keychain vault.
HTTP request and response capture is on by default, with 30-day body retention, subject to your storage quota and settings.
New bodies are stored locally without encryption.

Masked in captures

Authentication headers are masked; longer request tokens keep only their final four characters.
Known credentials found in request bodies are replaced with SHA-256 fingerprints.
Bodies can still contain your prompts, files and tool data.

Figures as reported

Cost totals include only values the gateway reports, with coverage counts.
Gateway response-cache hits and provider prompt-cache tokens are shown separately.

Open source

Open source under the MIT License.

The app, its supervised Swift helper, the native transcript, the build and release scripts and every release’s validation record are at github.com/BelloWare/BelloAgent. Building, running and the test suite need no credentials; signing and notarization use ones the repository does not carry.

View the source on GitHub

BelloWare/BelloAgentMIT License

apps/macosThe SwiftUI and AppKit app, native transcript and tests
packages/swift-hostThe supervised Swift helper
scriptsBuild and release scripts
docs/validationEvery release’s validation record