AI workflow coach · local-first · tool-agnostic

Steer AI coding agents
before they make a mess.

A local-first CLI with plugins for Claude Code and Codex that coaches engineers and non-technical builders through a disciplined workflow — Research, Propose, Plan, Implement, Test — and meters what every session actually costs.

“Like a calm senior engineer sitting next to you in the terminal.”

~/my-project — bash
The foundation

What breaks agents

Everything your team works around is something that breaks an agent. Took audits your repo for exactly these gaps and frames every finding as a fixable thing.

No tests
Agent can't verify its own changes
Add tests before you start. Took tells you which.
Inconsistent patterns
Agent picks the wrong one and spreads it
Document your conventions in AGENTS.md
Stale context docs
Agent hallucinates APIs that changed
Took flags docs/playbook drift after every diff
No hermetic build
Agent can't tell if it broke anything
`took ready` surfaces build + lint gaps
Vague AGENTS.md
Agent invents rules you didn't intend
Took helps you write a machine-readable playbook
Context overrun
Quality collapses in the "dumb zone"
Took warns before you hit ~170K tokens

$ took ready — runs an agent-readiness audit on your repo and produces a prioritised fix list. Free, local, no account needed.

The methodology

The workflow loop

Research → Propose → Plan → Implement → Test. The six-step discipline that turns vague prompts into production-grade, verifiable code — and coaches you through every stage. Built on the foundation laid by Mihail Eric at The Modern Software.

R
Research

Produce the document the agent plans from. No research → no plan → slop.

P
Propose

Force alternatives. Only one solution considered is a red flag.

P
Plan

A reviewable spec with a verifiable done. The agent's contract.

I
Implement

Externalise progress on large tasks. Protect the context budget.

T
Test

Favour automatic verification. Missing tests = untrusted code.

Max
S
Secure

Pipp Health only. HIPAA/SOC2/HITRUST gate after Test.

RResearch

Before the agent touches code, you (or Took) generate a research artifact: current API docs, existing patterns, affected files. Without it the agent hallucinates.

PPropose

Took coaches you to produce at least two or three approaches with trade-offs before committing. Alternatives prevent tunnel vision and surface risk early.

PPlan

A plan document scopes the diff before it happens: files in scope, files off-limits, acceptance criteria. Took checks the agent stays inside the lines.

IImplement

Took guards each tool call — scope blowouts, dangerous commands, near-dumb-zone context warnings. Large tasks checkpoint progress to artifacts so nothing is lost to compaction.

TTest

Took's diff check catches missing test coverage and flags risky untested diffs. PR readiness scores tests as a first-class signal.

SSecurePipp Health

In regulated mode, a sixth gate runs compliance checklists against the diff. Loop Implement → Test → Secure until clean. A green diff never auto-passes what code can't prove.

Right-sizing

Two-axis governor

Over-coaching is a failure mode too. Took scales how closely it watches to the task and the operator — so a one-line fix never gets the full loop, and a healthcare migration never skips the Secure gate.

Axis 1

Rigor ↔ Task Size

A one-line fix gets a light touch. A cross-service migration gets the full loop. Took never over-coaches trivia or under-coaches complexity.

Trivial·rename a variable
Just do it
Small·add a field
Plan only
Medium·new feature
Full loop
Large·migration
Full loop + checkpoints
Axis 2

Protection ↔ Operator

Engineers get Took in Coach mode. Non-technical builders default to Guard/Strict. Regulated teams get the Secure gate. Right-sized, not maximum.

Engineer·experienced dev
Coach (warn-only)
Builder·non-technical
Guard (confirm)
Regulated·healthcare team
Strict + Secure
Team·shared rules
Policy-controlled
The honest meter

Know what your agents cost — without the marketing math

Took meters every session from the agent's own transcript and refuses to flatter you: costs are list-price equivalents, savings carry a named basis, and what Took caught is counted — never priced. No causal ROI claims, ever.

took cost — the local ledger
$ took cost
Cost $11.54 list-price equiv · 3.1M tokens · 14 sessions
# metered list price — not billed on a flat subscription
Prompt caching
saved $4.31 (~90% off re-read context) — caching working as intended
Recall
Context avoided via recall: 41.2k tok (~$0.12) · 6 recalls · est. vs dumping matched files
What Took caught
9 interventions (2 blocks · 3 corrections · 4 right-sized) — counted, never priced
By type
feature: $4.80 · fix: $2.31 · chore: $1.31
Top tasks
$3.95 tk-0007 (feature) checkout retry flow
Context pressure
avg 38.4k tokens/turn — 23% of the 170.0k smart window
# Coverage: 73% of spend is pinned to a task ($8.42 of $11.54).
$

The cost ledger

Real tokens and dollars per session, read from the agent’s own transcript — never estimated from prompt counts. Cost per task and per type via the git-native task graph, plus context pressure against the smart window. Figures are list-price equivalents, not billing claims.

Savings, with a named basis

Recall reports tokens avoided versus dumping the matched files; the cartridge reports a per-session estimate against its build baseline. Both are labeled estimates — never billing claims. Prompt-cache savings are credited to your provider’s caching, not to Took.

What Took caught

Blocks, warns, corrections, right-sizing, compactions — counted, never priced. There is no honest counterfactual for a blocked operation, so Took refuses to invent one.

The learning curve

took learn charts corrections, right-sized prompts, and waste week over week — computed from your own local event log, so you can see the habit change, not just be told about it.

The context-discipline trio

Recall

took recall

Pull only the symbols a task needs — a zero-dependency structural map serves budget-fit context slices instead of whole files.

Cartridge

took cartridge

Freeze repo knowledge into a slim, source-grounded artifact — faithfulness-gated, with a labeled per-session savings estimate.

Task graph

took task

A git-native task graph with commit-trailer lineage — the thing that turns raw spend into cost per task.

Optional cloud view

The dashboard — app.tookbot.com

Log in and sync to see the same ledger in the browser: an overview, the /cost ledger, and /learn trends — plus readiness, weekly reports, badges, and your profile. Sync is opt-in and privacy-tiered: derived metrics only — never code, prompts, or diffs.

Open the dashboard
The layers

One coach, three layers

Took is free for engineers. Pipp extends it for builders who want to understand what they're managing. Pipp Health adds regulated-industry compliance for healthcare teams.

Took

Engineer Core
Freelocal CLI · forever

Steer AI coding agents before they make a mess.

  • Workflow coaching
  • Pre-tool guard + diff review
  • Agent-readiness audit
  • Docs & playbook drift check
  • PR readiness score
  • Local XP, cost & learning ledgers
Install free

Pipp

Education Layer
$15per dev · Max

Build with AI and actually understand what's happening.

  • Plain-English narration of every agent action
  • Leveled, cited learning resources
  • Heightened scanning & guardrails
  • Concept tracking — stops re-explaining
  • Learning XP + profile
  • Free taste: took explain works for everyone
Learn about Pipp

Pipp Health

Regulated Add-on
In Maxno add-on

Ship healthcare software with compliance checks built into the workflow.

  • Secure phase
  • HIPAA / SOC 2 / HITRUST assessment
  • PHI/PII standing scan
  • Code-verifiable vs org controls
  • Retained compliance audit trail
  • Composable with Pipp
Explore Pipp Health
Claude Code + Codex

Install in your agent

Took ships as a Claude Code plugin — and a Codex CLI plugin from the same repo — hooking five agent events to coach without interrupting your flow.

claude-code — plugin install
# Claude Code plugin marketplace
$ /plugin marketplace add gettook/took
Found: Took · coaching hooks for Claude Code
$ /plugin install took@took
UserPromptSubmit hook · registered ✓
PreToolUse hook · registered ✓
PostToolUse hook · registered ✓
PreCompact hook · registered ✓
Stop hook · registered ✓
Took is watching. A coach, never an agent.
$

Running Codex CLI instead? The Codex install guide →

UserPromptSubmit
Right-sizes the prompt and coaches Research-first for non-trivial tasks.
PreToolUse
Guards scope blowouts, dangerous commands, and context budget warnings.
PostToolUse
Reviews the diff for missing tests, docs drift, and out-of-scope edits.
PreCompact
Encourages caching progress to an artifact before context compaction.
Stop
Ingests real token + cost usage from the session transcript into the local ledger; auto-syncs derived metrics when logged in.

Plus took statusline — one quiet line of prompt right-sizing + XP in your Claude Code status bar.

Took XP

Judgment rewarded, not tokens burned

Took XP rewards quality, judgment, and learning — never prompt count, diff size, or token volume. Right-sizing a task earns a Judgment bonus. Running the full loop when it was warranted earns Research, Plan, and Test XP.

Anti-gaming: per-category daily caps, evidence required (real artifacts, tests, surviving PRs), and improvement deltas rewarded over absolutes — so messy starters are rewarded for growth.

Research+25 XP

Generated a research doc before prompting

Plan+20 XP

Scoped the diff with a verifiable done

Guard+15 XP

Prevented a dangerous out-of-scope edit

Test+20 XP

All changed files have test coverage

Judgment+30 XP

Right-sized a trivial task, skipped loop

Learning+10 XP

Mastered a new concept via Pipp

Agent-ready ✓ score badge for your README
·
Compliance XP for Pipp Health
·
Portfolio export for non-technical builders

Start steering your agents.

Free local CLI. No account. No cloud. No surveillance. Just a calm senior engineer in your terminal — run took start for the 60-second tour.