@observee/agents

Abstracts tool/function calling across multiple LLM providers (OpenAI, Anthropic, Gemini, Ollama) through a unified schema-based interface. Translates provider-specific function calling formats (OpenAI's tools array, Anthropic's tool_use blocks, Gemini's function calling) into a normalized capability model, handling request/response marshaling and provider-specific quirks automatically.

Implements the Model Context Protocol specification to expose tools and resources as standardized MCP servers that can be discovered and invoked by MCP-compatible clients. Handles MCP transport (stdio, SSE), resource management, tool registry, and request/response serialization according to the MCP specification, enabling interoperability with Claude Desktop, other MCP clients, and MCP-aware frameworks.

Orchestrates agentic loops that repeatedly call LLMs, parse tool calls from responses, execute tools, and feed results back into the conversation context. Implements the core agent pattern with automatic tool call detection, execution, and result injection, supporting both streaming and non-streaming LLM responses, error handling for failed tool executions, and configurable stopping conditions (max iterations, tool call completion).

Handles streaming LLM responses and parses tool calls from streamed token sequences, enabling real-time display of agent reasoning and tool execution progress. Buffers streamed tokens, detects tool call boundaries (e.g., Anthropic's tool_use blocks in streaming), and yields partial results as they become available, supporting both text streaming and structured tool call extraction from incomplete streams.

Validates tool definitions against JSON Schema and provider-specific requirements, ensuring tools are compatible with the target LLM provider's tool calling format. Performs schema validation, parameter type checking, and provider-specific constraint validation (e.g., OpenAI's 4096-char description limit, Anthropic's input schema requirements), providing detailed error messages for schema violations.

Manages conversation history and context windows for multi-turn agent interactions, tracking messages, tool calls, and results in a structured format. Provides utilities for building conversation context, managing message ordering, and preparing context for LLM API calls, but does not include automatic context trimming or summarization; applications must manage context window limits explicitly.

Implements error handling for tool execution failures, including automatic retry logic, error context injection into agent loops, and graceful degradation when tools fail. Catches tool execution exceptions, formats error messages, and optionally retries failed tool calls with exponential backoff, allowing agents to recover from transient failures or adapt when tools are unavailable.

Provides TypeScript type definitions and generics for tool definitions, tool call responses, and agent outputs, enabling compile-time type checking and IDE autocomplete for tool parameters and results. Uses TypeScript's type system to enforce tool schema compatibility and provide type-safe tool execution handlers with inferred parameter types.

Provides integration points with LangChain's agent framework, tool definitions, and LLM interfaces, allowing Observee agents to be used within LangChain applications or vice versa. Implements adapters for LangChain's Tool class, AgentExecutor, and LLM interfaces, enabling interoperability between Observee's agent orchestration and LangChain's broader ecosystem.