agent-takkub weekly downloads statistics

🛩️ agent-takkub

Your AI dev team — in one desktop cockpit.

Prompt one _Lead_ agent. It plans the work, splits it across specialist teammates, runs them in parallel as real claude processes, and verifies the result — while you watch and steer.

100% local · runs on your logged-in Claude Code CLI · no SaaS middleware

🖥️ The Desktop Cockpit

One window: you talk to the Lead, and it spawns and drives specialist teammates (frontend · backend · qa · reviewer · devops · …) as live Claude Code panes.

✨ Why agent-takkub?

A single AI agent hits a wall on big work: context fills up, sub-tasks collide, and everything runs one-at-a-time. agent-takkub runs it like a real engineering team — a Lead you talk to, and specialist teammates it delegates to, each in its own isolated claude process, working concurrently.

Orchestration

| | | | :-- | :-- | | 🧠 Orchestrated teammates | Converse with the Lead; it spawns, tasks, and manages specialist panes (frontend, backend, qa, reviewer, devops, mobile, …) on demand — only the roles a job actually needs. | | 🔀 True parallelism | frontend and backend build a feature at the same time; QA always verifies last, against the real running stack. | | 🌿 Branch & worktree isolation | Parallel teammates each work on their own git branch in an isolated worktree — no commit races, no dirty-state collisions. You merge when ready. | | 👥 Fleet mode | One toggle scales a role into a fleet (frontend#1…#K) sized to your machine — for many independent features or sharded test suites at once. | | 🧭 Plan-first shards | takkub assign --role qa --plan --shards N — a planner splits a big browser-QA sweep into N buckets before fanning out, instead of guessing an even split. | | 🖥️ Steerable, always | Every pane is a live claude shell. Watch output in real time, interrupt, or type straight into any teammate. | | 🗂️ Multi-project tabs | One isolated Lead per project — no cross-talk. | | 🔒 100% local | No SaaS middleware. Everything runs on your machine, on your logged-in Claude Code CLI. |

Token efficiency & memory

| | | | :-- | :-- | | 📥 Pull-on-demand context | Reference docs, patterns, and CLI cheatsheets live outside the Lead's default prompt — it reads them only when a task needs them, instead of paying for them on every turn. | | 🧾 Per-role memory + L2 archive | Each teammate keeps a short, token-budgeted "learned notes" file; entries that get trimmed for space are archived (not deleted) to a searchable sibling file instead of being lost. | | 🔎 BM25 search (takkub search) | Full-text search over session logs and role-memory archives, ranked with Okapi BM25 — tokenizes both English words and Thai character trigrams, no external segmenter. | | ✅ Evidence-gated QA/review | A verify role's "done" report is checked for actual evidence (a file, a test run, an exit code) — reports with no evidence cited get flagged for the Lead instead of trusted at face value. | | 🧬 Structural code-intelligence (graft) | Symbol search, call-graph tracing, and signatures-only file views wired into code-reading roles (frontend/backend/mobile/devops/qa/reviewer/critic) — no LLM call, ~92% smaller than reading a large file whole. Fully optional: install with takkub doctor --fix (needs Node ≥ 20); a pane never gets the extra tools unless a project's graph has actually finished building. The graph lives outside the repo (~/.agent-takkub/graft-graphs, never committed); disable entirely with TAKKUB_SKIP_GRAFT_BUILD=1. |

Reliability & diagnostics

| | | | :-- | :-- | | 🩺 takkub doctor --live | Checks the running cockpit's spawn queue, CPU/RAM governor, resource waits, PTY queue depth, stale-write drops, and remote mirror — catches live wedges that config-only checks cannot see. | | 🛡️ Reliability v2 | Delivery IDs, pane-session generations, TTL, single-flight submission, bounded priority PTY queues, fair heavy-work admission, adaptive hidden-pane rendering, durable completion dedupe, and Windows Job Object cleanup. | | ⌨️ Multiline input that just works | Shift+Enter / Alt+Enter add a newline instead of submitting, provider-aware (works on claude/gemini panes; skipped where the underlying TUI would misinterpret it) — includes a macOS IME fix for Thai/CJK input dropped after switching input languages. (community contribution by @than-aa) |

🧠 One team, many model "brains"

Model diversity beats a single point of view. Takkub lets the Lead pull in a second, third — or sixth brain for planning, review, and cross-checks, and it never breaks if you don't have them installed.

| Brain | Backed by | Great at | | :-- | :-- | :-- | | 🟣 Claude | Claude Code CLI | The Lead + every specialist — build, test, review | | 🟢 Codex | OpenAI Codex CLI | Second opinion · refactor patterns · cross-checking a plan | | 🔵 Gemini | Google Antigravity (agy) | Long-context planning (reads the whole repo) · a third perspective | | 🟠 OpenCode | sst OpenCode | One CLI, 75+ model backends (GLM · DeepSeek · local Ollama …) | | ⚫ Cursor | Cursor CLI (cursor-agent) | Pick per-task from Cursor's own model roster |

Pick the model, not just the CLI. Every provider — and every *role* — can be pinned to a specific model from Settings → Providers & Roles, or from the terminal:

A role's own model wins over the provider default, so backend can run Codex on gpt-5.6 while reviewer runs it on something cheaper.

Never a hard dependency. If a provider isn't installed (or you've toggled it off), the Lead keeps the role — Claude transparently stands in, and tells you you've traded away model diversity. No refusals, no dead ends.

⚠️ Cursor is new in 1.0.27 — they spawn and take tasks, but its idle/busy screen markers are not calibrated yet, so prefer Claude/Codex/Gemini/OpenCode for roles you leave unattended.

⚡ Quick Start

⚠️ Install it globally — the -g flag matters. It provisions the isolated runtime and the Desktop launcher. A plain npm install agent-takkub (no -g) will not set the app up.…

Description from the npm listing.