OpenCode Arrives on 19 June 2026
On 19 June 2026, the team at SST — the open-source infrastructure company that had rebranded as Anomaly — released OpenCode, a terminal-based AI coding agent under an MIT licence. Within weeks of launch, OpenCode crossed 160,000 GitHub stars, making it the most starred open-source AI coding agent in the category's history by a wide margin. More than 7.5 million developers use it monthly as of early July 2026. The founding team — Jay (CEO), Frank (co-founder), Dax Raad (who joined SST in 2021 and became a co-founder), and Adam Elmore — built OpenCode to put the AI agent at the terminal level rather than inside an IDE, a design decision that defines almost everything distinctive about the tool.
Terminal-Native, Not IDE-Native
IDE extensions — GitHub Copilot, the Cursor plugin, Windsurf — operate within the host editor's permission model. They can suggest completions, modify open files, and run configured tasks, but they cannot arbitrarily execute shell commands, manage background processes, or interact with the full terminal environment. This is a structural constraint: the sandboxed execution model that keeps IDE extensions predictable also prevents them from functioning as true system-level agents.
OpenCode runs in the terminal as a first-class process with access to everything the developer's shell session can do — running scripts, calling CLIs, executing test suites, committing to git, and piping outputs between tools. An agent running at this level can observe build failures, parse test output, make corrections, and re-run validation without leaving a single interface. Conversation history is stored locally in SQLite on the developer's machine, not transmitted to any vendor cloud. The data stays on the device because the tool was architected that way from the start, not as an optional privacy feature.
Model-Agnostic: BYO Keys and Local Inference
OpenCode connects to more than 75 AI providers through a single unified interface. Developers bring their own API keys for Anthropic, OpenAI, Google, Mistral, and other providers, and pay only for the tokens they consume — there are no per-seat subscription fees to Anomaly. This means the cost of using OpenCode scales directly with usage rather than a flat monthly rate that accrues regardless of actual agent activity.
For teams with strict data governance requirements, OpenCode supports fully local inference through Ollama, llama.cpp, and LM Studio. A developer can run a capable open-weight coding model entirely on local hardware with zero data leaving the machine. The MIT licence means organisations can inspect, modify, and self-host every line of OpenCode's logic without needing permission from Anomaly or accepting an external privacy policy — a meaningful distinction for regulated industries and government software teams where vendor terms of service are a procurement obstacle.
Plan Mode, Build Mode, and Parallel Sessions
OpenCode implements two operating modes to separate the planning and execution phases of agentic coding work. In Plan mode, the agent reads the codebase and task description, then produces a structured plan listing which files it will edit, which commands it will run, and which dependencies it needs to resolve — before taking any action. The developer reviews and approves the plan. In Build mode, the agent executes the approved plan step by step, with high-privilege actions such as git commits, writes to sensitive directories, and file deletions requiring explicit human confirmation before proceeding.
This two-stage design addresses a failure mode common in coding agents that collapse planning and execution into a single undifferentiated loop: errors compound because the agent cannot surface its assumptions for review at the point where correction is cheapest. Separating Plan and Build gives the developer meaningful oversight before irreversible actions occur, not just a log of what happened afterward. OpenCode also supports running multiple agent sessions in parallel within the same terminal environment, allowing long-running coding tasks across different parts of a codebase to proceed concurrently without blocking each other.
What OpenCode Means for Indian Engineering Teams
The most significant attributes of OpenCode for Indian software teams are its MIT licence, local inference support, and BYO-key pricing model. Enterprise teams at Indian IT services companies and product startups frequently face data governance requirements that restrict sending client code to external cloud APIs. OpenCode's local-first architecture removes that constraint without sacrificing agent capability — a developer can run a larger open-weight coding model on a standard developer workstation and keep the entire session on-premises.
The BYO-key model changes the cost equation for teams with uneven agent usage patterns, where a small number of developers are heavy AI coding users and the majority use AI assistance occasionally. Per-seat subscriptions charge the same rate for every seat regardless of utilisation; OpenCode's token-only cost structure means lighter users cost proportionally less. For teams evaluating AI coding tooling at the enterprise level, OpenCode's combination of data sovereignty, open-source auditability, and usage-proportional pricing addresses the three objections that most commonly delay adoption in Indian enterprise procurement processes.
The Bottom Line
OpenCode launched on 19 June 2026 from the SST/Anomaly team — a terminal-native, MIT-licensed, model-agnostic AI coding agent that crossed 160,000 GitHub stars and 7.5 million monthly active developers in its first weeks. It supports 75-plus AI providers on a BYO-key model, runs local models via Ollama and llama.cpp for fully on-premises deployments, stores conversation history in SQLite on the local machine, and separates task planning from execution through Plan and Build modes with high-privilege action gating. For Indian engineering teams navigating data governance requirements and enterprise procurement constraints, OpenCode's MIT licence and local inference capability make it the most enterprise-compatible open-source AI coding agent currently available.
Frequently Asked Questions
What is OpenCode and who built it?+
OpenCode is an open-source, terminal-based AI coding agent released on 19 June 2026 by the team at SST, which has rebranded as Anomaly. The founding team includes Jay (CEO), Frank (co-founder), Dax Raad (who joined SST in 2021 and became a co-founder), and Adam Elmore. OpenCode is published under an MIT licence, is model-agnostic (connecting to 75+ AI providers), and stores conversation history locally in SQLite rather than transmitting it to a vendor cloud. Within weeks of launch it crossed 160,000 GitHub stars — the most starred open-source AI coding agent in the category — with 7.5 million monthly active developers.
How does OpenCode differ from GitHub Copilot and Cursor?+
GitHub Copilot and Cursor are IDE-native tools that operate within a host code editor's permission model. They can suggest completions and modify open files, but they cannot arbitrarily execute shell commands, manage background processes, or interact with the full terminal environment. OpenCode runs in the terminal as a first-class process with the same system access as the developer's shell session, enabling it to run test suites, manage git workflows, and observe and respond to build failures within a single interface. It is model-agnostic (no vendor model dependency), local-first (conversation data stored in on-device SQLite), and MIT-licensed — three attributes neither Copilot nor Cursor matches by default.
Can OpenCode run AI models entirely on a local machine without cloud APIs?+
Yes. OpenCode supports fully local inference through Ollama, llama.cpp, and LM Studio, allowing developers to run open-weight coding models such as GLM-5.2, Qwen 2.5 Coder, or Llama 3 entirely on local hardware with no data transmitted to any external service. This capability makes OpenCode viable in regulated enterprise environments — financial services, healthcare software, government contracting, defence-adjacent development — where sending codebase context to external cloud APIs is restricted by procurement policy or regulation. The MIT licence also means organisations can inspect, modify, and self-host OpenCode's codebase without accepting any vendor's privacy terms.
What are Plan mode and Build mode in OpenCode?+
OpenCode separates agentic coding tasks into two phases. In Plan mode, the agent reads the codebase and task description and produces a structured plan — listing files to edit, commands to run, and dependencies to resolve — before taking any action. The developer reviews and approves this plan. In Build mode, the agent executes the approved plan step by step, with high-privilege actions such as git commits, writes to sensitive directories, and file deletions requiring explicit human confirmation before proceeding. This two-stage approach externalises the agent's assumptions before irreversible actions occur, reducing the error-compounding failure mode common in agents that collapse planning and execution into a single undifferentiated loop.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.