MyAgentStudio — Agent Workbench

Stop editing AI agents blind. MyAgentStudio keeps a structured, always-visible view of your agent next to an agent-aware AI chat — the chat proposes changes, you see exactly what would change side-by-side with what's there today, and nothing is written until you click Apply.

• Structured, not blind — Role, Behavior, Guardrails, Output (and any custom sections) stay visible on screen at all times, not buried in a chat transcript.
• AI proposes, you approve — every chat-suggested change renders as a diff card; nothing lands until you review it and click Apply.
• Lossless import — anything the importer can't confidently classify becomes its own visible custom section, never silently dropped.
• Beyond the browser — an MCP server lets Claude Code (or any MCP client) list, pull, and push your own agents from the terminal.

🔗 Try it live — invite-gated beta; no code yet? use the Request access form on the landing page. Full source below, AGPL-3.0 — operated by ProcessMind Solutions.

See it in action

Import → chat edit → proposal card → apply → raw preview, all in one pass.

A reviewed proposal, applied

---

A local-first workbench for building and managing AI agents. An agent-aware AI chat sits next to an always-visible structured view of the agent it is editing, so you never edit blind. The platform targets Claude Code's .claude/agents/*.md subagent format and runs entirely on your machine using your own Anthropic API key — or use the hosted beta above with no setup at all.

Quick start

npm install

Create a .env.local file in the project root and add your key:

ANTHROPIC_API_KEY=sk-ant-...

# Auth — required (see "Auth setup" below)
JWT_SECRET=<at-least-32-random-characters>

# Only needed for the one-time bootstrap command (see below)
BOOTSTRAP_USER_EMAIL=you@example.com
BOOTSTRAP_USER_PASSWORD=<your-admin-password>

# Session lifetime — optional, defaults to 7 days (604800 seconds). Bounds: 60–7776000
# seconds (60s–90d). An invalid value refuses to start the process rather than silently
# falling back to the default.
SESSION_TTL_SECONDS=604800

# Google sign-in — optional. All three must be set together, or none at all; a partial
# set refuses to start. See "Google OAuth setup" below.
GOOGLE_CLIENT_ID=
GOOGLE_CLIENT_SECRET=
OAUTH_REDIRECT_BASE_URL=http://localhost:3000

Optionally override the default model (defaults to claude-opus-4-8):

ANTHROPIC_MODEL=claude-sonnet-4-6

To enable a second LLM provider (OpenAI-compatible, e.g. NVIDIA NIM, OpenAI, Groq):

OPENAI_COMPATIBLE_API_KEY=<your-key>
OPENAI_COMPATIBLE_BASE_URL=https://integrate.api.nvidia.com/v1
OPENAI_COMPATIBLE_MODEL=meta/llama-3.1-8b-instruct

All three vars must be set together to enable the provider. Switch between providers from System Settings (admin only) — no restart needed. See .env.example for more vendor examples and notes on which NVIDIA NIM models are actually callable on a free-tier key (not every model listed in NVIDIA's catalog is — see .env.example).

Start the dev server:

npm run dev

Open http://localhost:3000.

Auth setup (first run / new database)

The app requires a login session. On a fresh database, run the migration and bootstrap the admin account:

# 1. Run the migration + seed (also runs automatically on every npm run dev)
npm run db:seed

# 2. Set the admin's email and password (run once; re-run with --force to change credentials)
BOOTSTRAP_USER_EMAIL=you@example.com BOOTSTRAP_USER_PASSWORD=yourpassword npm run auth:bootstrap

After that, go to http://localhost:3000/login, sign in as the admin, and generate invite codes from System Settings to let others sign up.

HTTPS required for deployed instances. The session cookie is secure: true in production, which means it is only sent over HTTPS. An http:// deployment will silently drop the cookie and every request will appear unauthenticated.

Session lifetime

SESSION_TTL_SECONDS changes the lifetime baked into newly issued tokens only — it is not revocation. A JWT's expiry is set at signing time, so shortening the TTL does nothing to anyone already signed in; they keep their original lifetime until it expires on its own. If a session must be killed right now, the only immediate options are deleting or altering that user's row in myagent.db (getSession() re-reads the DB on every request), or rotating JWT_SECRET, which invalidates every session at once.

Google OAuth setup (optional)

MyAgentStudio can offer "Continue with Google" alongside password sign-in. It is entirely optional — leave the three GOOGLE_*/OAUTH_REDIRECT_BASE_URL variables unset and the button never renders. Signing in with Google still requires a valid invite code on first use; it is a second way to prove who you are, not a second way to get admitted.

To enable it:

1. Create an OAuth 2.0 client in the Google Cloud console (Web application type).
2. Register this exact redirect URI, built from OAUTH_REDIRECT_BASE_URL (no trailing slash, no path):

   <OAUTH_REDIRECT_BASE_URL>/api/auth/oauth/google/callback
   

3. Set GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET, and OAUTH_REDIRECT_BASE_URL in .env.local. All three are required together — setting only some of them refuses to start the process rather than half-working.
4. In production, OAUTH_REDIRECT_BASE_URL must be https:// (Google rejects non-HTTPS redirect URIs except for localhost).
5. For a closed beta, keep the Google Cloud consent screen in Testing mode and list your invite-gated users as test users — this restricts who can even reach the consent screen, on top of the invite-code gate on the MyAgentStudio side.

Google is told nothing beyond the standard OpenID Connect openid email profile scopes — no ongoing access, no refresh token is requested, and nothing Google issues is ever stored.

Email delivery setup (optional)

MyAgentStudio can email an invite code to whoever requested access, instead of the admin copying and sending it by hand. It's entirely optional — leave RESEND_API_KEY/EMAIL_FROM/APP_BASE_URL unset and codes are generated and copied exactly as before email existed.

To enable it:

1. Create a free account at resend.com, add your sending domain, and publish the SPF/DKIM (and recommended DMARC) DNS records it gives you. Disable open/click tracking in the dashboard — this app sends transactional email only.
2. Create an API key scoped to sending only.
3. Set RESEND_API_KEY, EMAIL_FROM, and APP_BASE_URL in .env.local. All three are required together — see .env.example for the exact shape and the two independently-optional EMAIL_REPLY_TO/ADMIN_NOTIFICATION_EMAIL vars.

A failed or not-yet-configured send never blocks the underlying action — the invite code is always created and shown, with a status line and a one-click resend for the fallback path.

Settings

Configurable from System Settings (/settings, admin only):

Setting	Default	Notes
liveLlmCalls	true	When off, every AI call is blocked and logged with no network traffic.
llmProvider	anthropic	Which vendor answers every AI call. anthropic uses ANTHROPIC_API_KEY; openaiCompatible uses OPENAI_COMPATIBLE_API_KEY + OPENAI_COMPATIBLE_BASE_URL. A provider with no key configured cannot be selected. Switching takes effect on the next call — no restart.
maxUsers	5	Maximum number of accounts (including the admin). Lowering it below the current count blocks new signups but does not remove anyone.
maxLlmCallsPerUserPerHour	15	Per-user hourly LLM call cap (rolling 60-minute window). The admin is always exempt.
chatMaxTokens	8192	Max tokens Prometheus may generate per chat reply.
chatHistoryTurns	10	How many prior chat messages Prometheus sees for context.
mcpWrites	false	When on, write-scoped MCP tokens can call push_agent (see "Console MCP access" below). Off by default — a deployment that never touches this setting behaves exactly as if the MCP server didn't exist.
liveEmailSends	true	When off, every outbound email is blocked and logged before any network request is made. Only matters once email is configured at all — see "Email delivery setup" below.
maxEmailsPerHour	50	Deployment-wide cap on outbound emails per rolling 60-minute window, counting only attempts that reached the provider.

The workbench layout

The UI is a four-pane IDE layout:

• Library (left, foldable) — your agent list, plus a "Shared with me" section for agents another user has granted you read-only access to. Import files, or redeem a share code, from the action bar at the bottom. (Grouping agents into collapsible groups is built underneath but not yet enabled in this release.)
• Custom Visualization (center top) — the currently selected agent split into its named sections (Role, Behavior, Guardrails, Output, and any optional or custom sections). Each section can be expanded, collapsed, or edited in place — unless the agent was shared with you, in which case this pane is read-only with a Copy-to-me action.
• AI Chat (center bottom) — type an instruction and ✦ Prometheus proposes a change to the description, sections, and/or config. Review the proposal card and click Apply (or Discard) — nothing lands until you do.
• Raw / Share (right, foldable) — a two-tab dock: Raw is a live read-only export preview of the agent as it would be written to a .md file, updated after every save; Share (owner only) is where you enable a share link, add people by email, and revoke access.

All four panels resize via drag gutters. Library and the right dock fold to rails to reclaim space.

Console MCP access

A console/CLI MCP client (Claude Code and equivalents — not Claude Desktop's GUI connector) can list, read, pull, and push your own agents over POST /api/mcp, authenticated by a Personal Access Token you generate in Account. See docs/user-guide.md for setup and docs/system-about.md §13 for the design.

Documentation

• User guide — how to import, edit, organize, and export agents.
• Project explanation — the product story: the problem, who it's for, how it works, how it was built.
• System About — the engineering reference: stack, architecture, data model, design principles.
• Roadmap — what's available today, coming next, and planned.

License

MyAgentStudio is released under the GNU Affero General Public License v3.0 (AGPL-3.0). Commercial licensing is available on request for anyone whose use case doesn't fit the AGPL's terms.

Copyright (C) 2026 ProcessMind Solutions.

Contributing

Issues and small pull requests are welcome. Larger contributions need a prior discussion first — as sole copyright holder, that's what keeps the commercial-licensing option above meaningful.