🎉OpenBlock reached #1 on Terminal Bench on 09/10/25View announcement

The Agent ThatWorks Like You

OB-1 brings frontier AI directly to your terminal. Context-aware, model-agnostic, and built for developers who ship.

Changelog

We ship every day

The frontier of agentic coding, right in your terminal.

December 2025
Feature

Value Delivery Tracking

Tracks concrete outcomes: files staged, commits made, PRs submitted via Graphite/GitHub CLI. Understands the difference between running git add vs git status for measuring real impact.

Improvement

Silent Background Auto-Updates

OB1 updates itself automatically on startup without interrupting your workflow. Shows a subtle 'updated' notification only after successful update.

Feature

Specialized Sub-Agents (Explore, Plan, General)

Delegate complex tasks to purpose-built agents: Explore for fast read-only codebase searches, Plan for architectural design, General for multi-step modifications.

Improvement

Intelligent Error Messages

User-friendly error handling for common issues like 413 (payload too large) - explains what happened and suggests solutions instead of raw API errors.

Improvement

File Extension Verification Guidelines

Prevents agent from calling file tools with incorrect extensions. Automatic verification before reading or writing files.

Feature

Session Memory with Cross-Session Learning

LLM-powered session extraction that creates persistent memory across conversations. Summaries from past sessions are injected into new prompts.

Feature

Full Expand Mode (Ctrl+E)

Completely bypass all output height constraints for the current view. Useful for reviewing massive diffs or lengthy error stacks.

Feature

Interactive Shell Focus (Ctrl+F)

Focus into embedded shell sessions for direct interaction. The agent pauses while you type commands directly, then resumes when you're done.

Improvement

Microcompact Context Management

Context window optimization that automatically compacts old tool results with "[Tool result compacted]" markers when approaching context limits.

Improvement

Real-Time Cost & Token Tracking

Async cost tracking with OpenRouter generation stats, token usage, and context utilization. Warns when approaching model limits.

Feature

Image-in-Prompt Support

Paste images from clipboard or drag screenshots into the prompt. Handles volatile macOS temp screenshots that disappear after drag-and-drop.

Improvement

Tool Response Truncation with Token Budgets

Intelligent truncation of large tool outputs based on token budgets, keeping the most relevant content while respecting context limits.

Feature

Todo List with Ctrl+T Toggle

Built-in task management with write_todos tool. Agent breaks down complex tasks, tracks progress, marks items complete. Toggle visibility with Ctrl+T.

Feature

Auto Mode with LLM-Based Task Classification

Intelligent model routing that analyzes your prompt complexity in real-time. Routes simple tasks to fast models (Haiku) and complex tasks to premium models (Opus/Sonnet).

Feature

Session Resume with Named Sessions & Thread Forking

Name sessions with /session name <tag> for easy recall. Fork at any point with the thread tool to branch conversations. Resume with /resume.

Feature

Thread Tool for History Navigation

Access and fork past conversation threads. Full thread history with search, continuation, and branching capabilities.

Feature

Deep Research Tool (Parallel API)

In-depth web research that synthesizes information across multiple sources. Handles complex questions requiring comparison and nuanced understanding.

November 2025
Improvement

Audio Notifications

Configurable sound alerts for task completion, errors, and attention-needed events. Toggle with /notifications command or OB1_NOTIFICATIONS env var.

Architecture

Context Window-Aware Model Selection

When using weighted presets or Auto Mode, OB1 filters out models whose context windows are too small for your current conversation.

Architecture

Extended Thinking Support

Configurable thinking budgets per model/agent. Sub-agents like Plan use thinkingBudget: -1 (unlimited) for deep reasoning on architectural decisions.

Integration

Web Fetch with Tavily API

Fetch content from any URL including localhost and private networks. Supports up to 20 URLs per request with custom processing instructions.

Integration

Multi-Provider Model Access via OpenRouter

Access 200+ models from OpenAI, Anthropic, Google, Meta, Mistral and more through a single unified API with automatic prompt caching detection.

Improvement

Binary File Processing Limits

Safe processing of images, PDFs, and other binary files with automatic size checking and limits. Groups files under 3MB per turn.

Improvement

Shell Output Token Efficiency

Intelligent output management for shell commands. Automatic use of quiet flags, temp file redirection for large outputs, and smart truncation.

Integration

Web Fetch & Research Tools

Fetch content from any URL including localhost and private networks. Supports up to 20 URLs per request with custom processing instructions.

Improvement

Ripgrep-Powered Search

Fast, optimized code search with automatic output limiting (max 20k matches). Supports pagination, regex patterns, and file type filtering.

Feature

Smart Replace with Context Matching

Precise text replacement requiring 3+ lines of context. Supports multiple replacements with exact match verification to prevent accidental changes.

Architecture

Project Convention Detection

Automatic analysis of existing code patterns, test frameworks, and styling conventions. Ensures all generated code matches project standards.

Feature

Memory System for User Preferences

Persistent storage for user-related facts and preferences. Remembers coding styles, common paths, and personal tool aliases across sessions.

Improvement

Glob Pattern File Discovery

Efficient file matching with modification time sorting. Supports case-sensitive search and respects .gitignore patterns.