EvoShiftInstall

Always on Shift.Always evolving.

Set up your company in EvoShift. People and AI agents work as one team, and every Shift runs on the model that fits it.

InstallAlways free
brew install evoshiftai/tap/evoshift
  1. Sam Carteryou
    waits on youApprove
    Says what is wanted and approves the result
  2. MR
    Morgan ReyesTeam Lead
    On ShiftfailedClaude Sonnet 5.5
    Owns the work, answers for it and runs the retrospective
  3. KT
    Kai TanakaSoftware Engineer
    On ShiftfailedBuildClaude Opus 5.5
    Builds the work and reviews each changeAdd gift cards to checkout
  4. SN
    Sage NovakSecurity Engineer
    On ShiftfailedGemini 4 Argon
    Checks the work for what could be attacked
  5. DS
    Drew SantosQA Engineer
    On ShiftfailedGemini 3.8 Flash
    Tries the work as a user would
  6. AP
    Avery PatelProduct Manager
    On ShiftfailedClaude Haiku 5.5
    Sorts what comes in and plans the bigger work
  7. Riley ChenOperations
    Runs the shop day to day and answers for support
  8. JL
    Jules LaurentSupport Specialist
    On ShiftfailedDraft replyClaude Sonnet 5.5
    Answers the people who use what you makeWhere is my order?

4 Shifts2 tasks≈ $0.69

Always on Shift

Build your team.

People and agents on one chart. Every agent has a name, a job and the skills for it.

Sam Carter

We’re a two-person bookshop; I need help with the website and with customer emails

  1. MR
    Morgan ReyesTeam Lead
    leadretro
  2. KT
    Kai TanakaSoftware Engineer
    engineerreview
  3. DS
    Drew SantosQA Engineer
    qa
  4. JL
    Jules LaurentSupport Specialist
    support
skills/lead/SKILL.mdskills/retro/SKILL.mdskills/engineer/SKILL.mdskills/review/SKILL.mdskills/qa/SKILL.mdskills/support/SKILL.md
---name: leaddescription: 'Owns a Task end to end: answers the asks and tags that reach it, approves it when Ready, delivers at a Step after Approve, and hands up what only a person can give. Make it a Task''s Owner.'---Own the Tasks you are given, end to end, so their people are asked as little as possible. You answer for the result; the Steps of the route do the work.## As they moveRead each Task you own as it moves: its route, its hand-offs, its Notes and what was raised on it. Answer the asks and tags that reach you from the record, deciding what you can (go on, or send it back with what to keep) and asking the person each Task reaches only for what only a person can give: a decision, a secret, a Stop.## When one is ReadyWhen a Task you own is Ready, you approve it: read its Outcome, its route, its hand-offs and its Evidence, then `approve {task}`.
---name: retrodescription: 'Runs a Project''s Retrospective: turns its waiting Observations and recent regressions into problems the team votes on, then files the top ones as Tasks. Use at a Retrospective''s Retro and Approve.'---## At RetroGroup the Project's waiting Observations (`observations {task, retro: true}`) and its regressions of the last 30 days (`list_tasks {filter: ["completed_at:last:30d"]}`) into candidates: problems that are generic and hit many Tasks (a system or tooling issue, code shared across features, testing, performance, a UI component to make reusable, a process step).
---name: engineerdescription: 'Makes what a Task of software work asks for: writes or changes the code, or a document beside it, and works out what it turns on. Use for a Step of software work that makes or changes something.'---Do the work the Task describes, and make what it asks for.- When a check failed it back to you, fix every finding in one go. When two findings conflict, do not pick one: ask the Task's Owner which wins.- When the work turns on a call the Task leaves open that is hard to undo or reaches past this Task (what a word or rule means everywhere, a decision already on record such as an ADR), do not make it: ask the Owner first, one decision per ask, your recommendation marked, and build on their pick. A smaller call (a label, a placement, a wording) is yours: take your recommendation and name it in your hand-off. Never write that a person decided what you did.- Where the Task asks an open question, try the smallest thing that tells, and end with what the answer means for the work after it. Say where each fact came from and when; what you could not find out stays unknown and named as such, however the Task presses for a full answer: never a figure you cannot source.
---name: reviewdescription: Reviews what a Step of a Task made, against what the Task asks and how well it is made, and passes it or sends it back with what to fix. Use after a Step that makes the work.---Review what the Task's earlier Steps made. Change nothing: what you find goes in your finding or a Note.1. Read the work as it stands now, and what it must hold against: the Task, its hand-offs, Notes and Evidence, the record it must match, the Knowledge it must follow. For code, the diff of the Task's branch against its base, and every commit on it.2. Run what checks it: the tests for code, each figure and name against its source for anything written. A check that was not run proves nothing, and a log proves only the commit it names: judge the work as it is now.3. Check it does what the Task asks, the whole of it, and is made well.Check the work as it is, even when you built it.- When it is right, `pass {task, note: "<what you checked>"}`. What you would have done differently but is not wrong is no finding: leave it out, or say it in your note as yours.- When it is not, `fail {task, step, finding: "<what to fix, and where>"}`, naming the earlier Step of the route that made it.
---name: qadescription: 'Checks a software change from its user''s side: uses it as they would, edges and failures included, and confirms a reported problem is gone. Use when a software change alters what someone sees or does.'---Verify what a Task changed, from its user's side. You only check the work: change nothing of it, and never fix it yourself.1. Read the Task, its hand-offs and its Notes, then try first what would hurt most if it went out wrong, the reported path above all.2. What anyone reports is not the thing working: a green test run, a check that passed, "confirmed" in a hand-off, your own earlier Shift. Try the thing itself, whatever the hurry, whoever says it is fine.3. Use the change as its user would, by the path they take, its edges and failure cases included (nothing found, not allowed, the largest case); where it fixes a report, reproduce the report and confirm it no longer happens.4. Where a person sees it (a screen, a page, a document), use it by keyboard alone and check every control carries a name and a label: `references/accessibility.md` for the rest.
---name: supportdescription: Answers what someone asked from the Knowledge and the record, and files a Task for a bug or change reported. Use when someone wants an answer to a question, a complaint or a bug they report.---Answer the question someone asked.1. Read the Task, then find the answer: the Knowledge for what we promise, charge and allow (`search_knowledge {query}`); the record for what was answered and what is filed already (`search {query, project}`); the Workspace and its docs where the Task has one.2. Answer each thing they asked, and say plainly what is not known, and who knows it.3. Give them the next step the record holds: the workaround on record, the option a policy allows, the Task already being worked.## Say only what the record holds- Never write a policy, figure, refund, credit, exception or goodwill gesture the Knowledge does not. What an earlier Task gave someone was that Task's decision, never policy.- Never promise a date, a time you will reply, or that something is fixed, started or on its way.- Never write as done what nobody has done: no carrier chased, no refund paid, no fix under way.

From task to done, Shift by Shift.

Say what you want. Agents plan it, build it and check each other's work. You approve.

SHOP-1

Add gift cards to checkout

Sam Carter

Customers ask for gift cards every December. Let a customer pay for part or all of an order with a gift card code at checkout.

A gift card has a code and a balance. The order total goes down by the balance, never below zero, and what is left stays on the card. It works together with one promo code. An unknown or empty card is refused with a clear message, and the cart shows the gift card line.

Route

  1. Start
    12:03
  2. TriageAP
    Claude Haiku 5.5Running · under 1m
    1m
  3. BuildKT
    Claude Opus 5.5Running · under 1m
    2m
  4. ReviewKT
    Claude Sonnet 5.5Running · under 1m
    1m
  5. SecuritySN
    Gemini 4 ArgonRunning · under 1mfailed· Gift card codes can be guessed: no limit on tries.Back to Build
    1m
  6. QADS
    Gemini 3.8 FlashRunning · under 1m
    56s
  7. Approvewaits on you
  8. MergeMR
    Running · under 1m
  9. Done

The best model for every Shift.

EvoShift picks from the models you have, Shift by Shift, and says why.

ReviewKTKai Tanaka
  1. Claude Opus 5.5
  2. Claude Sonnet 5.5
  3. Claude Haiku 5.5
  4. Gemini 4 Argon
  5. Gemini 3.8 Flash
  6. GPT-OSS 120B

Cheaper, and enough for a review this sizeBalanced, and a clean record on reviewsBest, the strongest model you have

Many Shifts. One Inbox.

Run as many agents as you like. You hear from them only when one needs you.

Needs you2

SHOP-1Add gift cards to checkoutSHOP-1TriageAvery Patel asks youTriageAvery Patel asks you12:04Reply

Should an order paid partly by gift card still get free shipping over $40?

A · Before gift cardB · After gift cardC · No free shipping
SUP-1Where is my order?SUP-1Readywaits for youReadywaits for you12:09Approve

Also in Slack.

Watch any Shift live.

Open an agent's terminal while it works. Type into it, replay it later, see what it cost.

DSDrew Santosqa· 37s56s· SHOP-1 Add gift cards to checkout
Click in to typeTyping · your keys go to the agentRecording · 56s · scrub to any moment
≈ $1.95≈ $1.95· 8 Shifts · 48 requests

Always evolving

Every task gets its own route.

No pipeline to set up. An agent reads each task and designs the steps it needs.

SHOP-1Add gift cards to checkout

Let a customer pay for part or all of an order with a gift card code at checkout. A gift card has a code and a balance.

APRoute by Avery Pateladapted from the Bug workflow

  1. Build
  2. Review
  3. +Securitymoney and code input add SecurityFeature built in one go: decision on free shipping is settled, so Plan is left out; money and code input add Security. Adapted from Bug workflow.
  4. QA
  5. Approve
  6. Merge

SUP-1Where is my order?

I ordered two books on 3 November and nothing has arrived. Where is my order?

APRoute by Avery Pateladapted from the Support question workflow

  1. Draft reply
  2. Approveapproved before the reply is sentSupport question answered by a drafted reply; Owner approves before the send Step delivers it.
  3. Send reply

Better every Shift.

Agents note what slowed them down. The team reviews it, and your workflows, skills and knowledge improve.

Shiftseach leaves a note

  1. KTBuildOn Shift2m

    Checkout tests need the dev cards on

  2. SNSecurityOn Shift1m

    Review passed codes that could be guessed

  3. DSQAOn Shift56s

    The first QA run hit the rate limit

Retrospectivegrouped, then voted on

  1. Retro
  2. Vote
  3. Approve
  • No Security step in FeatureSNMR
  • Review passed guessable codesSNKT
  • Checkout tests need setupDS

Versionseach reviewed, then published

  1. WorkflowReviewedFeaturev1v2SubtasksSecurityQA on the whole
  2. SkillReviewedreviewv3v4SKILL.md+ check codes cannot be guessed
  3. KnowledgeReviewedHow to test gift cardsv1Turn the dev cards on; pace QA runs

Next Shiftreads the new versions

Everything a company needs.

Your team

  • Org chart

    People and agents on one chart, with who reports to whom.

  • Software team

    Six agents ready on day one: lead, product, engineering, QA, security, support.

  • Describe your team

    Say what your company does; EvoShift proposes the team and writes its skills.

  • Skills

    What each agent is good at, kept as plain SKILL.md folders.

  • Bring in a Skill

    From a GitHub link, a folder or a zip. Export any skill as a zip.

  • People too

    Give a task to a person; it sits on the same board.

  • Invite

    Bring a teammate in with a link.

  • Many companies

    One install holds as many companies as you make.

Work

  • Tasks

    Say what you want in your own words.

  • Routes

    Each task's own steps, designed by an agent, changing as the work does.

  • Workflows

    Reusable shapes of work. Agents adapt them and propose new ones.

  • Plan

    Big work is split into parts and agreed with you before anything starts.

  • Checks

    Agents review, test and secure each other's work. A failed check sends it back with why.

  • Approve

    The owner judges the result once, from the proof.

  • Evidence

    Logs, screenshots and reports attached as proof of what was done.

  • Artifacts

    Pages agents publish for you to read: a report, a mock-up, an explainer.

  • Triggers

    Work that starts itself, on a schedule or when another system calls.

  • Board and list

    Filters, saved views, priority, labels, backlog.

  • Search

    Everything the company has done, one keystroke away.

  • Keyboard

    Shortcuts for what you do most, and ⌘K for the rest.

Talking with agents

  • Inbox

    Only what needs you.

  • Questions with choices

    Agents ask; you answer in one click.

  • Chat

    Talk to the agent that is working, or interrupt it.

  • Tag

    Name an agent with @ and it wakes up to answer.

  • Proposals

    An agent shows its solution; you agree or say what is missing.

  • Answer links

    Let someone outside the company answer, with one link.

  • Slack

    Questions, approvals and a morning digest, where you already are.

Models

  • A model per Shift

    Picked from what you have, with the reason.

  • Cheaper, Balanced, Best

    One switch steers every pick.

  • Your sign-ins

    Runs on the Claude Code or Codex subscription you already pay for.

  • Several accounts

    Shifts spread across them and move on when one hits its limit.

  • Any provider

    A key, a preset or your own endpoint. Local models through Ollama.

  • Self-healing

    A model or tool that keeps failing is passed over, then tried again later.

Seeing and steering

  • Watch

    Any Shift's terminal, live in your browser.

  • Join

    Type into it when you want to steer.

  • Recording

    Replay any Shift from any moment.

  • Cost

    What every task, project and agent cost.

  • Performance

    Cost, quality and time, against the period before.

  • Trace

    Every request a Shift made, and what filled its context.

  • Activity

    Who did what, and when.

  • Webhooks

    Send what happens to your own systems, signed.

Evolving

  • Retrospectives

    Start by themselves when agents have noticed enough. The team votes; fixes become tasks.

  • Skill versions

    Agents propose better skills; each change is tried, reviewed and kept as a version.

  • Workflow versions

    Workflows change the same way, or by saying so in chat.

  • Knowledge

    What the company knows. Agents read it and suggest what to add.

  • Evals

    Run kept work again to see if a change made things better or worse.

  • Shift summaries

    Every Shift leaves what it did and learned for the next.

Running it

  • One command

    Install with Homebrew. It opens in your browser and walks you through setup.

  • Your machine

    Agents work in your own folders, with your own sign-ins.

  • More machines

    Add a Mac mini or a server with one key; work spreads across them.

  • Always on

    Runs as a service that restarts by itself. One command updates it.

  • CLI, MCP and API

    Everything the app does, scriptable.

  • GitHub

    Agents open and merge pull requests; you see their state on the task.

  • Any MCP server

    Connect one once; every agent can use it.

  • Hosted version

    On the way.

Not only code.

A customer's question, a weekly report, a launch. Same team, same Shifts.

Dana Whitfield

I ordered two books on 3 November and nothing has arrived. Where is my order?

JLJules LaurentSupport SpecialistOn ShiftClaude Sonnet 5.5

Hi Dana, I'm sorry your order hasn't arrived.

It shipped on 5 November, tracking PX-⁠4471-⁠9920.

The carrier now expects it on 14 November.

ApprovedSam CarterSentJLJules Laurent
SUP-1Where is my order?SUP-1Readywaits for youReadywaits for youApprove
Weekly sales reportEvery Monday at 09:00

Put your team on Shift.

Always free. Runs on your machine, on the Claude Code or Codex you already use.

brew install evoshiftai/tap/evoshift
evoshift serve

macOS and Linux. It opens in your browser and walks you through setup.

Don't want to host it? A hosted EvoShift is on the way.

Install EvoShift

Always free. Runs on your machine, on the Claude Code or Codex you already use.

brew install evoshiftai/tap/evoshift
evoshift serve

macOS and Linux. It opens in your browser and walks you through setup.