ECAssistant.Core 12.9.13

ECAssistant.Core

The embeddable .NET agent library. Sessions, tools, memory, streaming inference — a few lines of code turn any .NET 8 app into an intelligent, tool-using agent.

Part of ECAssistant — small, lightweight, open source. Your models, your keys, your machine.

Install

Public on nuget.org — no token, no auth:

dotnet add package ECAssistant.Core

Hello, agent

using ECAssistant.Core.Composition;
using ECAssistant.Core.Engine;
using ECAssistant.Core.Session;

// One call wires config, model resolution, tools, memory and inference
var root = new EcaCompositionRoot(userConfigDir, args);
var services = root.Build();

// Create a session and register an output listener (streamed tokens + tool events)
var sessions = new SessionManager(services.Config, services.ModelPath, workingDir, services.Logger);
var session = sessions.CreateSession("main");
session.AddListener(myListener);          // implements IOutputListener

await services.SessionBuilder.BuildAsync(session, externalTools: null);

// Run the agent — it plans, calls tools, and streams its answer
var loop = new EDecisionLoop(session.Engine, session);
await loop.ExecuteInteractiveLoop("Summarize the docs in this folder");

A complete, runnable wiring example lives in ECAssistantConsole.

What you get

What makes it different

• Vision structure extraction (EVisionStructure) — point the agent at a screenshot, UI image, or PDF page and get back a fixed, versioned JSON schema (schemaVersion 1.0): elements (headers, labels, buttons, inputs...) with approximate bounding boxes, label↔control associations, and semantic groups. Never-null design: unknown enums map to Other, missing fields get defaults, dangling refs are stripped — downstream code can consume it blind. Server-side GBNF grammar enforcement makes the shape physically guaranteed, and a deterministic validator normalizes semantics on top.
• Grammar-forced structured decisions — the agent's act/answer/toolcall decisions are token-level constrained (GBNF), not prompt-asked. Valid tool calls with typed JSON Schema parameters, every time.
• Interactive checkpoints (EAskUser) — the model escalates genuine ambiguity to a real choice prompt instead of guessing; falls back to autonomous mode when unattended.
• Self-correction with loop detection — malformed outputs trigger error-feedback retries; long-range repeat loops are detected and stopped before they burn your budget.
• Thin by design — 2.8 MB tool, server fetched on demand; local-first privacy with a remote escape hatch in the same code path.

Architecture boundary (by design)

Core talks to the ECAssistantLLM server exclusively over OpenAI-compatible HTTP. It knows nothing about model loading or runtime internals — the server is a self-contained, separately versioned product. Point llm_provider at any OpenAI-compatible endpoint and Core doesn't care what's behind it.

• No in-process LLamaSharp — zero native model dependencies in your project
• No embedded blobs — even the ~170 MB server is fetched on demand at wizard time, never bundled
• No telemetry — nothing leaves your machine except the LLM calls you configured

Docs & architecture

• ARCHITECTURE.md — full class map, dependency flow, setup state machine
• API-INDEX.md — per-package API reference

Related repos

License

MIT — © 2026 SideDevEC