ByteRover

OpenClaw triggers security fears and honestly, that concern makes sense. When an agent can access your files, credentials, or run automation, the real risk is unintended behavior. ByteRover already gave OpenClaw persistent memory through our agent skill so OpenClaw understands and follow your intention better this also includes behavior and safety rules - the constraints you want an agent to always respect. The problem was how that memory ran. The setup still depended on a running CLI, which isn’t ideal once agents restart or run in the background. That’s why we added headless mode. With headless mode, ByteRover runs quietly in the background and persists clear safety rules, like: - don’t read .env or secret files - don’t access certain directories - don’t run actions outside approved paths Those rules stick around across restarts and apply whether you’re using OpenClaw for coding, automation, research, or other everyday tasks. Headless mode turns ByteRover into always-on guardrails, not just memory. If you’re excited about agents but uneasy about safety, this was built for you. 👉 ByteRover Headless Mode Skill in comments

We just added Agent Skill for ByteRover CLI in our 1.3.0 release 💫 This integration brings better token efficiency compared to MCP or rules, and more specialized ByteRover's context-related capabilities for the coding agents. Agent Skills are quickly becoming the standard extension of specialized capability for the agent. After ClaudeCode, major players like Cursor, Codex, GitHub Copilot, Google's Antigravity started to adopt Skills for their agents. Unlike rules or MCP, which are always preloaded into the context window, Skills are loaded dynamically when the agent determines that a particular Skill is relevant. Therefore, with ByteRover Skill, your coding agent can now keep context window clean while gaining specialized capabilities of ByteRover for context-related tasks. Skill is now available in ByteRover CLI 1.3.0. You can check our full release note down below! Check it out here: https://lnkd.in/guSDSUfG

ByteRover CLI can now be connected to coding agents via MCP Previously, static instruction files were the primary way to integrate ByteRover across agents. What we’ve consistently seen is that developers prefer using ByteRover CLI to actively manage, curate, and query context for their agents. That’s why we’re expanding our integration surface. With MCP, ByteRover CLI connects to agents through a standard interface without changing the core workflow: 🎯 Agents interact via ByteRover tools like brv_curate and brv_query, without relying on full rule files 🔀 Claude Code users can choose between Hook or MCP, depending on what fits best Check out the full release notes in byterover.dev

Stop treating your codebase like a PDF. We recently ran an experiment comparing standard Vector RAG against an agentic search approach using Context Tree (structured, intent-aware navigation instead of similarity search) on a production codebase of ~1,300 files. The results challenged the industry default. While Vector RAG is great for unstructured text, we found that embedding similarity fails to capture code relevance. The vector approach consistently retrieved noise: test files, deprecated code, and keyword matches that lacked semantic relationship to the task. When we switched to a domain-structured Context Tree approach, the metrics improved drastically: 📉 ~99% reduction in tokens per query 🎯 ~2× better precision on implementation-level questions ✅ ~2× better overall accuracy when balancing false positives and misses The most interesting finding? Vector RAG actually had decent recall, but it came at the cost of context pollution. By flooding the context window with similar but irrelevant code, the agent's reasoning ability degraded significantly. The takeaway: If your coding agents feel confused, hallucinate imports, or mix up file versions, the issue probably isn’t the model. It’s the search strategy. I wrote a deeper breakdown of a deeper breakdown of how we tested this, why context trees help, and where vector RAG still makes sense in the comment session 👇

ByteRover CLI 1.1.0: ClaudeCode Hooks & Sub-Agent Search 🚀 We are back to our weekly release cycle with two major upgrades focused on agent reliability and efficiency: - ClaudeCode Hook Integration ensures your agent maintains context dynamically by injecting instructions on every request. - Sub-agent Search tools reduce token costs drastically and improve response times by retrieving only the exact information needed, rather than loading the entire Context Tree. Read the full updates note here: https://lnkd.in/gB7QNmA6

ByteRover: 2025 Year in Review 2025 was the year of the coding agent. It was also the year we learned that agents still aren’t ready for real production work. “Build this in 1 hour” makes for great demos. But professional software development isn’t about demos. It’s about maintainable systems, clean architecture, and teams shipping together over time. Most agents still fall short there. Everyone shipped similar features. Very few focused on the real problem: 💢 developers still can’t manage, share, or reuse context across agents and teams. The conversation is finally shifting from better models to better context. ByteRover has been building for that shift from day one. 🧩 Our Year in Product 💫 ByteRover 1.0: Shared Memory Layer. We started with a simple bet: developers need a shared memory layer that works across agents and teams. We built on vector DBs, graph DBs, and MCP to make context portable and reusable. The response validated the thesis, strong interest from the developer community, teams adopting ByteRover as infrastructure, and inbound from S&P 500 companies. 💫 Our AI engine, Cipher, Goes Open Source We open-sourced Cipher, ByteRover’s core agent. It reached 3,000+ stars and countless forks, quickly becoming a foundation other memory solutions are now built on. That signal mattered to us. 💫 Context Composer & Git for AI memory We shipped the MCP Aggregator, giving developers a single interface to connect multiple MCP servers and compose clean, agent-ready context, and build "Git for AI memory" - a new workflow with context that allows teams to version control, and conflict resolve context just like Git for code. 💫 Finally, ByteRover 3.0: A New Engine Then we made the hardest decision of the year: we threw it all out and rebuilt from scratch. We moved away from MCP, vector DBs, and graph DBs entirely. The new architecture introduces an intelligent context-engineering agent, a file-and-folder schema we call the Context Tree, and a system optimized for precision, search accuracy, and token efficiency. 🧩 Our Year in Thinking: Vector DB was the wrong abstraction. Vector search is optimized for recall, but coding agents need precision. Vector DBs work well for long-term facts, while development context changes constantly as work evolves. Popular doesn’t mean correct. Developers need to curate, not just retrieve. Professional developers don’t want a black box. Human-in-the-loop isn’t a fallback, it’s the design. Context needs to be shaped intentionally, not inferred blindly. We also learned that context isn’t just for code. Developers use it for planning, documentation, reviews, and communication. The system has to support all of it. We chose the harder path. We could’ve stayed on the popular stack and shipped faster, but rebuilding the architecture was the only way to build what we actually believe in.

Today, we’re launching ByteRover 3.0. After two months of intensive beta testing, ByteRover CLI is now out of beta and fully production-ready. This launch brings our engineering milestones throughout beta period: a new interactive REPL interface, a massive upgrade in context precision through the Context Tree and Agentic Search. You are now experiencing a new level of precision in every context that coding agents are able to use. 👉 Check it out now at byterover(.)dev Please share with us your thoughts once you try it out.

We’re hiring: Agent Engineer at ByteRover Together with our team, let's build a better Context Engineer Agent! Who are we? We are a VC-backed start-up, solving one of the most interesting challenges in agentic coding today - context engineering. 💫 What you’ll work on: - Shape the core technology behind our Context Engineer Agent, an agent that learns continuously from the interactions between developers and coding agents to transform those interactions into structured, reusable context that dev teams can build upon. - You’ll tackle one of the hardest challenges in the agentic coding field today: optimizing the signal-to-noise ratio within the coding agent’s context window. 💫 What you get: - Top-of-the-market comp + equity - Working alongside experienced engineers and product folks who’ve shipped real things before, and are now aiming even bigger. - No politics, just real collaboration. We keep things open, honest, and direct, and just focus on getting things done. - Growing like crazy. We care about speed, not just in shipping code, but in helping you grow technically, professionally, and personally here. 💫 We’re looking for someone who: - Full-stack capable across the agent stack. You get the full picture, from LLM orchestration, vector memory (like Pinecone or Weaviate), tool/action execution (MCPs), to managing agent state. - High-velocity mindset. We ship weekly at ByteRover. We like small iterations, fast feedback, and improving constantly. - Fluent in one programming language. We’re currently working with JS/TypeScript, but we believe any good dev can pick up new languages when needed. - Is AI-native (Cursor, Claude, Windsurf, etc.) - Curious & adaptable. You’re not afraid to dive into unknown territory and learn fast. 📩 How to apply: send your resume and cover letter to andy@byterover.dev. If you have any question regarding the position, feel free to share too.

We are excited to release ByteRover CLI 0.4.0 with a new interactive REPL. You can now: - simply use slash commands for all operations - run multiple “brv curate” to curate multiple context at the same time. - track all curate and query execution by agents at one place. This new RELP interface delivers an extraordinarily fast, seamless, and agent-consistent context curating / querying experience. Check out latest blog to learn more! https://lnkd.in/ghR-ehER