You handle sign-in
You sign in yourself, in a plain browser window, including any MFA. Handhelm never types passwords or codes.
HUMAN-SUPERVISED BROWSER AUTOMATION
Let the work move.
You hold the helm.
Handhelm runs checked browser tasks in your own signed-in session. Sealed read-only flows can run unattended while the session is valid and return typed results. Irreversible actions remain behind approval gates; the phone approval service is not installed yet.
See the method Private development · not yet available
Models draft the steps. The runner executes them. Local approvals are not isolated from other processes running as the same user.
EXAMPLE RUN OPERATOR PRESENT
- 01 Check before action
- 02 Ask before consequence
- 03 Stop when uncertain
- 04 Keep a record
FROM DRAFT TO RUN
A careful flow has
four distinct moments.
Writing a flow and running it are separate jobs. A model can do the writing. The running never involves one.
- 01
Learn the page
Handhelm reads each page once, follows no links and blocks any write. It records what it expects to find: the address, the controls and the text.
OBSERVE - 02
Draft, then review
A model sees a minimized map of visible page structure and labels, plus the shape of a sample data file. Filled-in browser values and the file’s values stay local. Code rejects drafts that break the rules; you review the steps in plain words.
REVIEW - 03
Match before acting
On every run, Handhelm checks the address, the controls and the expected state of each page. If the page has changed, it stops instead of guessing.
VERIFY - 04
Pause at the approval gate
Submitting, paying, signing or deleting stops at an approval step. The request shows the site, control and values. The local CLI supports expiring, weekly-capped standing grants for exact reviewed flow versions and attested data files. These grants are same-user policy: any process running as the same user can create or use one. The separate passkey service is not installed.
APPROVE
WHERE MODELS FIT
Let a model write.
Make the boundary explicit.
In Handhelm, a page change never triggers another model call. If the page does not match, the run stops. A repair is checked and reviewed before the next run.
WRITES THE STEPSA model, or you
Drafts a flow from read-only page maps. Code checks it, and you review it once.
STARTS THE RUNAn agent, or you
Picks a reviewed flow, supplies the data and reads a plain result. Stops return a state-bound approval request; local approval is not isolated yet.
CARRIES IT OUTHandhelm only
Runs the reviewed steps exactly. No model takes part, so text planted on a page has nothing to instruct.
APPROVESLocal command line today
One-run approvals bind to the page state. Standing grants also bind exact reviewed flow versions, attested data digests, an account, expiry and a weekly budget. These are same-user policy; any same-user process can invoke the CLI. The separate passkey service remains uninstalled.
THE LIMITS ARE PART OF THE DESIGN
A good system
knows when to stop.
Handhelm is built to keep the operator in control of the browser, the account and the outcome. The boundary is visible before the moment arrives.
Read the operating principlesOPERATOR
Challenges mean stop
CAPTCHA, MFA prompts and rate limits are never bypassed.
Reviewed before it writes
A flow changes nothing until a person accepts that exact version. Any edit needs a new review.
Signing remains with the operator
The named person remains responsible for each filing. Patent signing helpers need further review; agents must not invoke them.
Evidence stays verifiable
When evidence is enabled, Handhelm records page checks and step outcomes. Before an irreversible action, it requires an evidence bundle with a masked screenshot and state-bound approval; a manifest verifies the saved files.
NO SHORTCUTS AROUND THE PERSONNO MODEL IN THE RUNNO ACTION ON A MISMATCH
BUILT FOR CAREFUL, REPEATABLE WORK
EARLY DEVELOPMENTUseful repetition.
Human judgment intact.
FORM WORKFLOWSRepeat the steps.
Repeat the steps.
Review every outcome.
Declared sequences for pages where the state must be checked and the person remains responsible for each filing or submission.
AGENT HANDOFFLet your agent start it.
Let your agent start it.
Know the boundary.
An agent can start a reviewed flow and read a plain result. Local standing grants allow bounded repeat approvals, but any same-user process can issue them.
SEARCH CONSOLE · READ ONLYReview properties.
Review properties.
Compare the next scan.
Read available API reports across multiple authorized properties, review evidence-backed SEO leads, then Scan again to compare the same scope. UI-only issue reports remain uncovered.
These are development areas, and nothing here is available yet. Handhelm is independent and is not affiliated with any government office or third-party platform.
CURRENTLY IN DEVELOPMENT
What is Handhelm?
Handhelm is a local browser-automation harness for repeatable tasks. It replays declared steps in the operator’s signed-in browser, checks each page before acting, and records outcomes. Models can draft flows at setup; deterministic code runs them. The project is in private development, with no public download or hosted service.
Current status
Sealed read-only flows can run unattended while the signed-in session is valid. They return typed results and stop for sign-in or challenges. Reviewed flows recheck each page; irreversible actions require approval.
Optional schedules are available but are not installed by default. Model-authored flow drafts are checked by code. On a practice site, the authoring loop produced a working flow in three rounds.
Approval boundary: the local CLI issues page-state-bound one-run approvals and supports standing grants for exact reviewed flow versions, attested data digests and a bound account, with expiry, revocation and a per-flow weekly budget. These remain same-user policy: any process running as that user can create or use them. The separate passkey service is uninstalled and disabled.
The current code includes a read-only Google Search Console API scanner for authorized properties. It reviews available Search Analytics data, sitemap counts, and selected indexed URLs across one or more properties. Scan again repeats the same saved properties, dates, and report selections, then compares a fresh local snapshot; it does not ask Google to crawl pages, submit a sitemap, or change validation state. The results are SEO review leads, not guaranteed diagnoses. Page Indexing and other UI-only issue reports remain outside this API connector's coverage. The code also includes offline USPTO patent-preparation tools, a small set of Trademark Center browser flows, and Kaggle flows. Patent Center browser flows remain planned.