Built for VS Code Repository-aware AI tasks with local controls and verifiable results. See how it works

Visible trust boundaries

AI reasoning. Local authority.

WebCodingAgent separates the eligible AI connection supplying model responses from the VS Code extension that can access and change your repository. No component receives more power than its job requires.

Three components, hard boundaries

Authority stays close to the thing it protects.

The wcagent service coordinates tasks. Desktop VS Code owns repository access and execution. For eligible web-account integrations, the browser connector transports prompts and responses without exporting provider cookies or credentials.

LOCAL

VS Code extension

Context · workspace trust · approvals · commands · mutations · redaction · recovery

CLOUD

wcagent service

Plans · workflows · risk policy · task state · audit events · coordination

BROWSER

Browser connector

Eligible provider exchange · provider session stays in profile · no repository access

Local execution

Requests do not become actions by themselves.

The extension independently validates every tool request before anything reaches the repository. Ask-for-approval mode pauses each command and mutation in the conversation; full access remains bounded to the active task and local policy.

01

Workspace confinement

Lexical and real paths must remain inside the repository, including protection against symlink escapes.

02

Conflict-safe writes

Existing-file edits require the SHA-256 returned by an earlier read and apply atomically.

03

Visible approvals

Commands and mutations can stop for explicit approval without losing the rest of the task.

04

Secret boundaries

Credential-shaped values are redacted before supported tool output reaches the cloud.

05

Cancellation

Long-running commands stay observable and can be cancelled from the same task UI.

06

Independent policy

Execution-environment limits are enforced again in VS Code, not trusted from a cloud label.

Browser authentication

Provider credentials stay in your browser.

For an eligible web-account integration, the connector uses a provider account already authenticated in your browser. Normal wcagent protocols return model text and bounded runtime metadata—not passwords, cookies, authentication headers, or full browser-session payloads.

Transport

Outbound pairing

The browser connector requests work over its paired channel; the wcagent service does not open an inbound path to your machine.

Scope

No repository access

The connector can manage validated worker tabs for the selected AI provider but cannot read or mutate workspace files.

Control

Explicit provider choice

You choose the eligible connection that supplies the task’s model responses, and connector activity remains visible and reversible.

Failure

No silent fallback

Signed-out, provider-limited, or uncertain sessions pause with a typed error instead of switching accounts or reporting success.

Provider controls

wcagent cannot override a provider.

wcagent does not increase, reset, conceal, or bypass usage limits. It does not defeat authentication, CAPTCHAs, safety systems, geographic restrictions, or other technical access controls.

Eligibility

Authorization depends on the provider

Use only a provider, account type, and connection method whose applicable terms permit its use with wcagent.

Separate product

Model access is not included

wcagent does not sell or resell access to an AI model. You supply and maintain your AI account separately.

Provider response

Limits pause the connection

If a provider refuses or limits a request, wcagent reports that result and cannot make the provider accept it.

Terms

Your provider agreement still applies

Provider availability, charges, usage limits, account settings, policies, and terms remain in force.

Evidence and governance

Know what the system knew—and what it did.

Policy decisions, approvals, local tool results, model routes, and completion evidence remain attached to the task. Hash-linked audit records make later review possible without pretending that an audit log replaces secure execution.

01
Requested and observed models stay separateThe interface does not claim a specific model unless the connector can observe it.
02
Completion holds are explicitMissing tests, failed builds, and unresolved specialist findings prevent a clean finish.
03
Recovery avoids accidental replayStable IDs and durable journals distinguish completed actions from work still safe to resume.