OpenHands v1.13 Adds a Context Meter That Finally Stops You From Blowing Your Token Budget

OpenHands v1.12.0 landed on August 7, 2026, and it didn’t add a flashy new agent mode or a redesigned UI. It did something more useful for beginners: it stopped making you guess which models cost money.

Three models — openhands/glm-5.2, openhands/deepseek-v4-flash, and openhands/minimax-m2.7 — are now free forever on OpenHands Cloud. The model picker shows orange Free badges next to them. Select one, and the badge sticks beside the model name in your settings. No API keys. No billing surprises. Just code.

Then v1.13.0 dropped on August 13 with the feature beginners have been begging for: a live context window usage meter with manual compaction. No more surprise “context full” errors halfway through a refactor. You see the ring fill up in real time, and when it gets tight, you hit Compact and keep going.

If you’ve been holding off on AI coding agents because every tutorial starts with “first, add your credit card,” this is your entry point. And if you’ve been burned by context limits before, v1.13 fixes the worst offender.

What OpenHands Actually Is (Two-Minute Primer)

OpenHands is an open-source AI coding agent (84,100+ GitHub stars, MIT license) that runs in your browser, as a desktop app (Agent Canvas), in your terminal, or self-hosted on your own infrastructure. You give it a task — “add authentication to this React app” or “fix the failing tests in api/users” — and it plans, writes code, runs commands, and iterates until it works.

Unlike Cursor or Copilot, OpenHands is provider-agnostic. Bring your own Anthropic, OpenAI, Google, or local models. Or use OpenHands’ own routed endpoints — including the three free ones.

Key numbers:

• 84.1k stars · 10.9k forks · 550+ contributors
• Web + Desktop (Electron) + CLI + Docker + Self-hosted
• MIT licensed — you own your data, your agents, your workflow

The v1.12.0 Change: Free Models You Can Actually See

Before this release, OpenHands had free tiers but the UI didn’t make them obvious. You’d pick a model, start coding, and maybe hit a quota wall later. v1.12.0 fixes that with display-only UI changes (no routing logic touched):

What you see in the model picker:

1. Orange Free badge next to each model in the dropdown
2. Orange info note at the top listing all free OpenHands model IDs
3. After selection, the badge stays pinned next to the model name in settings — so you never forget which one you’re on

This is display-only. The raw model IDs (openhands/glm-5.2, etc.) stay unchanged for routing. Adding or removing a free model is now a one-line config change.

The v1.13.0 Change: Context Meter, Artifact Previews, and Conversation Archive

v1.13.0 landed August 13, 2026 — a focused release that solves the biggest daily pain points for agent users.

Live context window usage meter with manual compaction

The headline feature: a context meter (ring visualization) in the chat UI that shows real-time context window consumption. When you approach the limit, a Compact button appears — one click summarizes the conversation, preserves file paths and decisions, and frees up space. No more “context full” errors mid-refactor. The meter draws from the foreground color (fixed in v1.13.0 after a brief regression), so it’s always visible.

Inline markdown artifact previews

Agent outputs now render markdown artifacts inline — code blocks, mermaid diagrams, tables, and formatted text appear as live previews in the chat stream. You see the architecture diagram or the generated spec while the agent is writing it, not after you scroll back.

Client-side conversation archive

Conversations can now be archived locally (client-side) instead of deleted. Archived sessions stay searchable, restorable, and don’t clutter the active list. This is a pure UI feature — no server changes — so it works instantly in self-hosted and cloud deployments alike.

Ready-for-dev issue readiness gate

A new issue readiness gate evaluates GitHub issues against type-specific criteria before an agent picks them up. For “bug” issues: reproduction steps, expected vs actual behavior, environment info. For “feature” issues: acceptance criteria, design links, scope boundaries. Agents only start work on issues that pass the gate — reducing wasted cycles on underspecified tasks.

Other fixes that matter

• Provider list untruncated: The Basic LLM provider dropdown no longer caps at 100 entries — all configured providers show.
• MCP automation surfacing: Non-MCP-installable integrations now appear in the UI instead of silently dropping.
• Touch-friendly overflow menus: Mobile and tablet interaction fixed.
• Launcher services spawn without implicit shell: Cleaner process tree, fewer zombie processes.
• Post-install message: New users see exactly how to start Agent Canvas after npm install.

Why This Matters for Beginners

1. Zero-Friction First Session

Install the desktop app (or run docker run -p 8000:8000 ghcr.io/openhands/agent-canvas:latest), open the model picker, pick a Free model, start coding. No account linking, no API keys, no “trial expired” emails.

2. Learn the Agent, Not the Billing

Most beginners quit when the first invoice arrives. With free models locked in, you spend your first weeks learning how to prompt, how to steer, how to review diffs — not debugging rate limits.

3. Real Models, Not Toy Demos

glm-5.2 and deepseek-v4-flash are production-grade. They handle multi-file features, write tests, debug failing CI, and refactor legacy code. You’re not sandboxed to “hello world” tasks.

4. Graduation Path Without Migration

When you outgrow the free tier (or want Claude Opus 4.1 / GPT-5), you add your own keys in the same UI. Same agent, same history, same workflow. No tool switch.

Quick Start: Your First Free OpenHands Session

Option A: Desktop App (Easiest)

1. Download OpenHands Agent Canvas from GitHub Releases (macOS .dmg, Windows .exe, Linux .AppImage/.deb)
2. Install and launch
3. Open Settings → LLM → Select OpenHands provider
4. Pick openhands/glm-5.2 (orange Free badge) → Save
5. Click New Conversation → Type: “Create a Python CLI tool that fetches GitHub trending repos and exports to CSV”
6. Watch it plan, write, test, and deliver — and watch the context meter in the top right as the conversation grows

Option B: Docker (Zero Install)

docker run -it --rm -p 8000:8000 ghcr.io/openhands/agent-canvas:latest
# Open http://localhost:8000 in your browser
# Same steps 3-6 above

Option C: Self-Hosted (Full Control)

# Clone and run locally — your data never leaves your machine
git clone https://github.com/OpenHands/OpenHands.git
cd OpenHands
docker compose up -d
# Access at http://localhost:3000

What You Can Build with the Free Models (Real Examples)

The Catch (Honest Assessment)

• Rate limits exist — Free endpoints share capacity. At peak hours, you may queue.
• No SLA — This is community infrastructure, not enterprise cloud.
• Context window varies — minimax-m2.7 leads; others are standard 128K-200K. v1.13’s context meter + compaction means you now see it coming and can compact before you hit the wall.
• Desktop app is Electron — ~300MB download, not a tiny CLI.
• Self-hosting needs GPU for local models — Free cloud models don’t, but BYO local models do.

Workaround: If you hit limits, switch between the three free models — they’re independent pools. Or add a $5 OpenRouter credit for burst capacity.

How OpenHands Compares (Beginner Lens)

Bottom line: OpenHands is the only major agent where a beginner can start today, free, with production models, no credit card, and graduate to BYOK/self-hosted without switching tools.

What’s Next for OpenHands (Roadmap Signals)

The v1.13.0 release shipped two major roadmap items: the context window usage meter with manual compaction and the ready-for-dev issue readiness gate — both called out in recent commits as active work. The commit velocity remains intense:

• 1,458 branches, 177 tags, 8,014+ commits — this moves fast
• Recently shipped: Agent Canvas rename, multi-arch Docker, context meter + compaction, ready-for-dev issue gates, artifact previews, conversation archive
• Active: MCP server support, skill system, agent-to-agent messaging, Slack/GitHub integrations

The free model clarification (v1.12) and context meter (v1.13) aren’t one-offs. They’re infrastructure for onboarding — making the first “wow” moment happen before the first paywall, and keeping you productive after the honeymoon.

Should You Switch? (Decision Checklist)

✅ Start with OpenHands if:

• You want zero-cost access to strong models
• You value open source and data ownership
• You want web + desktop + CLI from one codebase
• You might self-host later (team, compliance, air-gap)

❌ Stick with Cursor/Copilot if:

• You live in VS Code and never want to leave
• You need instant IDE integration (inline diffs, tab completion)
• Your team already pays for Cursor/Copilot seats

🤔 Run both — OpenHands for background/automation tasks (CI fixing, refactoring, test generation), Cursor for interactive editing. They’re complementary, not competitive.

Your First Hour Checklist

• Install Agent Canvas or run Docker
• Select openhands/glm-5.2 (orange Free badge)
• Give it a real task from your backlog — not a tutorial
• Watch the plan → act → verify loop in the transcript
• Review the diff before accepting
• Commit and push — you just shipped with a free agent

Related Reading

• Best Coding Agents 2026 — Decision Guide (Not Another Hype List)
• Claude Code vs Kilo Code: Which Terminal Agent Wins?
• OpenCode Deep Dive: The Skill-Driven Open-Source Coding Agent
• Coding Agent Pricing Guide 2026 — What You Actually Pay at Scale

Related articles

• Claude Code Alternatives in 2026: 12 Options Compared
• Oh My Pi v17.3.3 Fixes Gemini Reasoning Loops, Hashline Edge Cases, and TUI Rendering — What Beginners Need to Know
• Claude Code: Skills vs Subagents vs MCP — The 2026 Decision Guide

Want to compare the free OpenHands models against Claude, GPT, Gemini, and Grok side-by-side? aiFiesta gives you all major models in one place for $12/mo — including the ones OpenHands routes to.