Skip to content

Overview ​

Logician is a local-first coding agent built on a modular architecture. This page explains how the pieces fit together.

Design philosophy ​

  • Local-first runtime — sessions and tool execution stay local; the model endpoint can be local or hosted.
  • Streaming-first — provider text and tool progress are visible as they arrive.
  • Safe by default — edits use strict exact-text matching with CRLF/BOM preservation. Permission modes control how aggressively the agent acts.
  • Extensible — skills (SKILL.md files), plugins, and MCP servers extend capabilities without code changes.

Component architecture ​

Core packages ​

PackageResponsibility
log-coreAgent loop, harness, context, configuration, sessions, hooks, compaction, tools, and versioned client notifications
log-runtimeRuntime composition: capabilities (reasoning, delegation, tasks, ask, RAG tools, built-in tools, memory wiring, LSP, MCP, skills) plus orchestration (bridge, session, transcript)
log-eohEvolution of Heuristics — standalone optimization engine, wired into log-runtime's capabilities
memoriamStandalone Python memory engine (ecosystem/memoriam): SQLite-backed observation capture, consolidation, retrieval; embedded via its JSON-lines SDK worker
log-ragHybrid dense + BM25 retrieval, chunking, reranking, context budgets
log-autoresearchMeasured experiment loops — run, evaluate, keep or discard
log-evalOutcome-grounded evaluation runner for agent trials
tuiTerminal UI components, engine, layers, state management (packages/log-tui)

Key concepts ​

  • Agent loop — the core cycle: receive input → build system prompt → call LLM → parse response → execute tools → repeat.
  • Skills — SKILL.md files that inject specialized instructions into the system prompt when triggered.
  • Hooks — lifecycle callbacks (before/after tool calls, before/after LLM requests, etc.) for plugins.
  • Sessions — append-only JSONL conversation trees with bookmarks, branching, recovery, and compaction.
  • Trust model — permission modes (acceptAll, acceptEdits, ask, plan) control agent behavior.

Why local-first? ​

Logician keeps orchestration, tools, and session storage on your machine. Network exposure depends on the model endpoint, MCP servers, web tools, and plugins you configure.

This means:

  • Local model endpoints can keep model traffic on your machine
  • Secrets can remain in ~/.logician/.env rather than project files
  • Core editing and session workflows do not require a hosted Logician service
  • Full control over data retention and session history

MIT License