A memory for your agent’s browser work

Stop paying your agent to figure out the same website twice.

Agentic Routines remembers how your agent did a task, then replays it next time without asking the model again. Repeat chores run up to 3.5 times faster and cost up to 83% less.

Install, it’s free See the numbers

Works with

Claude Code
Codex
GitHub Copilot
OpenCode
and more soon

The same chore, run 30 times

By handWith Agentic Routines

tokens: −82%
time: −66%
cost: −81%

Cost of one YouTube description edit, 21 cents by hand and 4 cents replayed, times 30. Run 1 is your agent at full price. Tokens, time and cost for one run are medians of 5 runs with Claude Sonnet 5.5.

Made for the chores your agent repeats.

The first run is your agent, as usual. Every run after that replays what it learned: no reading the page again, no thinking, no waiting.

Once

A task you do one time.

Just ask your agent. There is nothing to save, and nothing lost.

Again and again

A chore you repeat.

Shipping a release, exporting a report, adding a record in a console. It saves on every run after the first.

Everyone

A chore your whole team does.

Record it once and share it, so nobody pays to teach it again.

Watch it once. Replay it every time.

Your agent does the task. Agentic Routines tidies the steps up and keeps them, then checks the page at every step on the way back.

1 · Record

Your agent runs the task.

navigate  play.google.com/console
click     Kavyam
click     Test and release
click     Testing
click     Internal testing
click     Create new release
… 19 more

25 steps, as your agent clicked them.

2 · Compile

The steps are tidied up.

navigate  …/releases/new  direct URL
upload    {{bundle}}
type      Release name  {{version_name}}
type      Release notes {{release_notes}}
click     Next
… 15 more

Navigation a link can replace is dropped. What changes between runs becomes an input. 20 steps.

3 · Replay

One call, no model.

routine_run("play-release", {
  version_name: "2.15.0 (215)",
  release_notes: "Dark mode for the reader.",
  bundle: "kavyam-2.15.0.aab"
})

Step 9 changed, so the model was called once. 20 steps, 0:31.

Your agent starts from zero every time.

Browser agents read the page, think, act, and read the page again, for every single step. Run the same chore tomorrow and you pay for all of it again.

539k: tokens to change one YouTube description
17: tool calls, and 6 screenshots sent to the model
0:56: start to finish, every time it runs
+19k: tokens of page snapshots left in your agent’s context afterwards

That is one small edit to a YouTube description, done by hand through Claude in Chrome. On a longer console task it took 44 tool calls, 910k tokens and 1:33, and one run in five said it was done without starting the rollout.

Measured, not estimated.

Up to 83% cheaper and 3.5 times faster than the same agent doing the task by hand. Each arrow goes from by hand to replay.

Same page as recordedTokensTimeCost

Edit a YouTube descriptionReal YouTube, signed in 539k → 97k−82%0:56 → 0:19−66%$0.21 → $0.04−81%

Release a build to testersReplica of a store console, one form 430k → 131k−70%0:46 → 0:15−67%$0.17 → $0.04−73%

Release with every settingThe same console, a long form 910k → 199k−78%1:33 → 0:27−71%$0.32 → $0.06−83%

After the site changedTokensTimeCost

Release a build to testersOne button renamed since the recording 430k → 198k−54%0:46 → 0:28−40%$0.17 → $0.06−66%

Release with every settingOne button renamed since the recording 910k → 305k−66%1:33 → 0:41−55%$0.32 → $0.08−76%

A button was renamed since the recording. That added 12 to 15 seconds and 1 to 2 cents: your agent fixed that one step and the fix was saved.

Best for chores you repeat. The first run still goes through your agent, and a task you do once is just as quick to ask for.

Claude Sonnet 5.5 through Claude in Chrome, against the saved routine, on one machine. Median of 5 runs each, a small sample. Costs are what Claude Code reported at list prices. The YouTube task ran on real YouTube; the other two on a replica of a store console. Token totals include context read back from cache.

It checks the page before every step.

Same page

It runs straight through.

No model, no tokens.

A step moved

Only that step goes to your agent.

It fixes it, and the fix is saved for next time.

It breaks

Your agent takes over.

From that step, with everything done so far.

It stops where you would.

Pauses before anything you can’t undo.

Publishing, paying, deleting. It waits for you, or for your agent to say yes.

Never saves a password.

What you type into a password field stays in the browser and shows up as a placeholder.

Learns from you too.

Do the task once yourself. It saves the same way an agent run does.

Routines other people keep working.

Publish a routine, or install one someone else maintains. Every routine is read by a person before it is listed, and you see each step before it installs.

Browse the directory Read the guidelines

Set up in two steps.

Add the extension

Chrome

Edge

Firefox

Brave, Arc and Opera use the Chrome version. Signing in with Google is optional.

Connect your agent

npm $ npx agentic-routines setup pnpm $ pnpm dlx agentic-routines setup bun $ bunx agentic-routines setup yarn $ yarn dlx agentic-routines setup

✓ Claude Code ~/.claude.json

✓ Codex ~/.codex/config.toml

✓ GitHub Copilot VS Code mcp.json

✓ OpenCode ~/.config/opencode/opencode.json

✓ Chrome extension connected

Paste this into Claude Code, Codex, OpenCode or any agent chat and it sets everything up for you.

Set up Agentic Routines on this machine.
1. Run "npx agentic-routines setup". If I use pnpm, bun or yarn, run it with "pnpm dlx", "bunx" or "yarn dlx" instead.
2. Run "npx agentic-routines doctor" and tell me which agents and browsers are connected and what is still missing.
3. If my browser has no extension yet, tell me which store page to open. Do not install anything else.
4. Tell me to restart this agent so it loads the browser tools.