Predictable
Known runtimes, dependable utilities, and repeatable setup across supported agents.
Open-source · Local workbench · Human-controlled
Your AI tools.
One clean workbench.
Build a dependable local workbench for multiple AI coding agents—then govern the runtimes, shared guidance, tool-server configuration, project instructions, and execution evidence around them without taking control away from you.
- Preview firstSee material changes before apply
- Local by designNo hosted control plane required
- Honest boundariesUnsupported means unsupported
Why AI Foundry Desk
More agents should not mean more workstation entropy.
Each AI tool arrives with its own runtime assumptions, installers, skills, profiles, and diagnostics. The result is powerful—but difficult to reason about.
AFD gives those tools a common foundation and a governed way to align what should be shared. Credentials, projects, conversations, memory, and agent-native behavior remain under their own control.
Reviewable
Dry-runs and explicit plans make consequential changes visible before they happen.
Recoverable
Drift stays visible, managed files are backed up, and rollback remains deliberately scoped.
How the workbench fits together
Three setup layers. Shared tools. One manager keeping them reviewable.
AFD organizes workstation setup into three practical layers, supported by a common toolbox and Agent Manager. Start with the machine, add the agents you choose, then apply reviewed capability bundles without losing sight of what changed.
03
Agent Manager recipes
Provision capabilities deliberately.
Plan a reviewed bundle of skills, profiles, tools, or local observability. Apply only the exact plan you reviewed.
02
Agent setup
Bring the tools you actually use.
Detect and install supported desktop agents and CLIs consistently. Authentication stays a manual step inside each product.
01
Foundation
Give every agent dependable ground.
Managed runtimes, PATH, package tools, diagnostics, and a compact toolbox create a predictable workstation baseline.
rgfdjqyqbatdelta
Common Agent Toolbox
Agent Managercatalog · review · sync · verify · recipes
Designed for a mixed desk
Claude · Codex · Pi · Grok · Hermes · Antigravity
AFD reports each integration’s real boundary. It does not manufacture parity where an agent lacks a stable profile, skill, or telemetry contract.
What you gain
A workbench you can inspect—not another black box.
The useful unit is not another installer. It is a manageable environment with enough evidence to understand its state and enough restraint to preserve what AFD does not own.
Operate
One interactive terminal workspace
Use afd tui to search and run all 74 workbench actions through guided inputs and structured results—or use the same application services from the CLI.
Configure
MCP servers, synchronized by scope
Discover, adopt, enable, disable, move, and synchronize selected server definitions across user and project scopes with redacted plans and hash-bound approval.
Govern
One project policy, tested across agents
Audit repository instructions, stage minimal adapters outside the project, run disposable cross-agent checks, and apply only with matching evidence.
Recover
Scoped repair and rollback
Backups precede managed replacement. Repair reconciles declared AFD state; rollback removes only effects recorded by the reviewed recipe.
Current release · v0.6.2
The whole workbench, now navigable from one terminal workspace.
The latest releases extend AFD without weakening its central contract: inspect first, bind approval to the exact plan, preserve unmanaged state, and fail closed when evidence is missing.
v0.6.2
All 74 actions in afd tui
Search capabilities, edit inputs without shell quoting, review structured results, and switch between keyboard-accessible responsive layouts while write actions keep the CLI’s confirmation, drift, receipt, and rollback protections.
Read the 0.6.2 release notes ↗v0.5.0
Scope-aware MCP configuration
Manage selected tool-server definitions across canonical user and project registries plus verified native agent files. Plans redact values, preserve unrelated settings, and roll back compound writes on failure.
MCP workflow ↗v0.4.0
Evidence-gated project harnesses
Consolidate repository guidance around one canonical policy, test minimal adapters in disposable workspaces, and require matching live evidence before transactional apply.
Harness workflow ↗Safety by design
Keep the human at the consequential boundary.
AFD helps you move faster without teaching your workstation to make invisible decisions on your behalf.
- Preview before apply. Important changes expose a dry-run or explicit plan.
- Preserve drift. Differences are reported, not silently normalized.
- Own only the declared surface. Repair avoids unrelated machine state.
- Keep sensitive context separate. Tokens, logins, projects, conversations, memory, and proprietary plugins are not shared.
Platform truth
Confidence should match the evidence.
Support labels reflect environments that have actually been exercised. Experimental means experimental, and unvalidated environments remain visible as contribution opportunities.
Validated experience
Windows x64
Complete Layers 1–3 experience, toolbox, Agent Manager, diagnostics, repair, backups, and verification.
Validated fixture
Ubuntu 26.04.1 on WSL2
Native Layer adapters, portable recipes, optional Docker host capability, diagnostics, repair, and verification within the documented x86_64 scope.
Experimental
macOS
Detection only today. Native adapters and independent validation remain open work—not a compatibility claim.
Need the exact boundary?Review the platform support matrix ↗
Get started
Inspect first. Apply when the plan makes sense.
Start from the stable release or install the npm preview, then begin with read-only commands. Administrator access is not required for the normal Windows setup.
Release status: v0.6.2 is the latest stable GitHub release. For easier evaluation, the public ai-foundry-desk@0.6.3-rc.1 npm package installs the afd command; pin the release-candidate version explicitly. Both channels require Node.js 24 or newer.
npm install --global ai-foundry-desk@0.6.3-rc.1
PS> afd init --dry-run
PS> afd doctor
PS> afd tui
PS> afd layer1 --dry-run
# Review the output. Apply only when ready.
PS> afd layer1 --apply
PS> afd layer2 --dry-runOpen source, inspectable by default
Build the workstation layer in the open.
AI Foundry Desk is MIT licensed. Read the architecture, trace every managed boundary, challenge the roadmap, or help validate another platform without weakening the portable core.