Tandem Browser

AdvisoryAudited by Static analysis on May 12, 2026.

Overview

No suspicious patterns detected.

Findings (0)

Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.

What this means

The agent could interact with websites in the Tandem Browser, including forms or pages tied to user accounts, if the user asks it to do so.

Why it was flagged

The skill intentionally gives the agent browser interaction abilities, including clicking and filling fields. This is expected for the stated purpose, but can have real effects if used on logged-in or transactional pages.

Skill content
Browse, snapshot, click, type, navigate ... mcporter call tandem tandem_snapshot_click ... tandem_snapshot_fill
Recommendation

Use this skill for trusted browsing tasks and require explicit confirmation before submissions, purchases, account changes, or public posts.

What this means

A compromised or malicious web page may have more opportunity to affect the local environment than it would in a sandboxed browser.

Why it was flagged

The skill documents running the browser without sandboxing on Linux. This is disclosed and may be required by Tandem, but it reduces browser isolation when visiting untrusted pages.

Skill content
`--no-sandbox` — always required on Linux
Recommendation

Avoid using the unsandboxed browser for unknown or high-risk sites, and run it in a contained user account or environment when possible.

What this means

Anyone or any process with access to that token may be able to control the local Tandem Browser API.

Why it was flagged

The skill discloses use of a local bearer token for the Tandem API. This is expected for connecting to the local MCP bridge, but it is still credential-based access to browser-control functionality.

Skill content
Auth token: `~/.tandem/api-token` (Bearer token)
Recommendation

Protect the token file, avoid sharing logs or screenshots containing it, and rotate it if it may have been exposed.

What this means

The safety of the skill also depends on the installed mcporter and Tandem Browser versions, which were not reviewed here.

Why it was flagged

The skill depends on an external mcporter binary and Tandem Browser components that are not included in the reviewed artifacts. This is a provenance and review-context limitation, not evidence of malicious behavior.

Skill content
Required binaries (all must exist): mcporter; No install spec — this is an instruction-only skill.
Recommendation

Install mcporter and Tandem Browser only from trusted sources and keep them updated.

NoteHigh Confidence
ASI10: Rogue Agents
What this means

A running daemon may continue making Tandem tools available for multi-step work until stopped.

Why it was flagged

The skill documents a persistent daemon mode and systemd service, but presents them as manual, user-directed options rather than hidden background behavior.

Skill content
Daemon mode (persistent, for multi-step workflows) ... `mcporter daemon start` ... Service is disabled (start manually when needed)
Recommendation

Start the daemon only when needed and stop it after use if you do not want ongoing browser-control availability.

NoteHigh Confidence
ASI01: Agent Goal Hijack
What this means

Malicious web page content could try to influence the agent if treated as instructions instead of untrusted page data.

Why it was flagged

The skill acknowledges that raw page HTML can contain prompt-injection content. The warning and recommendation to prefer safer content-reading methods are purpose-aligned mitigations.

Skill content
`tandem_get_page_html` | Last resort. Raw HTML, prompt-injection exposed.
Recommendation

Treat web page content as untrusted evidence and avoid using raw HTML unless necessary.