runtime 0.2.0 / self-hosted / single-machine

A browser agent that reports what actually happened, not what the function call asked for.

orvima re-reads the DOM after each meaningful step and returns verified: true or verified: false. If it cannot prove what happened, it does not upgrade uncertainty into success.

49tools exposed to MCP hosts
8product domains
1DOM read after every state change
$ browse_navigate --url https://app.example.com/account/settings
verified: true
$ browse_click --selector 'button.delete[data-account=all]'
classifier: destructive outward action
result: held for human approval
click: not executed

The evidence is in the execution path, not only in the launch page.

Ordinary execution returns a result after dispatch. orvima returns what it could confirm and keeps a transcript you can re-read.

What a typical MCP result sounds like

step 1: browse_navigate → ok

step 2: browse_click → ok

step 3: browse_click → ok

result: success

The transport itself completed. That does not say what the page displayed.

What orvima keeps after the run

step 1: browse_navigate → verified: true

step 2: browse_click → verified: false

step 3: gate_check → refused

state captured, approval card written, run stopped

Every edge stays in memory with evidence instead of disappearing into generated prose.

Tools shaped around execution truth.

Eight domains share one object model. The tool names below are the public interface MCP hosts actually call.

browse_navigatebrowse_snapshotbrowse_clickbrowse_typebrowse_wait_formail_composewhatsapp_sendvoice_speakmedia_transcriptgate_checkbrowse_extract_textbrowse_download_file
Browsenavigate, read, click, type, verify11 browser tools
Mailcompose and deliver from a local account4 mail tools
WhatsAppsend through a running local session2 message tools
Voicespeak and transcribe2 voice tools
Mediacapture, extract, download5 media tools
Math and sciencenumbers, units, analysis helpers6 compute tools
Gaterun, ask, or refuse before crossing outwardmultiple gate tools
Routertransport and control surfaceMCP plumbing

Safety gate

Refusal is cheaper than cleaning up a bad action.

The classifier labels actions as run, ask, or refuse. A high-risk click is stopped before execution, and the reason is visible instead of hiding in a context trail.

held-approval.txt
risk: deletes_all_account_data
reason: destructive account selector targets all data
action: browse_click
result: not executed

Human-reviewable, not only machine-readable.

The card preserves the selector, arguments, risk, and reason that a human would use to make the decision. A failed run can be audited later without reconstructing it from a deleted chat.

A classifier that asks in the right place reduces unobservable state changes far more than another prompt layer.

Install locally and keep the evidence where the run happens.

No cloud approval surface is required for the local execution path. The transcripts and run evidence stay next to the action that generated them.

$irm https://raw.githubusercontent.com/Sachitt-AV-08/orvima/main/install.ps1 | iex

The command is the same one documented in the repository and does not include legacy placeholder steps.