v1.4.2 Beta · Live

Build a Claude Code agent team, tailored to your codebase.

Foundry can read your project and write a complete .claude/ for you: agents, hooks, rules, behaviors. Edit everything Foundry helps you make in Forge, and reset to baseline at any time.

Free during beta · Windows 10/11 · Bring your own Anthropic API key
Foundry's team page showing five generated agents with their 2x2 taxonomy robots.

Your team after Foundry runs. Five agents matched to a real billing service: security, review, history, compliance, and tests. Every card pulls its 2×2 robot from role + kind.

New in 1.4

Schedule your agents to run themselves.

The Routines bucket turns any agent on your team into a scheduled task that runs in Claude Code Desktop. Compose agents, declare the connectors the routine needs, and Foundry writes a SKILL.md Desktop picks up automatically.

  • Connector-aware. Foundry checks the routine's required connectors against your installed MCPs (Gmail, Slack, Notion, graphify, anything local stdio) and warns before deploy.
  • Test before you ship. Every routine has a copy-pasteable Test command. Run it from your terminal once to confirm behavior, then click Deploy.
  • Project or global scope. Tag a routine for one project to keep it scoped, or tag global to surface in every Foundry project's bucket.
  • Failure-recovery built in. If a connector disappears after a routine deploys, the card shows a banner and reconnect link inline.
~/.claude/scheduled-tasks/morning-triage/SKILL.md
---
name: "Morning Triage"
description: "Scan inbox, summarize Slack DMs"
tags: [global]
---

## Required connectors
- claude.ai Gmail
- claude.ai Slack

## Composed agents
- inbox-triage
- slack-summarizer

## Schedule
cron: "0 8 * * 1-5"
→ Weekdays at 8:00 AM
Changelog

Latest releases.

The three most recent. See full build notes →

v1.4.2-beta Auto-updater fix + Graphify/MemPalace are optional May 2026
Fixed
  • Auto-updater works for prerelease users again. The v1.2.1 "channel pin" fix was actually the bug — pinning the channel to latest caused electron-updater's tag-discovery step to skip every prerelease in the feed silently, so the updater threw ERR_UPDATER_NO_PUBLISHED_VERSIONS and no banner ever appeared. Every prerelease user from v1.2.1 through v1.4.1 was affected. Dropping the pin lets the updater fall through to latest.yml correctly.
Added
  • Auto-updater is now observable. The update banner renders all five states (checking, available, downloading with progress %, downloaded, error with Retry). A log file at %APPDATA%/Foundry/logs/updater.log captures every updater event for diagnosing silent failures. The initial-check delay dropped from 15s to 5s.
  • Settings: an "App updates" panel with installed version and a manual "Check for updates" button.
  • Settings: an "Optional MCP servers" panel for installing Graphify and MemPalace from outside the wizard, with the same streaming install experience.
Changed
  • Graphify and MemPalace are optional now, not required. The full-screen install gate that blocked the app on launch is gone. Both retiered from required to recommended; users discover and install them from Settings when they want them. Both degrade gracefully at runtime: the MemPalace bridge is fire-and-forget safe and the Scholar prompt has codebase-graph signals sealed off via the existing review boundary.
v1.4.1-beta Windows install-gate hotfix May 2026
Fixed
  • Install gate no longer hard-locks Windows users whose Python is reachable only via the py launcher. The graphify and mempalace install routes were running the registry's bare pip install … and python -c … commands directly, bypassing the python-runtime normalization the rest of Foundry uses. On Windows machines where Python is installed via the python.org installer's default path (bare pip isn't on PATH, but py -3 -m pip is), preflight passed but the actual install steps exited with code 1 and users were stuck on the gate. Both install routes now detect the canonical interpreter and rewrite each step before running.
Added
  • "Skip for now" escape hatch on the install gate's error state. Defense in depth so users with broken Python setups aren't permanently locked out. Skipping persists across launches; users run with degraded features (no codebase analysis, no semantic memory) and can re-install from Settings any time.
v1.4.0-beta Routines: scheduled tasks for Claude Code Desktop May 2026
Added
  • New Routines bucket in Forge. Compose existing agents into a routine, set a schedule, pick required connectors (cloud MCPs like Gmail, Slack, Notion or local stdio servers), and choose project or global scope. Routines deploy to ~/.claude/scheduled-tasks/<slug>/SKILL.md for Claude Code Desktop to pick up.
  • Connector-aware scheduling. Foundry queries claude mcp list to detect what is installed and shows a status chip per required connector. Missing connectors warn you before deploy. If a connector disappears after deployment, the routine's card shows a failure-recovery banner.
  • Active/Draft lifecycle. Build a routine as a Draft, click "Test it" to copy a runnable command with the correct CWD, Deploy when ready. Undeploy or edit any time without losing the definition.
  • Four starter templates ship in the Routines bucket.
  • Scholar can propose routines via a new propose_routine tool inside the Routines bucket conversation.
  • Project vs. global scope on every routine via user-chosen visibility tags.
  • "Keep my routines" toggle in reset. Routines are machine-global, so the reset/nuke dialog now has an explicit opt-out before wiping them.
  • Apify capability registered as a recognized connector.
Fixed
  • HTTP-transport MCP apiKeySet detection corrected. Status chips now reflect configured keys correctly.
  • Local stdio MCP servers health-check correctly. Connector detection now passes projectPath, so local stdio servers no longer always appear as unavailable.
  • test-it command sets CWD correctly. The copied test command uses cd <path> && rather than --add-dir, so it runs in the expected project context.
v1.3.0-beta Graphify guidance reaches the orchestrator May 2026
Added
  • SessionStart hook tells the orchestrating Claude to query graphify before grepping. Without it, Claude saw graphify as registered MCP tools but had no instruction to use them proactively, so it defaulted to grep/read across unfamiliar files. Hook is managed, idempotent, and removed cleanly on reset.
  • .graphifyignore is written to your project on first scan. Mirrors the in-process skip list (node_modules, .next, dist, build, vendor, etc.) so the post-commit graphify update skips the same directories. Idempotent: never overwrites a file you have edited.
  • Post-commit summary regen. A new shim and a markered Foundry block in .git/hooks/post-commit keep the auto-loaded codebase summary fresh after every commit. Closes the staleness loop: graphify rebuilds the graph on commit, the shim immediately refreshes the summary file Claude sees on every turn.
Changed
  • The auto-loaded codebase summary now addresses the orchestrator directly. Previous wording read as delegation-gated and steered the main thread away from using graphify itself. The new copy is a direct project rule: query graphify before exploring 3+ unfamiliar files, local tool, no token cost.
Fixed
  • Settings to "Refresh code graph" now refreshes the summary file too. Previously the manual refresh button rebuilt the graph but left the auto-loaded summary stale until the next full wizard or analyze run.
  • Wizard-init scan path now installs graphify's git hooks for parity with the analyze path. Without this, a project initialized via the wizard never got post-commit auto-rebuild.
v1.2.3-beta Documentation Enforcement gains a procedural runbook April 2026
Added
  • Documentation Enforcement now scaffolds a step-by-step runbook. Enabling the toggle drops a third doc-bootstrap file alongside the index and standards: documentation-runbook.md, a five-step procedure (blast radius, update vs. create, write, index update, stamp) with a concrete exit condition on every step. Existing users see no overwrites; the runbook appears on next re-apply or fresh enable.
Changed
  • The always-loaded rules file is now a thin pointer layer. Methodology no longer lives in the rules file; it points at the index, the standards (WHAT), and the runbook (HOW), plus the load-bearing rule that topic content lives in your docs folder, never in CLAUDE.md. Pre-commit checklist consolidated into the runbook so there is exactly one canonical version.
  • The post-write hook nudge now references the runbook by path. When Claude finishes writing a structural change, the advisory points at the runbook procedure rather than asking it to "update or create a doc."
v1.2.2-beta Shared context imports healed April 2026
Fixed
  • CLAUDE.md managed block now references shared resources that actually exist. The block was hard-coding a reference to a file Foundry never writes, so every freshly-init'd project shipped with a broken import that Claude Code silently skipped. The @-ref list is now derived from the resource manifest, so it always matches what is on disk.
  • Shared-resource writes now auto-refresh the managed block. Adding a new shared resource via any Foundry init or agent path immediately updates the @-refs in CLAUDE.md without waiting for the next Save Team.
  • Empty-roster projects keep their shared @-refs. A project with shared context docs but no agents now correctly retains those imports instead of stripping the entire block.
Added
  • App-startup project sweep. On launch, Foundry rebuilds the CLAUDE.md managed block for every known project from its current resource manifest. Heals any project that existed before the auto-refresh wiring landed.
v1.2.1-beta Auto-updater fix April 2026
Fixed
  • Auto-updater now reliably detects and delivers updates. A channel-detection bug in electron-updater caused the updater to look for a beta.yml manifest that was never published, silently reporting "no update available" on every check. This has been present since the first beta. The fix pins the channel to latest.yml and enables prerelease-to-prerelease update delivery.
v1.2.0-beta Cleaner reset and full init visibility April 2026
Added
  • Foundry self-introduces at init. A context note is written into Claude Code's per-project memory directory so agents working in a Foundry-managed repo immediately know which .claude/ files are machine-local and which are committed source. Stops agents from trying to commit runtime files.
Fixed
  • Existing agents show up in the Forge during init. Agents already on disk in .claude/agents/ now appear in the agents bucket alongside Scholar's proposals, tagged "existing" and deduped by name. Previously an on-disk agent could go invisible during init.
  • Reset and uninstall leave zero trace. The # >>> foundry >>> fence is stripped from .gitignore and Foundry's self-introduction is removed from Claude Code's per-project memory. CRLF-safe on Windows.
v1.1.1-beta Documentation Enforcement: modular docs scaffold April 2026
Changed
  • Documentation Enforcement scaffolds a real docs structure on enable. The toggle now writes two starter files into your docs folder if absent: a skeleton README.md index and a full documentation-standards.md methodology doc. The .claude/rules/foundry-doc-currency.md rules file is now a thin pointer (~30 lines) instead of a self-contained reference. You own the starter docs after first write; Foundry never overwrites or removes them.
  • Existing users: the new shape applies on next toggle re-apply. Resave the toggle in Foundry Settings to pick up the scaffold.
v1.1.0-beta Cost-aware hooks, install-time linting, zero-hook fresh installs April 2026
New
  • Firing-rate badges. Every hook starter template carries a cost-profile badge (session-start, per-tool-call, etc.) so you see how chatty a hook is before installing.
  • Static-text linter. When a PreToolUse or PostToolUse hook just echoes constant text, the editor surfaces a soft-warn banner suggesting SessionStart instead.
  • Hook prerequisite probe. Templates can declare required files; if any are missing on disk, the install flow warns before proceeding.
  • Stale toggle ID detection. Global Settings flags behavior toggles removed in a Foundry update so you can clean them up.
Changed
  • New hooks default to SessionStart. Cheapest-fires-first by default; PreToolUse remains available for gating patterns.
  • doc-currency moved to SessionStart. Previously fired on every read-type tool call; now fires once per session.
  • Standing Order Gate installs conditionally. Empty teams now write zero hooks; removing agents cleans up the obsolete registrations.
Fixed
  • Fresh installs no longer ship redundant hooks. Prior versions auto-installed 4 system hooks + 3 helper scripts at every init regardless of agent state.
  • Settings drift on agent removal. Old hook registrations no longer persist in settings.json pointing at deleted scripts.
v1.0.1-beta Patch: reset hygiene, Electron data bundling, Scholar UI polish April 2026
Fixed
  • Reset cleanup is complete. Reset Project now also clears foundry-baselines.json, so baseline state doesn't drift across a reset.
  • Behavior cleanup strips everything. Disabling a behavior toggle now removes all four artifact types (settings.json hook entries, settings keys + sidecar, and .claude/hooks/foundry-*.mjs files).
  • Packaged builds load all data. The Electron installer now bundles data/ runtime files into the standalone server, fixing bucket registries that silently failed to load.
  • Scholar UI fills the column. The collapsed FAQ rail uses a single icon and fills its height; the chat panel now fills the full viewport.
  • Doc-currency behavior, three small fixes. Custom stamp templates are validated, stdin handling is fixed, and .mjs hook scripts deploy correctly.
Changed
  • Auto-updater follows the org. Release publishing and the in-app updater now point at try-foundry/tryfoundry-releases.
Removed
  • Re-scan project (legacy). The dead Re-scan path is gone from both entry points.
v1.0.0-beta Baselines, templates, and path-scoped behaviors April 2026
Headline Features
  • Per-item baselines. Every Foundry-managed item auto-captures its first deployed state. One click to reset if a Scholar revision goes sideways; one click to update when you settle on a new good state.
  • Templates everywhere. Every bucket (hooks, agents, triggers, skills, rules, standards) now has a Pick from library modal with curated starters, plus an inline pill row in each editor for fast inserts.
  • Path-scoped behaviors. All 29 behavior toggles support file glob scoping. A "Write tests in TDD style" rule fires only when Claude touches test files, keeping prompt payloads lean for focused work.
Kebab menu on a Foundry-managed hook showing Set baseline and Reset to baseline actions
Pick from library modal in the hooks bucket showing nine curated starter templates
Quality and Reliability
  • Reliable persistence. Standards and rules save correctly across both the wizard and disk-mode editor.
  • Stable identities. Hooks and triggers carry stable IDs so baselines survive a reload.
  • Atomic writes. The .claude/ tree writes atomically; no more half-written files on crash.
  • Failure transparency. When Scholar generation fails, you see why instead of a silent fallback.
  • Browseable reasoning. Mid-flow Scholar reasoning is accessible via a dedicated modal.
  • Better diffs. Agent diffs now surface model, effort, and disallowed-tools changes that previously applied invisibly.
v0.9.0 Forge editor, 2x2 taxonomy, and interview overhaul March 2026
New
  • Forge editor. A persistent dashboard for modifying your team config after initial setup. Every bucket editable in one place: hooks, agents, behaviors, triggers, skills, MCPs, and project context.
  • 2x2 agent taxonomy. Agents now carry a role (Specialist or Guardian) and a kind (Worker or Scholar). Shown on every roster card and in every generated file.
  • Interview overhaul. The onboarding interview is now a real conversation: 8 to 12 focused questions, developer-skill pills, and a natural back-and-forth instead of a form.
  • Domain exclusions. Per-domain rules for where agents operate. Keeps test-only agents away from production paths.
Fixes
  • MemPalace bridge deduplication removed dev console error spam.
  • Wizard post-creation handoff state corrected after navigation changes.
  • Tooling detection in the scope step now reports specific package names.
v0.8.0 Initial release: wizard, Graphify, and first agents January 2026
First Release
  • Six-step wizard. Analyze, Scope, Interview, Brief, Approve, Install. From empty repo to running team in one pass.
  • Graphify integration. Local codebase graph built on every commit. Agents query it instead of reading raw files, around 71x fewer tokens for structural questions.
  • MemPalace memory. Decisions and conventions accumulate across sessions, searchable by memory-capable agents.
  • Behavior toggles. Initial set of 20 toggles covering commit hygiene, planning gates, documentation checks, and standing-order controls.
  • First agent roster. Code Reviewer, Security Guardian, Architecture Historian, Compliance Auditor, and Test Runner. All generated from your interview with role and kind already set.
Forge

Every surface. One editor.

After Scholar builds your team, Forge is where you manage it: agents, hooks, rules, standards, behaviors, triggers, MCPs, skills. Most buckets live-sync to your .claude/ directory; skills and MCPs are read-only in this beta.

Forge editor showing the Agents bucket with 5 generated agents and DETECTED and FOUNDRY badges
Behaviors bucket with 5 of 29 toggles enabled, Living-Doc Workflow expanded showing path-scope
Behaviors: 29 toggles. Enable globally, per-project, or scoped to specific file globs.
Standards bucket with 3 compliance standards and DETECTED and FOUNDRY badges
Standards: compliance frameworks stored as one file each, with full traceability.
Triggers bucket with 3 triggers: src-test-runner, pre-commit-security, post-build-audit
Triggers: file watchers, git hooks, and npm scripts that run agents outside sessions.
Skills bucket with 3 skills: graphify, migration-reviewer, stripe-helper
Skills: slash commands and reusable workflows available to every agent.
MCPs bucket with 5 MCP servers: postgres, github, stripe, code-graph, memory-palace
MCPs: project-scoped server connections that extend what agents can access.
Domain Exclusions bucket with 5 deny patterns blocking rm -rf, DROP TABLE, .env writes
Domain Exclusions: tool-call patterns agents are permanently blocked from running.
Download Foundry 1.4 beta
Free during beta · Windows 10/11 · Bring your own Anthropic API key
Foundry Scholar AI
Meet your Scholar

Scholar reads your project, then proposes the team.

Foundry's Scholar AI skips templates entirely. It interviews you about your codebase, scans your repo locally via Graphify, and writes custom agent prompts matched to how your project actually works. Every suggestion comes with visible reasoning.

The problem

Configuring Claude Code is tedious.

Hand-writing agent .md files, behavior rules, hook scripts, and standing orders takes hours of YAML, copy-paste, and trial-and-error before you've shipped anything.

The answer

Scholar AI designs the team.

Foundry analyzes your codebase locally via Graphify, interviews you for ten minutes, and proposes a complete config: agents, hooks, rules, and standards, explaining every choice.

The result

You approve. It ships.

Nothing touches .claude/ until you say so. Foundry-managed items get baselines you can reset from any item card. The whole config is plain Claude Code. Edit it forever, inside or outside of Foundry.

What it builds

Real Claude Code config, generated for your project.

Foundry's output isn't a black box. It's standard .claude/. Every file is plain markdown or JSON you can read, edit, and version-control yourself.

Scholar collects your context first
Project Context panel showing interview answers: developer experience, project clarity, communication style, caution level, and compliance frameworks
Developer experience, project clarity, communication style, caution level. Everything Scholar learned is stored in .claude/foundry-context.json and browseable any time.
.claude/{rules, standards, hooks}
rules/
├─ api-design.md (scope: src/api/**)
├─ test-conventions.md
└─ commit-style.md

standards/
├─ soc2-controls.md
└─ no-hardcoded-secrets.md

settings.json hooks:
├─ PreToolUse/Bash → block force-push
├─ PostToolUse/Edit → run prettier
└─ Stop → desktop notification
Path-scoped rules, file-per-standard storage, and one-click hook templates. Every artifact traces back to where it came from.
Solo and team work

Your compliance docs become enforceable rules.

Upload your security policies, coding standards, or regulatory frameworks. Scholar extracts the rules that matter, wires them into agents, hooks, and permissions, and stamps full traceability from agent to source line. SOC 2 ships with a curated starter template; for HIPAA, PCI-DSS, GDPR, ISO 27001, and any other framework, Scholar extracts enforceable controls from the policy docs you upload.

Team / Enterprise

Shared codebases · Compliance review · Policy enforcement

Upload your security policies, coding standards, or regulatory framework. Scholar extracts enforceable rules and wires them into agents, hooks, and permissions. Every artifact traces back to the source doc that required it. No audit-time surprises.

  • Standards & policy extraction from your own docs
  • SOC 2 starter template included; other frameworks extracted from your own policy docs
  • Full traceability: agent → rule → source line
  • Portable team blueprints. Share across projects, onboard new devs

Solo / Personal

Personal projects · Freelance · Learning

Skip the team controls and Foundry runs as a streamlined wizard tuned for solo workflows. Same Scholar, same Forge, just with the compliance scaffolding stripped away.

  • No compliance overhead. Scholar focuses on velocity and common mistakes
  • Behavior presets tuned for solo development
  • Same local Graphify knowledge graph and MemPalace memory
  • Scholar chat available throughout the Forge to refine any proposal
Foundations

What's under the hood.

The free building blocks and the intelligence that ties them together.

◆Scholar, your AI configurator

Scholar reads your code, conducts the interview, then works alongside you in the Forge editor. Ask it to add, revise, or explain any proposal. Every suggestion comes with a justification.

⎔Graphify knowledge graph

A local semantic graph of your codebase, updated on every commit. Agents query it instead of reading raw files, using ~71× fewer tokens. Free forever, runs entirely offline.

❋MemPalace memory

Architectural decisions, conventions, and pain points accumulate across sessions. The Scholar files new knowledge into MemPalace. Agents with the memory-palace capability can search it across sessions.

▣Behavior presets & toggles

Plan-before-acting, conventional commits, doc-currency checks, standing-order gates. Each toggle includes an explanation and example so you know exactly what you're enabling. Presets bundle sensible defaults; per-toggle overrides let you fine-tune.

◎2×2 agent taxonomy

Every agent has a role (Specialist or Guardian) and a kind (Worker or Scholar). Consistent mental model across the roster, visible in every surface.

▶Open architecture

Everything Foundry generates is standard Claude Code configuration: plain .md files, hooks, settings.json. No lock-in. Nothing touches disk until you approve it.

The flow

How it works.

From clean repo to running team in a single sitting. Graphify scan time depends on codebase size.

1
Analyze
Foundry scans your codebase locally via Graphify. No tokens spent, no data leaves your machine.
2
Scope
Solo or team? Guided or expert? That shapes every step and what Scholar proposes.
3
Interview
Scholar asks focused questions about your workflow and pain points. Nothing generic.
4
Forge
A collaborative workspace where you chat with Scholar to shape agents, behaviors, hooks, triggers, and skills before anything gets written.
5
Approve
One final confirmation. Nothing is written to .claude/ until you say so.
6
Install
Foundry writes your team configuration. Every artifact gets a baseline restore-point automatically.
Prefer to drive everything yourself? Skip the wizard and build your team by hand in Forge. Same edit surface, same install pipeline, no Scholar in the loop.
Download Foundry 1.4 beta
Free during beta · Windows 10/11 · Bring your own Anthropic API key

Foundry is free while in beta.

Free to download and use during the beta period. Each build is time-gated to 30 days, with the next one available before the previous one expires. If you find Foundry useful, donations on Ko-fi help fund development, and beta donors will get credit toward the purchase price at the eventual 1.0 stable launch.

Download Foundry 1.4 beta
Free during beta · Windows 10/11 · Bring your own Anthropic API key