Axis Workflow Axis Workflow

Your project.  Your computer.

AI is replaceable. Your project is not.

Models can come and go,
but your work needs to stay.

The Axis Workflow™ is a free, open-source workflow for your AI agent to follow.  It keeps your plan, tasks, decisions and knowledge base on your computer, in plain markdown files, so any AI can pick up where the last one stopped.

GitHub

Your project folder always stays with you.
AI models read from your folder and follow the Axis rules.

What is Axis?

Keep your AI. Add a workflow.

Axis adds what’s missing in your AI platform: a cross-platform method to run your project across multiple sessions, different models, and remote computers.

  • Not a model.Axis brings no AI of its own.  Your AI does the thinking.
  • Not an app.No account, no install, no server.  One folder of plain files.
  • A workflow.Instructions, records and checks that every agent follows.

Where Axis fits

  • YouThe direction
  • Your AIThe labor
  • AxisThe discipline
  • Your folderThe work

How does Axis work?

Axis manages a project, not a chat.

Axis records everything – but Agents do it in a smart way that saves tokens.

Axel, the Axis agent, holding a record card

  • A plan, with tasks

    Instead of a thread too long to scroll

    Your agent drafts a plan, organizes initiatives, and tracks specific tasks along with their status.  Decisions for you are surfaced and tracked separately, along with reminders and action items.

  • Decisions, with reasons

    Instead of settled questions coming back

    Every decision is logged in a permanent record, along with the logic and reason it was made.  If you change your mind, the new decision is added beside the old one; it’s never overwritten.

  • Memory, with tracking

    Instead of re-pasting the brief

    Type Start Axis and your agent reads the plan, the tasks and where you left off before it answers.  It picks up where you stopped, on Monday or on another computer.

  • A review, without the hype

    Instead of answers nobody checked

    For work that matters, a separate AI session runs a cross-examination to challenge the assumptions, logic and evidence.  Findings go on the record, and you decide what to do next.

All records live in a folder on your computer, as plain files that you can read.

Why is Axis better?

Axis works smarter, not more.

You can always throw more compute at a problem – but that’s expensive.

Built in the open.

Axis is an open-source project.  Every file is available on GitHub for you to inspect, and contributors are welcome!  If you find Axis useful, please give it a star on GitHub – that helps others to find the project.

View on GitHub

Better models. More possibilities.

Don’t start over every few months. With Axis, your workflow will automatically route your work to the best alternative: harder tasks to frontier models, and routine tasks to cheaper and/or free local models.

Keep up.
AI will improve.

We are still at the start of the AI revolution. Axis has the flexibility to adapt to new models and new features as they arrive.

Setup takes a minute.
There is no account, and nothing to install.

Take a tour

Want a closer look?

More questions?

TourA chat is not a project.

A chat is not a project.

Projects run for weeks.  They need a plan, a record, a memory and a second opinion.

A new courseIllustrative chat
  1. You: Where did we put the module plan?  I’ve been scrolling up for ages.
  2. You: Wait, didn’t we settle the reading list last week?
  3. AI: Maybe.  I can’t see what we decided, or why.
  4. New chat
  5. You: Here’s my brief again:
  6. You: Is the new outline right?  Has anyone checked it?
  7. AI: It looks right to me.
Sound familiar?  Four signs you’ve outgrown the chat.  Each one has a fix below.

Sign one

The plan is somewhere up the thread.

A project keeps a plan and tasks instead.

Your agent writes the plan down and keeps every task with its status and a definition of done.  Decisions that only you can make wait on one list of follow-ups, so nothing important hides in the scroll.

In the chat

Where did we put the module plan?  I’ve been scrolling up for ages.

In the project

Sign two

Settled questions keep coming back.

A project logs each decision, with its reason.

When a choice is made, the reason goes on file with it.  Logs are written once: if you change your mind, the new decision is logged beside the old one, not written over it, so nobody reopens a settled question by accident.

In the chat

Wait, didn’t we settle the reading list last week?

Maybe.  I can’t see what we decided, or why.

In the project

Sign three

Every new chat needs the brief again.

A project remembers between sessions.

Type Start Axis and your agent reads the plan, the tasks and the last snapshot before it answers.  It picks up where you stopped: on Monday, in another AI tool, or on another computer.

In the chat

New chat

Here’s my brief again:

In the project

Sign four

Nobody checked the answer.

A project gets a second opinion.

AI tends to oversell its first answer.  For important work, a separate agent with a fresh context cross-examines the assumptions, the logic and the evidence.  Its findings go on the record, flattering or not, and you decide what changes.

In the chat

Is the new outline right?  Has anyone checked it?

It looks right to me.

In the project

Who it’s for

For knowledge workers with serious projects.

Work that runs for weeks, where AI does the legwork and you make the calls.

All four fixes live in one folder on your computer, as plain files you can read, and they work with the AI tool you already use.

Ready when you are

Make it a project.

Sixty seconds to start.  Free and open source, under the MIT License.  No account, app, server or install.

Download and unzip.  Open the folder with your AI tool.  Type Start Axis.

Download Axis

TourKeep your AI.  Add a workflow.

Keep your AI. Add a workflow.

The AI tool you know, plus a written way to run a project across sessions, tools and computers.

  1. Not a model.
  2. Not an app.
  3. A workflow.

Bring your AI

Not a model.

Your AI does the thinking.

Keep using the tool you know.  Axis brings no AI of its own, so it isn’t tied to one vendor or one model.  At every start, it checks what your tool and model can do, and adapts.

Keep your files

Not an app.

One folder.  On your computer.

Plain instructions, small scripts and the project record, beside your own work.  Nothing runs in the background and nothing phones home.  Open any file in a text editor.

  • No account
  • No install
  • No server
  • No telemetry

Add the workflow

A workflow.

Pick up where you left off.

Research, a course, a business plan: real projects take more than one chat.  Axis gives every agent the same written practices, so each session works the same way and picks up the thread.

Discipline lives in the folder, not in anyone’s memory.

What your agent does, every session

  1. Plan the work.

    A direction, priorities and a next step.

  2. Track the tasks.

    What’s in progress.  What’s waiting on you.

  3. Keep the record.

    Decisions and their reasons, on file.

  4. Resume with context.

    The record is read at the next start.

  5. Cross-examine important work.

    A separate agent challenges the assumptions.

Why it matters

Dependable results need a dependable workflow.

You want your agents following a consistent workflow, not improvising.

When every session improvises, different parts of the project end up working at cross-purposes.  With Axis, every agent follows the same practices, so the pieces fit together and the results hold up.

Where it fits

You, your AI, the workflow, your folder.

You set the direction and make the calls.  Axis keeps every agent working the same way.

  1. DirectionYouSet the goals, review the work, make the calls.
  2. ThinkingYour AI toolClaude, ChatGPT, Codex, Gemini, Cursor or a local model.
  3. WorkflowAxisThe practices it follows: plan, track, record, resume, cross-examine.
  4. RecordYour folderPlan, tasks, logs, snapshots and wiki, on your computer.

Ready when you are

Same AI. A better way to work.

Free and open source, under the MIT License.  No account, app, server or install.  Just your AI tool and a folder you control.

Download and unzip.  Open the folder with your AI tool.  Type Start Axis.

Download Axis

TourStop briefing your AI.  Let it brief you.

Stop briefing your AI. Let it brief you.

The costly part of AI isn’t getting an answer.  It’s getting a project back on track.

The recap tax

Right now, you do the briefing.

A long thread buries the decisions that mattered.  A fresh chat forgets them entirely.

So every session starts with you scrolling back, pasting the brief again and re-explaining what you already settled.

Coming back to a project

  • One long thread

    Finds the last decision
    If you scroll
    Knows why it was made
    Somewhere above
    Knows what is open
    Partly

    To get goingThe whole thread, read again

  • A fresh chat

    Finds the last decision
    No
    Knows why it was made
    No
    Knows what is open
    No

    To get goingYour recap, typed again

  • With Axis

    Finds the last decision
    Yes, it is on file
    Knows why it was made
    Yes, the reason is on file too
    Knows what is open
    Yes, and who owns each step

    To get goingA slim read, then the briefing

Our assessment.

As you work

It goes on file.

Keeping the record is part of the work, not a chore for later.

Each time a choice is made, three things go into your project folder: the decision, the reason, and the next step with its owner, you or the agent.

When you return

It reads first.

Not the whole history.  Just what matters.

You type Start Axis.  Before it answers, your agent reads the plan, the latest snapshot and what is waiting on you.  Everything else waits behind a one-line index until the work needs it.

Before the first reply

  1. The plan and the tasksPLAN.md · TASKS.md
  2. The latest snapshotSnapshots/
  3. What is waiting on youFollowups/
  4. Everything else: an index first, the detail only when needed

Before any work

It briefs you.

Where things stand, why, and what needs you.

Orientation takes seconds, not a page of recap.  You get straight back to the work, on Monday, after a holiday, or in a different AI tool.

Why trust it

A record you can open, not a memory you can’t see.

The briefing is only as good as the record behind it.

With Axis, that record is plain files in your project folder, on your computer: readable by you, correctable by you, and readable by any AI tool you open the folder with.

In the briefingOn file
  • DecidedLogs/Each decision, written once
  • WhyLogs/The reason, beside the decision
  • Waiting on youFollowups/What only you can decide
  • Next for meTASKS.mdEvery task, with a status
  • Where things standSnapshots/The project at each save

Illustrative: one project folder, read at the start of a session.

Ready when you are

Start in sixty seconds.

Free and open source, under the MIT License.  Set the project up once; next time you type Start Axis, your agent reads the record first and picks up where you left off.

Download and unzip.  Open the folder with your AI tool.  Type Start Axis.

Download Axis

TourMake your AI show its work.

Make your AI show its work.

Getting an answer is easy.  Standing behind it means seeing how it was reached.

Axel, the Axis agent, holding up a record

Your agent’s answerIllustrative

Question: Is the course outline ready to share?

  1. The planOn file, with “done when” for every task
  2. The reasons6 decisions logged, 1 reversal, 1 mistake
  3. The challengeCross-examined: 3 objections, 2 accepted
  4. The limitsNot verified: one module 5 reading

Answer: Yes, with one reading marked as not verified.

The rule

An answer only counts if you show how you got it.

Every math teacher had that rule.  Axis applies it to the AI you already use.

The work is kept where you can check it: in plain files, in your own project folder.  Four things go on file, and the rest of this tour shows each one.

The plan

The plan is on file before the work starts.

Every task has a status, and a definition of done.

Your agent writes down the goal, the tasks that reach it, and what “done” means for each one.  You read it and correct it in plain words.  At any moment you can see what is finished, what is next, and what is waiting on you.

TASKS.md · a new courseIllustrative

Write the module 3 reading listActive

Done whenEvery reading has an edition, a page range and a reason for being on the list, and you have approved it.

The reasons

Every decision that matters is logged, with its reason.

So are the reversals, and the mistakes.

Each entry is written once.  When the plan changes, a new entry supersedes the old one and both remain, so you, a colleague or your agent’s next session can trace why the work went the way it did.

  1. Decision

    Teach module 3 before module 2.

    Reason: students need the vocabulary first.

  2. Reversal

    Move the case study to week 5.

    Reason: the guest speaker can only come that week.

    Supersedes 2 Oct

  3. Mistake

    A reading cited the wrong edition.  Corrected.

    Lesson: check the edition before citing.

Illustrative · Written once.  Corrections are added, never written over.

The challenge

A second agent cross-examines the important work.

Its critique is filed exactly as written, flattering or not.

AI tends to oversell its first answer.  So a separate agent, started with a fresh context, challenges the assumptions, the logic and the evidence.  Then you decide what changes.

The limits

It says plainly what it could not verify.

A confident answer can hide a guess.

Your agent names what it checked and what it could not, so the gaps are in front of you before anyone relies on the work.  Type ^review for a dated report you can file, print or send.

Not verifiedIllustrative

  • Module 5 reading: the page range is not checked.  The library copy has not arrived.
  • Guest speaker’s date: confirmed by email only, not yet in the calendar.

Waiting on youApprove the outline.

Ready when you are

Start in sixty seconds.

Axis shows its work, too.

Every rule your agent follows is a plain text file in your project folder: read it, change it, share it.  Free and open source, under the MIT License, with no account and nothing to install.

Download and unzip.  Open the folder with your AI tool.  Type Start Axis.

Download Axis

TourSo what’s the catch?

So what’s the catch?

You need to pay for your own AI.

Axel, the Axis agent, pointing at the project folder

Is it really free?

Yes.  Free and open source, under the MIT License.

Use it at home, at work or for clients.  Change any rule, share it with anyone, build it into your own tools.  There’s nothing to buy and nothing to cancel, and because your copy is a folder of files, no one can reprice it, revoke it or switch it off.

Fine printAxis is free, but your AI tool may not be, and keeping records and running checks adds some usage to your plan.

Where’s my work?

All of your work goes into plain files, in a folder on your computer.

Plans, tasks, notes, logs and the wiki are markdown files you can open in any text editor.  Any AI tool can open the same folder, so you can switch tools without exporting anything.

Who holds your work
Your computer
Everything.  The folder is the project.
Your AI provider
What its agent reads as it works, under your provider’s terms.
The makers of Axis
Nothing.  There’s no Axis server and no telemetry: nothing phones home, because there’s nowhere to call.

Fine printYour computer holds the only copy until you make another.  Back it up however you like, or version it with git.

What do I get?

Your own AI agent, following a workflow: it plans, keeps records, remembers between sessions and cross-examines what matters.

Most AI work evaporates when the chat ends.  With Axis, your agent writes down the plan, the decisions and the reasons as it works, and reads them back at the start of every session.

Fine printEach session starts with a short, fixed read of the plan and records before the agent replies.  It’s a small, predictable cost, and the reason you don’t have to re-explain your project.

Is it right for me?

Yes, if you have a serious project.

Axis is built for work that runs for weeks, where AI does the legwork and you make the calls.  Where another tool will serve you better, we’d rather say so now.

Axis is a good fit for

  • A thesis or a research project
  • A court case and its filings
  • A new course
  • A business plan or a new product
  • Taxes, mail and family admin

Better served elsewhere

Building software
Better to use Cursor or Visual Studio
Sharing knowledge across a company
Better to use Confluence or Notion
Asking a quick question
Better to use a typical chat session

Fine printBefore you commit a large project, skim the limitations in the documentation.  It comes in the download, and it’s on GitHub.  We want you to find the right fit.

How do I try it?

Sixty seconds to start.  Ten seconds to stop.

  1. Download and unzip.Now you have a project folder.  Rename it for your project.
  2. Open the folder with your AI.Claude Cowork or Claude Code, ChatGPT Work or Codex, Cursor, or Gemini CLI.
  3. Type Start Axis.Your agent finds the workflow and helps you set up the project.
  4. Changed your mind?Just stop.  Nothing was installed anywhere else and there’s no account to close.  Keep your files; remove the workflow files if you like.

Fine printIf you are on Windows, use WSL or Git Bash to get every feature.

Ready when you are

So what’s the catch?

You bring your own AI.  That’s it.

Free and open source, under the MIT License.  No account, app, server or install.  Read every file on GitHub before you start, if you like.

Download and unzip.  Open the folder with your AI tool.  Type Start Axis.

Download Axis

TourIt starts with two words.

It starts with two words.

Start Axis

First, what it is

The Axis Workflow is a free, open-source workflow that your AI agent follows.  It lives in a folder of plain files on your computer and keeps your plan, tasks, decisions and records ready for whichever AI you bring.

Open that folder with your AI tool and type the two words.  Your agent finds the workflow and gets to work.

Download Axis

Free · MIT License · Nothing to install

Axel, the Axis agent, pointing at the Axis files in a project folder
your-project/
Your workThe things you make
_Axis/The workflow and the recordPlan · Tasks · Logs · Snapshots
Illustrative project folder

Bring the AI you already use.  Claude, ChatGPT, Codex, Gemini, Cursor, local models.

The first time

Start Axis

A name.  A little background.  A few goals.

That’s enough to begin.  Your agent helps you set up the project and turns your goals into a plan.  Everything else can change as you go.

Every time after

Start Axis

Pick up where you left off.

Before it answers, your agent reads the plan and the last snapshot, then checks what’s waiting on you.  You get back to work without rebuilding the story.

While you work

Start Axis

Then just talk to your agent.

The work moves. The record keeps up.

Your agent plans the next steps, tracks tasks and logs decisions with their reasons.  Important work can be cross-examined by a separate agent.  You set the direction and make the calls.

Anywhere else

Start Axis

Different tool. Same two words.

Type ^save, then open the same folder with another AI tool, or copy it to another computer.  The next agent reads the same plan, tasks and records, checks what it can do, and carries on.

Illustrative: three AI tools, one project folder.

Ready when you are

One download. Two words.

Free and open source, under the MIT License.  No account, app, server or install.  Your work stays yours.

Download and unzip.  Open the folder with your AI tool.  Type Start Axis.

Download Axis

Deep diveHow Axis reads smarter, not more.

Fewer tokens. Every answer still correct.

With two-pass loading, Axis reads the project’s index first, picks the records it needs, and opens only those, instead of reading broadly.

In our long-session benchmark on Claude, that used about 38% fewer tokens, and every one of 240 hard questions was still answered correctly.

  • 38%fewer tokens on ClaudeOur 45-turn, 5,000-record benchmark
  • 100%of answers still correct240 hard questions
  • Claude and Codexboth testedWe recommend it only where it helps

How it works

Read the index first. Then only what the job needs.

Axis keeps every project as short index entries plus detailed records.

Two-pass loading makes the AI use that structure deliberately, on every request.  It reads less, so it spends fewer tokens, and the material it does read is the material that matters.

Standard reading

The AI reads broadly to find what the request needs.

Two-pass loading

The AI reads the index, then opens only the records the request needs.

  1. Pass oneLook only at the index: task titles, and the first line of each log and note.  List the records the request actually needs.
  2. Pass twoOpen exactly those records, and do the work from them.

Illustration of the method.  The measurements follow.

The results

Fewer tokens in every matched Claude test.

Two-pass loading used fewer tokens than standard reading in all four of our matched Claude tests, saving 21% to 56%.

Each pair used the same test project and the same questions, so the method was the only difference.  Tokens counted are all the context the model processed, including cached context.

Tokens per session, in millionsClaude Sonnet · shorter is better
  • Standard reading
  • Two-pass loading

The long sessions (Phase B)

Two test projects of about 5,000 records, 45 turns per session.

Phase B in total: standard reading 80.0 million tokens, two-pass loading 49.3 million.

  • 38%fewer tokens
  • ~11%faster to finish
The four matched comparisons
Test (Claude Sonnet)Project sizeTurns per sessionStandard reading (M tokens)Two-pass (M tokens)Token savingAccuracy, both methods
Phase B, test project 2~5,000 records4533.721.337%20/20
Phase B, test project 1~5,000 records4546.428.040%20/20
Phase A, test project 2~1,500 records159.54.256%10/10
Phase A, test project 1~1,500 records1510.78.521%10/10

No trade-off

Less reading. No loss of accuracy.

Accuracy stayed at 100% across 240 hard questions, including outdated decisions, look-alike names and facts buried deep in long records.

Both methods answered every question correctly, so the gain is efficiency, not intelligence.  Two-pass loading does not make Claude smarter or more accurate: it reaches the same answers while reading less.

  • Outdated decisions

    Questions about decisions that were later replaced, where only the current one counts.

  • Look-alike names

    Questions about records with similar names, where the AI has to open the right one.

  • Facts buried deep

    Questions whose answer sits deep inside a long record.

Claude and Codex

We recommend it only where it helps.

We test every optimization on more than one AI platform.  Two-pass loading helps Claude; it did not help Codex, so Axis does not recommend it there.

When a project runs on Claude, Axis suggests turning two-pass loading on.  On Codex, it does not.

On Claude

38% fewer tokens

and about 11% faster, in the long Phase B sessions

Axis will suggest it

On Codex

3% more tokens

and about 19–20% slower, overall

Axis will not suggest it

The fine print

What the numbers do and don’t say.

These limits travel with every number in this deep dive.

Two-pass loading is one of the context-management methods Axis is testing, toward a larger goal: managing what the AI reads as a project grows, instead of letting its context pile up until the host compresses it.

What we measured

  • Axis’s own benchmark: synthetic project records, one Claude model (Sonnet), four matched comparisons.
  • Savings varied from 21% to 56% by project.  38% is the result on the larger, longer sessions, not a guarantee.
  • “Tokens” means the context the model processed.  It is not a dollar figure: on a subscription, it shows up as using less of your plan’s allowance, not as a lower bill.
  • The benefit is specific to Claude.  It did not help Codex.

What we don’t claim

  • That Axis makes Claude more accurate or smarter.  Every method scored 100%; the gain is efficiency only.
  • Savings on your bill, in dollars or percent.  It is not “38% cheaper”.
  • That it helps every AI model or host.
  • That it solves context loss in very long conversations.  Our sessions rarely reached the point where hosts compress their context, so that is still being tested.
  • That it is proven or guaranteed, or what it will do on your project.  Nothing has been measured on customer projects yet.

Axis benchmark, October 2026: synthetic projects, Claude Sonnet, four matched comparisons (savings 21–56%).  Your results will vary.  Two-pass loading is an optional setting, available from Axis 2.04.

Ready when you are

Start today. Two-pass loading is now out in Axis 2.04.

It is an option of the new Context Management setting: when your project runs on Claude, Axis suggests turning it on, and on Codex it does not.  Free and open source, under the MIT License.

Download and unzip.  Open the folder with your AI tool.  Type Start Axis.

Download Axis

Download and try Axis

Quick Start

Free and open source, under the MIT License.  No account, app, server or install.  Just your AI tool and a folder you control.

Download Axis GitHub

  1. Download and unzip.

  2. Open the folder with your AI.

  3. Type Start Axis

  • Your files always stay on your computer.

  • It’s yours forever.  Nobody can switch it off.

  • It’s easy to leave.
    Nothing to cancel.