Goose Just Became the Only AI Agent That Show

You run an AI coding agent for an hour. The session ends. You check your API dashboard — $47.32. For what? Which messages? Which model calls? Which tools?

Most agents shrug. Goose v1.46.0 doesn’t.

Released 11 hours ago, this is the most transparency-focused release in the open-source agent space. Per-message cost breakdown. Cache token accounting. Streaming shell output so you see work happening in real-time. Plus 20+ new providers including Azure AI Foundry, OllamaCloud, and Fireworks AI.

The Feature You’ve Been Begging For: Per-Message Receipts

Per-message usage stats UI — tokens, cost, TTFT (time to first token), tok/s — displayed inline for every single message.

Per-message usage/cost tracking with derived session totals — you see the running tab accumulate in real-time.

Track cache tokens for accurate cost reporting — cached prompt tokens (which cost ~10% of fresh tokens) are now counted separately. No more overestimating your bill.

This isn’t a dashboard you open in a separate tab. It’s in the chat. Every assistant response shows:

✓ 1,234 tokens • $0.0034 • TTFT 1.2s • 45 tok/s • 💾 892 cached

You know exactly what each interaction cost. Before the session ends. Before the credit card bill arrives.

Streaming Shell Output: Watch the Work Happen

Streaming shell output while commands run — no more staring at a spinner while npm install or cargo build runs for 3 minutes.

You see stdout/stderr stream in real-time, interleaved with the agent’s reasoning. If a test fails, you see the red output the instant it appears. If a build hangs, you know immediately.

This changes the mental model from “delegate and pray” to “pair program with visibility.”

20+ New Providers: Run Anything, Anywhere

Provider	What It Unlocks
Azure AI Foundry	Enterprise multi-LLM with compliance
OllamaCloud	Dynamic model discovery for local-first
Fireworks AI	Fast inference for Llama, Mixtral, Qwen
Friendli	Optimized serving for open models
Celeris	High-throughput inference
Sakana AI	Fugu API for specialized models
iFlytek Spark / Astron MaaS	Chinese-market models
Together AI	Cheap open-model hosting
Perplexity	Search-augmented generation
Alibaba/Qwen via DashScope	Strong coding models
Databricks AI Gateway	Lakehouse-integrated LLMs
Scaleway	European sovereign cloud
NEAR AI	Web3-native agents
xAI SuperGrok	Grok models via OAuth
Vercel AI Gateway	Edge-deployed model routing
AWS Bedrock (OpenAI GPT via Mantle)	Enterprise GPT on Bedrock
Azure AD / Entra ID	Corporate SSO for OpenAI
Meta Muse Spark 1.1	Meta’s latest via Models API

Total provider count: 30+. If it runs an LLM, Goose probably connects to it.

The Architecture Upgrades You’ll Feel

Feature	Why It Matters
Unrolled agent loop	Cleaner reasoning traces, easier debugging
Cache-safe request assembly	Append-only turn context = no context corruption
Structured summary output with templates	Compaction you can actually read
TUI diff viewer	See changes side-by-side in terminal
Resizable sidebar with persisted width	Your layout survives restarts
Worktree-aware directory switcher	Git worktrees feel native
Chat history search in nav panel	Find that session from three weeks ago
Group chat sessions by project	Automatic project clustering
Collapsible project group headers	Clean sidebar at scale

Skills & Extensions: The Platform Deepens

Hooks with PreToolUse denial — intercept and block tool calls before they run (security, policy, guardrails)
/goal command — agent self-evaluates before finishing (did I actually solve the task?)
/status slash command — session health at a glance
/model slash command — switch models mid-session
--edit session flag — edit conversation before forking
goose review — local code review command (runs in your repo, no cloud)
Open plugins generalization with skills — plugin = skill = portable capability
Summon extension: peek mode — watch async background tasks without blocking
Load global hints from ~/.agents/AGENTS.md — portable instructions across agents
Dynamically refresh skill instructions each turn — skills stay current
CLI to list skills with token counts — audit your context budget
Image read tool — agent sees screenshots, diagrams, UI mockups

Enterprise & Observability Ready

GenAI semantic convention attributes for OTEL — standard telemetry
Enrich root span with gen_ai attributes — traces show model, tokens, cost
ACP session/new supports Hidden session via _meta.hidden — background agents
Stable agent event message identity — correlate logs across systems
TLS support for ACP serve — secure agent-to-agent comms
Bind MCP apps to trusted ownership metadata — supply chain security
Proactive OAuth token refresh — no re-auth mid-session

Desktop & Mobile Polish

Reconnect desktop ACP sessions after sleep/connection loss — laptop closes, opens, session lives
Termux detection in installer — musl portable build for Android
Option to disable automatic update downloads — air-gapped environments
Encrypted Nostr session sharing — portable, private session transfer
Import sessions — migrate from other tools
Hugging Face OAuth support — one-click auth
goose://new-session and goose://resume deep links — launch from anywhere
Initial prompt with goose://new-session — parameterized session starts

SDK & Extensibility: Build On Top

UniFFI SDK for cross-language bindings — Python, Kotlin, Swift, Go
Provider bindings MVP to goose-sdk with Python wheel — pip install goose-sdk
Kotlin FFI with Maven publish — Android/JVM first-class
Create source archive as release artifact — reproducible builds

How to Update

# Linux/macOS (official installer)
curl -fsSL https://github.com/block/goose/releases/download/v1.46.0/download_cli.sh | bash

# Or download binary directly
# https://github.com/block/goose/releases/tag/v1.46.0

# Verify
goose --version
# Should show 1.46.0

Why This Matters

Goose v1.46.0 isn’t just “more features.” It’s a philosophical shift.

Every other major agent (Claude Code, Cursor, Codex, Copilot) treats cost as an afterthought — a monthly surprise on your credit card. Goose makes cost visible, per-message, in real-time.

Combined with 30+ providers, local-first architecture, and open extensibility, this is the first agent that feels like infrastructure you own rather than a service you rent.

Full changelog: github.com/block/goose/releases/tag/v1.46.0

Updated August 13, 2026 — covers v1.46.0 released August 12, 2026

Goose Just Became the Only AI Agent That Shows You Every Penny It Spends