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.
Public on nuget.org — no token, no auth:
dotnet add package ECAssistant.Core
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 result = await session.Orchestrator.ExecuteMultiStep("Summarize the docs in this folder");A complete, runnable wiring example lives in ECAssistantConsole.
| Capability | What it means for your app |
|---|---|
| Agent engine | Multi-session orchestration, sub-agents, self-correction, task planning |
| Model-tier adaptive harness | One harness, every model size: small models get step-by-step scaffolding, tighter sampling and strict recipes; large models get a slim profile with more headroom — automatic via model_tier.mode |
| Typed per-tool outputs | Tools speak for themselves: each tool renders its own model-facing projection (builds collapse to verdict + parsed errors, never raw logs) — smaller context, sharper next decisions. Override one virtual method on your custom tools |
| Dataflow toolchains | A later tool call can reference an earlier call's output with {{0}} in the same decision — sequential execution and argument substitution without model round-trips (taught to large models only) |
| Post-edit verification | Every file-modifying edit is followed by a build/test gate (tier-aware depth); failures are fed back to the model to fix, before you ever see the result |
| Playbook memory | Successful multi-step goals are captured as reusable playbooks and replayed on similar future tasks |
| Context pinning | Goals, decisions and the touched-file map survive compaction — long sessions don't lose the plot |
| 11 built-in tools + MCP | Shell, file I/O, code editing, git, dotnet, sub-agents, vision structure, handoff — permission-gated (approve once / always this session / deny — session-scoped, never persisted). Plus MCP server support: connect any external tool server via stdio or HTTP/SSE — zero dependencies, config-driven |
| Custom tools | Implement one interface, register it. That's the whole API. |
| Memory | Vector memory (embeddings) + daily notes + curated long-term memory |
| Local or remote LLM | GGUF via the bundled LLM server, or any OpenAI-compatible endpoint — identical code path |
| Model catalog | Data-driven and pulled live from GitHub at setup time (embedded fallback) — add/update models without code changes |
| First-run wizard | Provisions server + models interactively; nothing downloads at chat time |
- 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 toOther, 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.
- Dataflow chains, grammar-free —
{{N}}output references ride inside plain string args, so the GBNF grammar doesn't change: small models keep single calls, large models compose multi-step pipelines in one decision. - MCP (Model Context Protocol) client — connect any external tool server via stdio subprocess or HTTP/SSE. Zero NuGet dependencies — pure JSON-RPC 2.0. Tools discovered at runtime, wrapped as native
EToolBaseinstances. Per-server approval policy, per-tool whitelist/blacklist,{{keychain:name}}secret resolution. Image content flows through the existing vision pipeline. - Fuzzy tool edits — the code editor tolerates imperfect match text: exact → whitespace-tolerant → line-anchored matching with indentation restoration, and a structured ambiguity error instead of guessing.
- Thinking never leaks — empty or malformed model turns are retried with corrective feedback instead of surfacing the model's internal reasoning to your users.
- Thin by design — 2.8 MB tool, server fetched on demand; local-first privacy with a remote escape hatch in the same code path.
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
- ARCHITECTURE.md — full class map, dependency flow, setup state machine
- API-INDEX.md — per-package API reference
| Repo | What it is |
|---|---|
| ECAssistant | Landing repo & docs |
| ECAssistantLLM | Self-contained local LLM server (also standalone) |
| ECAssistantTUI | Reusable terminal UI library |
| ECAssistantConsole | Reference CLI host — dotnet tool install -g ECAssistant.Console |
MIT — © 2026 SideDevEC