@toolkit-cli/toolkode

v1.10.0

Published

21 hours ago

The AI engineering terminal. Free.

0High
0Medium
0Low

roseyballs

toolkode

The AI engineering terminal. Free.

Write code. Run agents. Ship faster.

Install

npm i -g @toolkit-cli/toolkode

Or via Homebrew:

brew install aaronmrosenthal/tap/toolkode

Usage

# Start the TUI
toolkode

# Start in a specific directory
toolkode /path/to/project

# Continue the last session
toolkode --continue

# Run a one-shot prompt
toolkode --prompt "fix the failing tests"

# Start the HTTP server
toolkode serve

What's new in v1.10

v1.10 is the biggest jump since the context rewrite. From v1.6 through v1.10, Toolkode grew from "good terminal agent" into a much more capable engineering environment: it understands long-running work better, remembers more across sessions, audits its own codebase, watches CI, and starts to coordinate across people as well as agents.

v1.10 — Hive

CI awareness is built in. Start a watcher with /watch ci, open the dashboard with /ci, and Toolkode will keep an eye on recent GitHub Actions runs. Failed runs can be pulled, summarized, and diagnosed directly in the terminal instead of bouncing between browser tabs and logs.

Cross-session recall is now real. Toolkode writes session digests from structured memory and lets you search them with /recall. Past work is no longer trapped in old transcripts — goals, decisions, touched files, and summaries are available when you need to resume context weeks later.

Local team presence is visible. Toolkode now advertises active files over the local network and flags conflicts when two people are working in the same area. It is lightweight, private, and degrades gracefully when mDNS is unavailable.

Mission Protocol is finally first-class. /mission now runs through the normal command system instead of a special-case prompt path. That means the feature is reachable, visible in session history, and behaves like the rest of Toolkode's command surface.

v1.9 — Cortex

Code intelligence landed. /search adds semantic code search, /impact estimates blast radius from a file or symbol change, and /health scores the codebase across complexity, dead code, type safety, and dependency freshness. These are fast, practical tools for real repos, not toy demos.

The sidebar became operational. Health status can now stay visible while you work, turning analysis from a one-off command into ambient signal.

v1.8 — Total Recall

Memory stopped being manual. Toolkode now auto-extracts durable decisions, goals, and constraints from structured session memory and persists them to disk as project memory files. The memory layer also injects the MEMORY.md index back into context so saved memories actually influence future turns.

Claude Code memory import is supported. Existing memory files can be read and imported directly, which makes migration easier and lets teams keep continuity instead of starting from zero.

Memory security is much stricter. Path validation, containment checks, and safer memory file handling were hardened so the memory system is more trustworthy in real projects.

v1.7 — Mission Control

Sessions became editable. You can branch, retry, edit, and copy from message history instead of treating every conversation as a straight line.

Mission Protocol arrived. Toolkode gained a typed mission state machine, structured worker lifecycle, synthesis, and verification primitives for multi-agent coordination.

Generation can be suspended cleanly. Session suspend brought background-style generation tracking into the product without forcing you into a separate workflow.

v1.6 — Signal

Context use became smarter. Cache Resonance added visibility into prompt-cache behavior so you can see when repeated work is actually being reused instead of paid for again.

Navigation got faster. Pathfinder made in-session search much quicker, and onboarding plus sidebar polish reduced the amount of UI friction between "I know what I want" and "Toolkode is doing it."

New commands since v1.6

/mission    structured multi-agent execution
/search     semantic code search
/impact     blast-radius analysis
/health     codebase health report
/watch ci   start CI polling + diagnosis
/ci         open the CI dashboard
/recall     search past session digests

Features

Multi-provider — Anthropic, OpenAI, Google, OpenRouter, GitHub Copilot, local models via Ollama
Subagents — Spawn parallel agents for complex tasks with team and batch tools
Mission Protocol — Structured multi-agent coordination with first-class /mission
CI watcher — Watch GitHub Actions with /watch ci and inspect failures in /ci
Memory intelligence — Auto-extracted memory, Claude Code import, and cross-session /recall
Code intelligence — Semantic search, impact analysis, and codebase health scoring
Team presence — Local network presence and file conflict detection
MCP servers — Connect external tools via the Model Context Protocol
Vim mode — Full vim keybindings in the prompt editor
Skills marketplace — 90K+ community skills from skills.sh
Themes — 59 built-in themes
Spec-kit workflow — /specify, /plan, /tasks, /implement
Cron/loop — Schedule recurring prompts within a session
Private by default — No telemetry, no data collection, no hidden network calls

Configuration

Config file: ~/.config/toolkode/config.json or toolkode.json in your project root.

{
  "provider": {
    "anthropic": {
      "apiKey": "sk-..."
    }
  },
  "model": "anthropic/claude-sonnet-4-6"
}

SDK

npm i @toolkit-cli/sdk

import { createClient } from "@toolkit-cli/sdk/v2"

const client = createClient({ url: "http://localhost:4096" })
const session = await client.session.create({})
await client.session.prompt({
  sessionID: session.data.id,
  parts: [{ type: "text", text: "hello" }],
})