इस प्रोजेक्ट के बारे में

BareAgent Node.js ke liye ek halka-phulka, zero-dependency agent orchestration library hai jo AI agents ke liye think→act→observe loop pradan karti hai. LangChain, CrewAI, ya AutoGen jaise bhaari frameworks ke viparit, BareAgent ek monolithic framework ke bajay composable building blocks par kendrit hai, aur iska core kisi bhi cheez ko import nahi karta. Core capabilities: Loop: Kendraki think→act→observe chakra jo goal poora hone tak chalta hai, policy governance, context engineering, aur transcript trimming ke liye opt-in seams ke saath. Planner & runPlan: Goals ko step DAG mein todta hai aur dependency awareness aur prati-step retry ke saath steps ko parallel waves mein execute karta hai. recurse (RLM): Ek recursive decomposition primitive jo kathin tasks ko todta hai, sub-tasks ko fan-out karta hai, results verify karta hai, aur final answer synthesize karta hai. Model-driven spawning, forced fan-out, aur data-driven partitioning support karta hai. Corpora par honest counting scan-based retrieval ke through karta hai jo model ke bataye numbers par bharosa karne ke bajay code-counting karta hai. Memory: Swapnable Store interface (JSON, SQLite, ya litextx) ke through sessions ke beech persistent storage aur recall. Governance: Optional bareguard integration sabhi LLM aur tool calls par ek single policy gate pradan karta hai, audit logging, budget caps, aur human approval checkpoints ke saath. Resilience: Backoff aur jitter ke saath retry, circuit breaker, aur provider failover. Providers: OpenAI-compatible APIs, Anthropic, Gemini, Ollama, CLIPipe, aur custom providers ko support karta hai. Sabhi normalized shape mein honest token/cost metrics return karte hain, including prompt-cache tier reporting. Tools: Koi bhi function tool ho sakta hai. Built-in web browsing (barebrowse), mobile control (baremobile), aur shell execution. MCP Bridge IDE configs se MCP servers ko auto-discover karta hai. Cross-language: Python, Go, Rust, Ruby, ya Java se stdin/stdout par JSONL ke through subprocess ke roop mein chal sakta hai. Streaming: Real-time progress ke liye structured JSONL events. Evaluator & refine: Output ko predicate, rubric, ya agentic critic se judge karta hai, bounded generate→evaluate→regenerate loop ke saath. Scheduler: Cron ya relative triggers, restarts ke beech job persistence ke saath. Checkpoint: Bring-your-own transport ke saath human approval gate. Spawn: Shared audit log aur budget ke saath child agents fork karta hai. Defer: Baad mein execution ke liye actions ko queue karta hai waker dwara. SkillRegistry: Meta-tool catalog ke through on-demand skills surface karta hai. stash: Finished work ko live context window se bahar compact karta hai ya token pressure ke niche auto-fold karta hai. Library itni halki hai ki use pooree tarah samjha ja sake, lekin itni complete bhi hai ki pahiya dobara invent na karna pade. Iske saath zero required dependencies ship hoti hain; optional peers mein governance ke liye bareguard, SQLite storage ke liye better-sqlite3, aur scheduling ke liye cron-parser shamil hain. Licensed under Apache 2.0.