The flintspark platform
The flintspark platform is where teams teach, verify, run and audit browser workflows in one place. Flint, flintspark's agent, learns a task from one demonstration and runs it on the fastest route that works. By default it asks a person before anything it cannot undo, and every approval goes into a tamper-evident audit log.
- TeachA person does the task once
- VerifyTwo verified runs before a skill is trusted
- RunThe fastest route that works, step by step
- AuditApprovals and safety events, hash-chained
What is in the platform?
Seven parts. They share one record of what Flint learned, what it did and who approved it.
- Teach
Record a task once, review it, and trust the skill after two verified runs.
- Skills
Versioned units of work learned once and reused across workflows.
- Workflows
Skills, checks, questions and approvals that wait for people and survive restarts.
- Approvals
A signed, single-use grant before any irreversible action, by default.
- Policies
Per-site rules for what Flint may do. The model never decides.
- Audit trail
Run records for review and a hash-chained log of approvals and safety events.
- Execution routes
An API, a page address, recorded actions or an AI agent: the fastest that works.
For the whole loop in order, read how flintspark learns and runs a workflow. For the controls together, read trust and safety.
Where does it run today?
On one operator's machine, in a dedicated browser profile that Flint uses only for its own work. The operator logs into each site once, with their own password manager. Flint never handles passwords: password fields are masked in the page before anything is recorded.
Cloud browsers, so Flint can run away from a person's own machine, are on the roadmap.
Deployment
- Today
- One operator's machine built
- Browser
- A dedicated profile, used only by Flint
- Sign-in
- The operator, once per site
- Later
- Cloud browsers roadmap
Which models does it use?
Each role in a workflow uses the model chosen in configuration, so a team can switch providers without changing workflows.
Model calls go through a model gateway, OpenRouter, with data collection off. The model sees only redacted text and screenshots, and never decides what is allowed: policy does.
Model per role
- Planning a runset in configuration
- Learning a skill from a recordingset in configuration
- Acting on a pageset in configuration
- Checking what a page showsset in configuration
Gateway: OpenRouter · data collection off · redacted inputs
What is not built yet?
These are not in the platform today, and each is labelled as it stands.
- Cloud browsers are on the roadmap.
- A credential vault is on the roadmap.
- Single sign-on and accounts for several people are on the roadmap.
- Use inside private internal web apps is on the roadmap. Flint has been shown on public sites and official APIs.
- Pharma incoming-lot release is in build, and no run of it has passed yet.
- flintspark holds no certifications. Each is under consideration; see trust and safety.
See Flint run
Bring one workflow you'd rather not do by hand. We'll show Flint learning one like it, and stopping before anything it can't undo.
We reply from hello@flintspark.ai.