AI Foundry Desk v0.6.2

AI Foundry Desk v0.6.2

• Explore all 74 AFD actions from one interactive terminal workspace with afd tui, organized by
  outcome so you can discover the right workflow without memorizing commands.
• Search capabilities, edit inputs without shell quoting, review structured results, and move between
  keyboard-accessible responsive layouts. High-contrast, monochrome, ASCII, reduced-motion, and
  mouse-free modes keep the interface usable across terminals and accessibility needs.
• Keep control of every change: read-only actions run directly, while writes show their scope and
  require the same explicit confirmation, drift checks, receipts, and rollback protections as the
  CLI.
• Receive checksummed assets from a release path that is now exercised on pull requests before a tag
  can trigger public publication.

AI Foundry Desk 0.5.0 — One MCP setup across your agents

One MCP setup, across the agents you use

AI Foundry Desk 0.5.0 introduces a safe, scope-aware way to keep MCP server configuration
consistent. You can discover an existing definition, adopt it once, sync it to selected agents,
enable or disable it globally or for one project, and move it between user and project scope without
manually editing every client.

What this gives you

• One canonical user registry and one optional project registry instead of agent-by-agent drift.
• Native user and project synchronization for Claude Code, Codex, Antigravity, and Grok, native
  user synchronization for Hermes, and explicit opt-in Pi support through pinned
  pi-mcp-adapter 2.31.0.
• Project-only disable overrides, so a globally useful server can stay off in a sensitive project.
• Atomic scope moves: AFD creates the destination, reconciles native files, and removes the old
  ownership in one transaction.
• Redacted discovery and dry-run plans with a content-derived confirmation token.
• Automatic rollback if any required write fails, and stale-plan rejection if an agent changes a
  file after preview.

AFD preserves unrelated JSON, TOML, and YAML settings and never copies OAuth or login state.
Secret-like headers and environment values must be references to existing environment variables;
their values are not printed in plans.

Run afd catalog to see the MCP capability of each installed agent and scope. Default all-agent
operations fail closed when an adapter is not proven or consented: Pi requires an explicit
--enable-pi-adapter preview before AFD declares its pinned third-party adapter, and Hermes does
not currently have a stable project-scoped MCP surface. You can
deliberately target the verified subset with --agents.

Start with a preview

afd mcp status --scope effective --project .
afd mcp sync --scope effective --project . --agents claude-code,codex,grok --dry-run

Review the paths, hashes, capability blockers, and approval token before applying the same command
with --confirm <plan-token>.

AI Foundry Desk 0.4.0 — One project policy, consistent agents

AI Foundry Desk 0.4.0 — One project policy, consistent agents

AFD 0.4.0 helps you replace a collection of drifting, agent-specific instruction files with one
reviewable project harness. Codex, Claude Code, Pi, and other selected agents can share the same
project policy without maintaining several copied versions of it.

See the problem before changing the project

afd harness audit inventories the repository without writing to it. It shows which policy is
canonical, which agent files are pointers or independent instruction sets, where instructions are
duplicated, and which discovery claims are not yet verified.

The resulting plan is available in a concise human view or structured JSON. Every proposal is bound
to the exact policy content, Git revision, and visible workspace state, so a reviewed decision cannot
quietly drift onto a newer project state.

Review a real proposal, not a promise

AFD generates minimal adapters that point to the canonical policy rather than copying its rules.
You can render them into an external staging directory and inspect every proposed creation, managed
update, or legacy removal without touching the repository.

If an existing agent file contains distinct project-owned guidance, AFD preserves it and blocks the
definitive apply. You decide how that guidance should be reconciled; AFD does not discard it.

Test every selected agent before apply

The live harness test builds a disposable policy-only workspace and starts fresh, bounded,
read-only sessions. Each selected agent must independently identify the same canonical policy and
return the same fingerprint. Missing commands, unsupported safety contracts, inconsistent behavior,
timeouts, or write attempts fail the gate instead of being silently skipped.

Apply with evidence and keep a way back

Applying a harness requires both the exact reviewed plan token and its matching passing live test
evidence. The update is transactional, followed by hash and workspace verification, and recorded in
a private per-user receipt outside the project.

Rollback is equally scoped: it refuses a project that has drifted since apply, restores the exact
prior bytes, and records what was restored. Conservative cleanup can remove proven redundant thin
pointers such as an unnecessary AGY.md, but never removes a file based on its name alone.

A practical, incremental workflow

1. Audit the repository.
2. Review or externally stage the exact plan.
3. Run the disposable multi-agent test and save its evidence.
4. Confirm the plan token to apply.
5. Verify the result or use the receipt for exact rollback.

The workflow and agent coverage are documented in docs/PROJECT-HARNESSES.md. Installing or
upgrading AFD still does not apply workstation layers, start telemetry, or modify a project.

Windows x64 remains the fully validated environment. The checksum-verified portable CLI bootstrap
remains available for the documented POSIX path; macOS support is experimental.

AI Foundry Desk 0.3.1 - A workstation that stays dependable after updates

AI Foundry Desk 0.3.1 — A workstation that stays dependable after updates

AFD 0.3.1 focuses on a simple outcome: your tools should keep working consistently in a normal console, inside Codex tasks, and when you start a fresh session.

Updates no longer leave agent tasks stranded

You can continue updating tools normally with WinGet. AFD now recognizes when an update replaced a tool and unintentionally removed the access an isolated Codex task needs. It explains the exact problem and gives you a reviewable recovery path instead of asking you to reinstall tools, freeze versions, or route everyday commands through AFD.

One clear health check

afd doctor now gives you one place to confirm that the workstation is genuinely usable—not merely configured. It checks the effective identity, runtime versions, executable paths, and whether the core toolchain can actually run in the current context.

That makes it easier to distinguish three different situations:

• the normal user environment is healthy;
• an isolated agent task lacks access to a specific tool;
• optional services such as local observability are simply not active.

Safer recovery with less hidden behavior

AFD does not take over WinGet, replace your tools, or silently change permissions during startup or diagnosis. If recovery is needed, you first see the exact proposed paths and read/execute grants. Applying the repair is a separate, explicit decision and keeps a rollback snapshot.

More predictable project work

Normal-user sessions and isolated agent tasks now share the same project dependency baseline. This removes a class of confusing failures where the repository worked in one context but tried to rebuild or discard dependencies in another.

Private observability when you want it

This release also establishes the optional local observability experience for understanding agent runs and usage. It remains private, local-first, and explicitly activated. Installing or upgrading AFD does not start telemetry, background services, or workstation layers.

When observability is not active, AFD now reports that state clearly instead of returning a low-level connection error.

What remains under your control

• You choose when third-party tools are updated.
• You review permission repairs before applying them.
• You choose whether local observability and autostart are enabled.
• Installing AFD does not automatically apply Layer 1 or Layer 2 changes.

Windows x64 is the fully validated environment for this release. The checksum-verified portable CLI bootstrap remains available for the documented POSIX path; macOS support is still experimental.

Use the checksum-verified install or upgrade command in the README, then run afd doctor.

AI Foundry Desk v0.1.2

AI Foundry Desk v0.1.2

Windows x64 includes the full PowerShell bootstrap and Layer 1 doctor/fix. The separate POSIX bootstrap is validated for the portable CLI cycle on Ubuntu 26.04.1 LTS under WSL2 x86_64; native Linux Layers remain unavailable. macOS detection is experimental and unvalidated. Installing the CLI never applies Layer 1 or Layer 2.

AI Foundry Desk v0.1.1

AI Foundry Desk v0.1.1

Verified Windows x64 bootstrap release. Installing the CLI does not apply Layer 1 or Layer 2.