While the major AI labs were locked in their usual race to announce the flashiest new capabilities this winter, a small infrastructure company chose a different lane entirely. Anomaly Innovations—better known among developers for serverless tools like SST—shipped something that, at first glance, looked almost quaint: a terminal-first coding assistant. Open-source. No proprietary model. No vendor lock-in.
The tool, called OpenCode, arrived in early February with a quiet proposition that might prove more consequential than the feature war raging around it: What if developers could use whichever AI model they wanted, swap between them freely, and never send their code to anyone's servers unless they chose to?
It's a question the industry seemed content to ignore until recently. Now, perhaps, not so much.
A Market That Outran Its Own Hype
AI coding assistants have moved past the early-adopter phase with startling speed. GitHub reported that its developer platform had grown to roughly 180 million users as of Octoverse 2025, with someone new joining every second—a drumbeat of growth that makes the old "will developers adopt AI?" debate feel quaint. Stack Overflow's survey data suggests somewhere around 84% of developers now use AI tools regularly, though nearly half remain openly skeptical about accuracy. That's adoption with asterisks.
A UK government trial involving over 1,000 tech workers found productivity gains averaging about an hour per day when using tools like Copilot and Gemini Code Assist. Whether that hour represents real output or just faster code generation—quality versus velocity—remains an open question that most vendor case studies prefer not to examine too closely.
Market sizing estimates land all over the map, as they tend to in fast-moving sectors. One research firm pegged AI code assistants at $3.9 billion in 2024. Another calculated the broader AI coding platforms market at $6.12 billion for 2025. Enterprise AI infrastructure spending hit $86 billion in Q3 2025, according to industry trackers, and IDC has projected that agent deployments could exceed a billion by 2029—a number that sounds absurd until you consider how many developers now treat autocomplete as a baseline expectation.
What the disparate numbers agree on: this isn't a niche anymore.
What Anomaly Built, and Why It Matters
OpenCode doesn't look like the polished chat interfaces most coding assistants favor. It's built for the terminal first—text-based UI, desktop app, and integration through the Agent Client Protocol (ACP), which allows it to work inside JetBrains IDEs, Zed, and other compatible editors. The architecture supports OpenAI, Anthropic, Google, AWS Bedrock, Groq, Azure, OpenRouter, and local models through unified configuration files.
For teams that want to run their own models on-premises, or swap providers based on cost or performance, that flexibility isn't a nice-to-have. It's the whole point.
The system includes Language Server Protocol integration for code intelligence, tool execution, and multi-session workflows. Developers configure agents per-project using opencode.json files. Built-in sub-agents handle different tasks—analysis versus modification, for instance—and the architecture supports the Model Context Protocol (MCP), tapping into a growing ecosystem of structured tools. (The documentation warns about token overhead when connecting certain MCP servers, particularly GitHub's—a detail that suggests the team has actually used this in production.)
Anomaly also offers "OpenCode Zen," an optional managed gateway with billing, bring-your-own-key support, and team controls. It's for enterprises that want the infrastructure handled but not the lock-in. The project moved repositories in September 2025, with active development now under the anomalyco GitHub organization.
Nothing about this screams "disruption." Which might be precisely the point.
The Standards That Could Change Everything
The Agent Client Protocol matters more than a standards announcement typically would. Launched by JetBrains and Zed last October and formalized with a registry in late January, ACP is an open standard allowing any coding agent to plug into any compatible editor. For JetBrains users, this means running external agents—Claude, Codex, OpenCode—without needing a JetBrains AI subscription, assuming the agent speaks ACP.
GitHub moved in parallel. At Universe 2025 in late October, the company announced Agent HQ, designed to orchestrate third-party and first-party agents across GitHub and VS Code. By February, Claude and Codex were integrated directly into the platform's UI.
The message from infrastructure providers: walled gardens are out. Modularity is in.
Maybe. These standards battles have a way of looking inevitable right up until they're not. But the timing aligns with something enterprises have been demanding more loudly—governance. Companies are implementing what amounts to admission control for agents: logging, decision trails, token budgets, sandboxed execution. A research dataset released in February documented nearly 933,000 agent-authored pull requests across major tools—Codex, Devin, Copilot, Cursor, Claude Code. Media reports suggest Stripe runs agents that submit over 1,000 pull requests weekly, though the company hasn't confirmed those figures publicly.
When code contributions from non-human actors reach that scale, auditability stops being theoretical.
The Privacy Problem No One Solved
OpenCode's positioning—what February coverage described as a "privacy-first architecture" that doesn't store code or context—speaks to developer anxieties that keep surfacing in surveys. It also arrives as security incidents accumulate in ways that can't be ignored. Over 29 million secrets leaked on GitHub in 2025, a surge observers have tied partly to AI uptake and looser coding discipline. Academic research published last April demonstrated that attackers could manipulate external references to induce AI tools into generating vulnerable code—a technique dubbed HACKODE.
The regulatory environment is tightening in parallel. The EU AI Act entered force in August 2024, with obligations for general-purpose AI providers kicking in last August. A voluntary code of practice published in July guides transparency, copyright, and safety—OpenAI stated publicly it intended to sign. In the U.S., the Copyright Office issued a report in late 2024 reaffirming that human authorship remains central to copyright claims on AI-assisted works. Litigation around GitHub Copilot's training data continues as of last spring.
These pressures favor tools that let organizations control model choice, data flow, and audit trails. Open-source architectures offer compliance teams visibility that proprietary systems don't—a selling point for regulated industries, or anyone who's read a security incident report lately.
Where This Goes
The coding assistant market is consolidating around a few capabilities that are starting to look essential: context gathering (Sourcegraph's Cody enabled "agentic context gathering" by default mid-year), multi-agent coordination (Anthropic's latest model launched with "Agent Teams" features in early February), and execution autonomy (Cognition's Devin reached a reported $10.2 billion valuation by September). Gartner projected in 2023 that half of enterprise software engineers would use ML-powered coding tools by 2027—a forecast that already looks conservative.
OpenCode's approach—open-source, multi-model, terminal-native—positions it as infrastructure rather than product. Whether that carves out meaningful share against well-funded competitors depends on how quickly standards like ACP and MCP gain traction, and whether enterprises actually prioritize control over convenience. They say they do. Purchase orders will tell a different story.
The fact that Anomaly chose to launch this the same week Anthropic shipped its agent coordination features suggests the team believes the window for open alternatives is narrow.
They might be right. Or the market might decide convenience beats sovereignty every time. That's the bet developers are about to make, whether they realize it or not.
