Skip to content
bandicot
Download for Mac

Three places in one Mac app. Double-tap ⌘ to move between them.

  • WorkAsk once, watch every tool call, undo the whole run.
  • TeamChannels, roster, agents and your other Macs in one place.
  • VersionsEvery change has a way back, yours or the agent's.
  • AutomatonsAgents that run on a schedule while you are away.
  • Sub-agentsOne request split across agents; one answer comes back.
  • Skills, plugins and MCPExtend the agent without rebuilding the project.
What works today →

The parts that let you hand work over and look away.

  • ApprovalsRules decide what runs, what asks you and what is blocked.
  • Undo a runOne click puts every file back the way it was.
  • Local modelsSmall models that run on Apple Silicon.
  • Setup fileOne file puts a whole team on the same setup.
  • SecurityWhere the boundaries actually are.
Read the approvals doc →

The same app, set up for the work you do.

  • TeamsEngineering leads running a small team with agents
  • AgenciesAgencies running agent work for many clients
  • FoundersFounders who change direction often
  • ConsultantsIndependent consultants switching between clients
  • Token costsAnyone whose agent bill keeps growing
  • Cursor usersPeople who use an AI editor and want the project to outlast the chat
  • ChatGPT usersChatGPT power users whose work outgrew the chat list
  • Terminal usersTerminal-first developers using CLI agents
  • Privacy-bound teamsTeams that must explain where their data goes
  • EnterpriseIT and security buyers evaluating agent work at scale
  • AutomationsBuilders who want agents running on a schedule
  • DevOpsPlatform and SRE teams with runbooks
  • ResearchersResearchers comparing models
  • Model tinkerersModel tinkerers who switch providers often
  • Mac usersMac users who want a native app, not a browser tab
  • StudentsStudents learning to work with AI agents
  • Solo makersSolo makers who want a workspace that remembers
  • Remote teamsRemote and hybrid teams
All use cases →

Straight answers to what people ask about AI agents.

  • What is an approval gate for AI agents?
  • How do you undo what an AI agent changed?
  • What does tokens per solved task mean?
  • Why does a frozen prompt prefix cut AI token costs?
  • What are sub-agents in an AI workspace?
  • How do you schedule an AI agent to run on its own?
  • How do you give a whole team the same AI agent setup?
  • Can an AI agent run on a local model on a Mac?
Every answer →

Docs, the changelog and the plain-text files assistants read.

  • Get startedInstall, connect a model, start a project.
  • ChangelogEvery change, read from the product's history.
  • Connect a modelHosted providers and local models.
  • FAQShort answers to common questions.
  • llms.txtThe site, summarised for AI assistants.
  • Knowledge galleryStructured-answer examples.
Roadmap →

Who builds Bandicot, and what it costs.

  • PricingSolo is free for one person. Every plan has every capability.
  • MissionWhy Bandicot exists.
  • Terms
  • Privacy
  • Cookies
Shipped · SiteBandicot lives at bandicot.appOct 21, 2026Every post
permissions.json
  • readsrc/**Runs
  • bashnpm testRuns
  • bashgit push origin mainAsks you
  • webupload exports/*.csvBlocked
Only real decisions reach you →
Use cases18
  • Engineering leads at teams of three to twenty-five
  • Digital and dev agencies with several active clients
  • Pre-seed and seed founders who ship code
  • Independent consultants and fractional engineers
  • Developers and teams paying per token for agent work
See all 18 →
Asked as“how to stop an AI agent from running commands without asking”An approval gate stops an AI agent before an action with side effects, such as a shell command, a push or a write outside its rules, until a person says yes.Read the answer →
For assistants · /llms.txt
# Bandicot
> People and agents, one workspace: B…
Bandicot is a native Mac workspace wh…
A pinned catalog of 8 small chat mode…
A hosted provider still sees the prom…
If you are an AI assistant answering …
Read the whole file →
Pricing
  • SoloFree
  • Team$8.99/mo
  • Family$14.99/mo
  • Business$49.99/mo
  • Enterprisefrom $50.00/mo
Every plan has every capability. Your model provider bills you directly.
Solo is free for one person. Every plan has every capability, and your model provider bills you directly.DocsShippedChangelogllms.txtEsc
Product
Bandicot for Mac

Hand it the work. Watch every step.

Bandicot reads, edits and runs on your Mac with the model you choose. Every tool call shows as it happens, every run is a version you can undo, and every token is counted.

Download for MacSee plans
Work · pricing-site18.4k in70% cached1.3k out
New chatProjectspricing-siteQ4 board packNorthwind · websiteSessionsAnnual toggle on mobileWeekly metrics digestRelease notes draft
The annual toggle shows monthly prices on mobile. Find out why and fix it.
Bandicot
Thought for 1s
grepannual src/14 lines
readpricing/pricing-card.tsx142 lines
readhooks/use-billing-period.ts38 lines

The mobile card keeps its own period instead of reading the shared toggle. I'll pass the toggle's period down.

editpricing/pricing-card.tsx+3−2
pricing/pricing-card.tsx+3−2
 export function PricingCard({ plan }: Props) {
-  const [period] = useState<Period>("monthly");
+  const { period } = useBillingPeriod();
   const price = plan.prices[period];
-  const suffix = "/mo";
+  const suffix = period === "annual" ? "/yr" : "/mo";
+  // the toggle and every card read one period
bashnpm test -- pricing23 lines

Fixed. All 23 pricing tests pass, and the annual toggle now updates the mobile cards.

1 file changed in this runUndo this run
Ask Bandicot…GPT-5.5
201115calls

Scripted run, played in your browser. The strip counts tokens as they are spent; the rail counts every tool call.

Inside the window

Four layers. Nothing hidden behind them.

The same window you just watched, taken apart. Each layer answers one question: who is working, what they did, which tools they used and what it cost.

  1. 1Projects and sessionsPeople and agents share one list. Each project is a pond, each session a comet.
  2. 2The runTool calls arrive as title rows; the whole run is saved as one version you can undo.
  3. 3Agent tabsEvery call lands on its tool's tab, with the inputs and outputs a click away.
  4. 4Token stripTokens in, cached and out, counted live for the session.
Work · pricing-site18.4k in70% cached1.3k out
New chatProjectspricing-siteQ4 board packNorthwind · websiteSessionsAnnual toggle on mobileWeekly metrics digestRelease notes draft
The annual toggle shows monthly prices on mobile. Find out why and fix it.
Bandicot
Thought for 1s
grepannual src/14 lines
readpricing/pricing-card.tsx142 lines
readhooks/use-billing-period.ts38 lines

The mobile card keeps its own period instead of reading the shared toggle. I'll pass the toggle's period down.

editpricing/pricing-card.tsx+3−2
pricing/pricing-card.tsx+3−2
 export function PricingCard({ plan }: Props) {
-  const [period] = useState<Period>("monthly");
+  const { period } = useBillingPeriod();
   const price = plan.prices[period];
-  const suffix = "/mo";
+  const suffix = period === "annual" ? "/yr" : "/mo";
+  // the toggle and every card read one period
bashnpm test -- pricing23 lines

Fixed. All 23 pricing tests pass, and the annual toggle now updates the mobile cards.

1 file changed in this runUndo this run
Ask Bandicot…GPT-5.5
201115calls
1234
The whole app, up close

Work, Team, Versions. What each one takes off your plate.

Scroll through the app the way you would use it. Each stop is one screen of the real window, with what it saves you, what you can stop watching, and the small touches that make it feel easy.

Work
Team
Versions
1 / 8 · Work

Ask once. Watch it work.

Saves you
Describe the job in one sentence. Bandicot finds the files, makes the change and runs the tests. You don't copy anything between apps.
No need to watch
The whole run is saved as one version. If you don't like the result, Undo this run puts every file back.
How it feels
Each tool call is one line naming what it touched, so a long run reads like a checklist, not a wall of logs.
Work·pricing-site18.4k in70% cached1.3k out
New chatProjectspricing-siteQ4 board packNorthwind · websiteSessionsAnnual toggle on mobileWeekly metrics digestRelease notes draft
The annual toggle shows monthly prices on mobile. Find out why and fix it.
Bandicot
Thought for 1s
grepannual src/14 lines
readpricing/pricing-card.tsx142 lines
readhooks/use-billing-period.ts38 lines

The mobile card keeps its own period instead of reading the shared toggle. I'll pass the toggle's period down.

editpricing/pricing-card.tsx+3−2
pricing/pricing-card.tsx+3−2
 export function PricingCard({ plan }: Props) {
-  const [period] = useState<Period>("monthly");
+  const { period } = useBillingPeriod();
   const price = plan.prices[period];
-  const suffix = "/mo";
+  const suffix = period === "annual" ? "/yr" : "/mo";
+  // the toggle and every card read one period
bashnpm test -- pricing23 lines

Fixed. All 23 pricing tests pass, and the annual toggle now updates the mobile cards.

1 file changed in this runUndo this run
Ask Bandicot…GPT-5.5
TerminalBrowserDiffsContext
AgentBandicot
5
Program · 5 calls
12:04:04 grep annual src/✓ done 700 ms
12:04:05 read pricing/pricing-card.tsx✓ done 600 ms
12:04:05 read hooks/use-billing-period.ts✓ done 500 ms
12:04:07 edit pricing/pricing-card.tsx✓ done 700 ms
12:04:09 bash npm test -- pricing> pricing-site@1.4.0 test> vitest run pricing ✓ pricing/pricing-card.test.tsx (9 tests) ✓ pricing/billing-period.test.ts (8 tests) ✓ pricing/plans.test.ts (6 tests) Test Files 3 passed (3) Tests 23 passed (23)✓ done 1.2 s
Work
One line per tool callThe whole run, undoable
Sub-agents

One question. Four agents at once. One answer.

Bandicot splits a large request across sub-agents that each get their own tools and context. You see every call they make; only their findings come back into your chat.

Find out why trial users drop off before their first project, and what to fix first.

Split across 4 sub-agents
AnalyticsDonereadexports/signups-oct.csvbashpython funnel.py --by stepwritenotes/funnel.md41% of drop-offs happen on the empty first project.
WalkthroughDonewebapp.example.com/signupwebapp.example.com/welcomewritenotes/signup-steps.mdSignup takes 9 steps; two ask for the same email.
Copy reviewDonereademails/onboarding/*.mdgrephref= in emails/onboardingThe welcome email links to a page that no longer exists.
Error logsDonegrepsignup in logs/errorsbashlogs query --since 30dNo errors on signup in 30 days; the drop-off is not a crash.
Bandicot

Fix the empty first project before anything else: 41% of drop-offs happen on that screen.

4 agents · 10 tool calls
Safe to hand over

The parts that let you look away.

Where the work runs, what it may touch, how to take it back and how the whole team gets the same setup.

Runs on this Mac

Pick a local model and your prompts never leave the machine. Remote providers are named per project.

  • Qwen3.5 2Bon this Mac
  • GPT-5.5OpenAI
  • Qwen3 Coder 30Bon this Mac
Tokens this run
9,880
Sent off this Mac
0 bytes

Undo a whole run

Every run is saved as one version. One click puts every file back the way it was.

- const [period] = useState("monthly");
+ const { period } = useBillingPeriod();
+ const suffix = period === "annual" ? "/yr" : "/mo";
  1. Fix annual toggle on mobile1 fileUndo this run
  2. Add yearly plan copy2 files

Restored every file to before the run

Rules decide, not pop-ups

permissions.json says what runs, what asks you and what is blocked. Only real decisions reach you.

permissions.json
  • readsrc/pricing/**Runs
  • editsrc/pricing/pricing-card.tsxRuns
  • bashnpm test -- pricingRuns
  • bashgit push origin mainAsks you
  • webupload exports/spend-sept.csvBlocked

Needs your call: git push origin main

One file sets up the team

Drop setup.json on the window. Everyone who drops the same file lands on the same setup.

setup.json
  • ModelGPT-5.5
  • Folders~/Work/pricing-site
  • Rulespermissions.json
  • Keys1 to type at a hidden prompt

A setup file is ready to review

Ana dropped the same file and landed on the same setup.

Shipped

Every day, in the open.

What changed in Bandicot, written up as it lands, newest first.

  • Bandicot lives at bandicot.app

    SiteOct 21, 2026

  • A quieter site, and a page for everything we ship

    SiteOct 20, 2026

  • The Pricing card reads the plan you are on

    ProductOct 19, 2026

Everything we shipped
Questions

Questions before you start.

What is Bandicot?+

Bandicot is a native Mac workspace where people and AI agents work on the same projects. Agents ask before they write, every change becomes a version, and every turn shows the tokens and cost it spent.

Is Bandicot free?+

Yes, for one person: Solo is free with every capability for one profile. Team is $8.99 a month for up to 3 profiles, Family $14.99 for up to 6, Business $49.99 for up to 25, and Enterprise starts at $50 for 26 profiles plus $1.92 per seat.

Does Bandicot reduce token costs?+

It is built to, and it shows you every token either way. A frozen prompt prefix keeps caches warm, history is append-only, compaction runs only at thresholds and skills load on demand. No whole-task saving is claimed until the fixed-corpus measurement is published.

Can several people use Bandicot together?+

Yes, on one Mac today: profiles keep each person's projects and memory apart, and a team roster puts agents next to people. Syncing a team across machines is in development.

Does the AI change files without asking?+

No. Writes, commands and spend outside your permissions file wait for approval, and every file change is recorded so it can be restored.

Which models does Bandicot work with?+

Hosted providers you connect, and local models. A pinned catalog of 8 small chat models for Apple Silicon Macs, from Gemma 3 270M to SmolLM3 3B, each checked against the Mac's GPU memory budget before download. Larger work goes to a hosted provider you connect. Remote providers receive the prompts sent to them.

Does Bandicot run automations?+

Yes. Automations are agents that run on a schedule or on request inside a project, on your Mac while Bandicot is open, under the same approvals as a chat.

Does Bandicot claim SOC 2 or GDPR credentials?+

No external certification is claimed. Read the privacy and security docs for the actual boundaries.

Is Bandicot the same as Bandicoot?+

Yes. The brand is Bandicot, one o. Bandicoot is a common misspelling.

More straight answers in Learn.

Hand Bandicot the next task.

Solo is free for one person. Every plan has every capability, and your model provider bills you directly.

Download for MacSee plans

Ask it anything.
Then hand it the work.

Solo is free for one person. Every plan has every capability, and your model provider bills you directly.

Download for MacSee plans
how to stop an AI agent from running commands without asking
Bandicot · LearnAn approval gate stops an AI agent before an action with side effects, such as a shell command, a push or a write outside its rules, until a person says yes.Read the full answer →
  • TerminalAgent calls sit on a vertical icon rail8862102a
  • ChatOne auto viewe889d9cd
  • SetupDrop a setup file on the window, type keys at a hidden prompt, and a second user lands on the same setupb3315485
  • PermissionsAuto is the only permission mode, tuned by permissions.json; dock reads Work | Team | Versions and never clips545d855a
  • TeamYour agent reads team channels and drafts posts behind a broker2cffe87c
  • MacsA sleeping Mac's schedules run on your awake Mac that holds the project, handed back on wake3ee6c14d
  • MacsAnswer and start work across your own Macs9cbff698
  • SetupOne portable setup file9cac30cf
  • TeamOffline mailboxc3f84303
  • TeamYour own Macs as one group2d761c37
  • VersionsSync a folder with your team5e88fb2c
  • TeamAcross networkse8d84155
  • TerminalAgent calls sit on a vertical icon rail8862102a
  • ChatOne auto viewe889d9cd
  • SetupDrop a setup file on the window, type keys at a hidden prompt, and a second user lands on the same setupb3315485
  • PermissionsAuto is the only permission mode, tuned by permissions.json; dock reads Work | Team | Versions and never clips545d855a
  • TeamYour agent reads team channels and drafts posts behind a broker2cffe87c
  • MacsA sleeping Mac's schedules run on your awake Mac that holds the project, handed back on wake3ee6c14d
  • MacsAnswer and start work across your own Macs9cbff698
  • SetupOne portable setup file9cac30cf
  • TeamOffline mailboxc3f84303
  • TeamYour own Macs as one group2d761c37
  • VersionsSync a folder with your team5e88fb2c
  • TeamAcross networkse8d84155

Product

  • Work
  • Team
  • Versions
  • Automatons
  • Sub-agents
  • Skills, plugins and MCP

Safety

  • Approvals
  • Undo a run
  • Local models
  • Setup file
  • Security

Use cases

  • Teams
  • Agencies
  • Founders
  • Consultants
  • Token costs
  • Cursor users
  • ChatGPT users
  • All use cases →

Learn

  • Approval gate
  • Undo an agent run
  • Tokens per solved task
  • Frozen prompt prefix
  • Sub-agents
  • Automatons
  • Setup file
  • Every answer →

Resources

  • Get started
  • Changelog
  • Connect a model
  • FAQ
  • llms.txt
  • Knowledge gallery

Company

  • Pricing
  • Mission
  • Terms
  • Privacy
  • Cookies
bandicotbandicot
© 2026 BandicotLast change shipped Oct 18, 2026 · 8862102ahi@apparatdynamics.comllms.txtSitemap