dassi

Dassi is an AI browser agent built by Omnify Labs in San Francisco that runs inside the Chrome, Edge, or Brave side panel. It completes tasks and recurring workflows — filling forms, extracting data, updating CRMs, triaging inboxes — using the accounts users are already signed into, with no passwords handed over and no integrations to wire up. The homepage states it is trusted by 30,000+ Chrome users and rated 4.8/5 on the Chrome Web Store.

What It Is

Dassi sits in the browser's side panel and accepts plain-English instructions. Users can describe a task once, or walk the agent through it while it watches, and Dassi takes over: clicking buttons, navigating pages, filling forms, and reading documents. When a task recurs, it can be saved as a workflow with a schedule, approval rules, a run history, and persistent memory so corrections carry forward instead of being forgotten at the end of a chat session. The agent itself runs on the user's machine inside the browser sandbox, not on Dassi's servers.

How the Workflow Model Works

Dassi distinguishes between one-off tasks and owned workflows:

• Tasks are ad-hoc requests — research a supplier, pull numbers from a dashboard, fill a form.
• Workflows are tasks Dassi has taken ownership of: they have a schedule (daily, weekdays, every few days, monthly), saved plans in plain language, defined approval gates, and a run history.
• When a scheduled run needs a human decision, Dassi pauses and pings the user via Telegram or email; the user can approve from a phone without returning to their desk.
• Scheduled runs execute in the user's own browser, so the browser must be open when a run fires; if it was closed, Dassi runs the job shortly after it reopens.

Cost and Caching Architecture

Dassi is built around two mechanisms designed to reduce model costs on repeated runs:

• Caching: Within a run, Dassi keeps the job's context byte-stable so every step after the first reads from the model provider's cache — the site claims up to 90% off input tokens on Claude.
• Learned site skills: As Dassi works a site, it writes itself a small JavaScript library for that site. Subsequent runs call a function instead of re-deriving selectors from screenshots and HTML, shifting work from expensive model calls to cheap local execution.

The site states that work you repeat gets cheaper the longer you keep it, and that a correction to a workflow — wrong supplier, wrong column, wrong tone — is stored in persistent memory and carries into the next run.

Model Flexibility

Dassi supports Claude, GPT, Gemini, Grok, DeepSeek, Kimi, and anything on OpenRouter. Users can bring their own API keys or use Dassi's included managed credits, and can switch models at any time with no lock-in. The site frames the model as "the engine" and the workflow as "the thing you keep."

Enterprise Deployment

Dassi offers an enterprise configuration with several security-oriented properties:

• The agent runs inside the browser sandbox and cannot read the local filesystem, install software, or escalate OS privileges.
• Page content, conversation history, and documents stay on the employee's device and flow directly to the organization's own AI gateway (AWS Bedrock, Azure OpenAI, GCP Vertex, or self-hosted endpoints) — never through Dassi's servers.
• SSO support covers Okta, Azure AD/Entra ID, Google Workspace, OneLogin, and any SAML 2.0 or OIDC provider, with SCIM for provisioning.
• Admins can pool credits and allocate them by team, department, or project with monthly caps.
• Tamper-evident audit logs cover every prompt, action, and credit consumed by user and timestamp, exportable to Splunk, Datadog, or any SIEM.
• A central skill library preserves employee-built workflows when staff leave.

The founders previously shipped the WhatsApp Business API at Meta and worked on Microsoft Word, and the About page cites WhatsApp's privacy-first architecture as a direct influence on Dassi's design.

Update: Current Status

Dassi is in active beta, with the homepage labeling the product "Beta" and offering free credits for 7 days to new accounts with no card required. The site lists a changelog page and an active blog, and the pricing page references annual billing as the default display, with plans ranging from a developer bring-your-own-key option up to a high-usage tier. The Chrome Web Store extension ID is publicly listed, and the product supports a CLI in addition to the browser extension.