OpenExecutive

AI-powered virtual executive team — a single coherent executive persona backed by 8 specialist agents (FastAPI + Next.js).

Automatically checked · GitHub checked 2026-11-17

OpenExecutive

Overview

Open Executive

CI License: Apache 2.0 Python 3.11+ Next.js 16

Open Executive is designed to transform leadership and management. Highly configurable, it can be deployed at any management level. Out of the box it supports approval levels, spend approval ranges, integration with corporate knowledge systems, and defined governance for how it interacts with human colleagues and other AI systems. Open Executive can also act as a busy leader's digital twin: with As you switched on, it drafts email in their own voice, from their own Gmail or Outlook, and sends only what their Approvals let go on its own. The rest waits for them to tap Send.

It meets people where they already work, with integrations for Slack, Discord, Telegram, Google Chat, email, Google Workspace, Microsoft 365 (Outlook mail and calendar, OneDrive), Notion, Confluence and any MCP-compatible AI tool, with more on the roadmap. Open Executive will always be open source.

A managed cloud offering is coming (https://openexecutive.ai), where you can get access without deploying anything yourself.

Demo

Open Executive demo video

A walkthrough of Open Executive in action — watch on YouTube.

What It Does

Developed by ForeAct Labs, Open Executive provides a single coherent executive voice backed by nine specialist AI agents, plus a triage agent that sorts incoming work:

The Executive at the center, with nine specialists around it: strategy, finance, people, legal, operations, marketing, product, sales and board

  • Chief Strategy Officer — competitive analysis, M&A, market positioning, OKRs
  • Chief Financial Officer — financial modeling, fundraising, unit economics, cash flow
  • Chief HR/People Officer — hiring, compensation, performance, culture
  • General Counsel — contracts, IP, employment law basics, compliance
  • Chief Operating Officer — process design, vendor management, operational scaling
  • Chief Marketing Officer — GTM strategy, brand, communications, PR
  • Chief Product Officer — roadmap, prioritization, product strategy
  • Head of Sales — pipeline and qualification, founder-led sales, pricing conversations, proposals, forecasting
  • Board Communications Director — board decks, investor relations, governance

All responses come from one consistent executive voice. The internal agent architecture is never exposed to the user. Beyond Q&A, the system maintains episodic memory of past decisions and initiatives across sessions, and a built-in scheduler can proactively surface follow-ups and time-sensitive actions.

What you can let it do

  • Home opens with your brief: what needs you, and Done for you, what it did for you in the last day.
  • As you drafts replies to your inbox in your own voice, from your own mailbox. Each teammate can turn it on for their own mail, which stays private to them.
  • As me lets it act in its own name: book a meeting, follow up with someone, update a sheet, answer a customer. Every step is listed.
  • Approvals decide what goes on its own and what waits for your yes, the same on every channel. Pick Careful, Balanced or Hands-off, add Your rules ("always wait" or "goes on its own" by amount, person or company), and keep a feature In training while you check its work. Anything that waits comes to you as an Approve card. As you, As me and Approvals sit together on the Delegate page.
  • Departments, People and Goals tell it who decides what. A cost goes to the person whose approval range covers it.
  • Tasks, Workflows, Documents and the Watch list hold the work it is doing, the jobs it runs on a schedule, what it delivered, and what it keeps an eye on.

The Executive can also use your web browser to get things done on websites for you. With the Open Executive extension in Chrome, it looks at the page you're on, clicks, types, fills in forms and opens pages until the task is done: "sort this search by best sellers and list the top five", "fill in this vendor form from the contract". It follows your Approvals, never types passwords or card numbers, hands a sign-in or a code back to you, and stops when you press Stop. Saved workflows can use it too. See docs/browser_extension_setup.md.

Architecture

User message
    ↓
Executive Orchestrator
    ↓ tool use → parallel specialist calls
CSO / CFO / CHRO / GC / COO / CMO / CPO / Sales / Board
    ↓ each specialist retrieves relevant context from ChromaDB
Built-in MBA knowledge + Your company documents
    ↓
Synthesized executive response

Knowledge — Two retrieval layers per specialist call: (1) built-in MBA-level Markdown (packages/core/openexecutive/knowledge/builtin/, git-tracked) seeded into ChromaDB at startup, and (2) your uploaded company documents chunked and stored in a separate company_docs collection. RAG context is injected into the user turn, never the cached system prompt.

Episodic memory — After every response, a background claude-haiku-5-5 pass extracts key decisions, initiatives, and advice into SQLite. The next session opens with a <past_decisions> block so the Executive remembers what it recommended last month.

Always in the loop — Each person can turn on "Keep track of what happens" (Settings → Memory). The Executive then keeps private, dated notes of what they said, in chat where it can confirm it's them (the web app signed in, their own Slack or Discord, Telegram with a webhook secret) and in As you replies they send. Every note rests on their own words, checked word for word. Only they see it, in Memories → History, where they can correct, pin or forget it. Notes are read back only to them, in a private chat, and expire after 90 days by default (the owner can change this). The owner's morning brief and evening digest include what the owner's own notes say is due, and anyone with notes on gets a short reminder on the day something they promised by email is due. It needs no setup and no extra service.

Scheduler — A built-in job runner claims due actions via UPDATE … RETURNING to prevent double-firing. The API must run as a single instance; do not horizontally scale it without gating the scheduler first.

Prompt caching — The system prompt is structured so the Executive persona, company profile, and knowledge index are cached separately (up to 85% cache hit rate after the first few turns). No dynamic content ever goes in a cached block.

See docs/architecture.md for the full design.

Tech Stack

Bring your own model. Open Executive runs on Claude, on any model OpenRouter offers (GPT, Gemini, Llama, DeepSeek, Grok and more), or on models you host yourself with Ollama, LM Studio, vLLM or any OpenAI-compatible server. You can mix them: keep the Executive on one model and move any specialist to another in the Agent Council. Out of the box it uses Claude; see Configuration and Running on Local Models to change that.

LayerChoice
ModelsYour choice, per agent (see above)
BackendPython 3.11 + FastAPI
Package manageruv
Vector storeChromaDB (local, embedded)
Episodic memorySQLite
Web UINext.js 16 (App Router) + Tailwind
LicenseApache 2.0

Repo Layout

openexecutive/
├── packages/
│   ├── core/
│   │   └── openexecutive/
│   │       ├── orchestrator/     # Executive persona, routing loop, Approvals
│   │       ├── agents/           # Specialist agents + triage
│   │       ├── knowledge/        # ChromaDB store + RAG pipeline + built-in knowledge
│   │       ├── memory/           # Company profile + episodic memory
│   │       ├── people/           # The People list: who the Executive can reach and wait on
│   │       ├── departments/      # Department state, goals and memory
│   │       ├── delegation/       # As you: writing from a person's own mailbox
│   │       ├── briefing/         # The morning brief and Home
│   │       ├── onboarding/       # Wizard state machine + profile builder
│   │       ├── prompts/          # Persona + domain prompts + cache manager
│   │       ├── providers/        # Claude, OpenRouter and local model backends
│   │       ├── api/              # FastAPI app + routes
│   │       ├── mcp_server/       # Open Executive as an MCP server
│   │       ├── integrations/     # Slack, Email (Gmail or Outlook), Telegram, Google Chat, Discord
│   │       ├── scheduler/        # Background job runner (single-instance)
│   │       ├── alerts/           # Proactive alert system
│   │       ├── monitoring/       # Watching status pages, feeds and prices
│   │       ├── audit/            # Audit logging
│   │       ├── architecture/     # Content for the /architecture page
│   │       ├── guide/            # Content for the User Guide page
│   │       ├── workflows/        # Multi-step workflow definitions
│   │       ├── browser/          # Using a person's own browser via the extension
│   │       ├── evals/            # Eval scenarios (_scenarios/) + judges
│   │       └── cli.py            # Click CLI
│   ├── ui/                       # Next.js 16 web UI
│   └── extension/                # Chrome extension (load unpacked or build your own)
├── evals/                        # LLM-as-judge runner
├── fixtures/                     # Demo company fixtures (profiles, docs, rosters)
├── scripts/                      # Operator scripts (mailbox sign-in, caller keys, PR checks)
├── docker/                       # Dockerfile(s) + docker-compose.yml
└── docs/                         # Architecture + deployment docs

Quick Start

# Clone the repo
git clone https://github.com/ForeactLabs/OpenExecutive.git
cd OpenExecutive

# Install the backend and web app dependencies
make install

# Set your Anthropic API key
cp .env.example .env
# Edit .env and add ANTHROPIC_API_KEY=sk-ant-...

# Start everything
make dev

Trying it on your own computer needs no sign-in setup. While AUTH_GOOGLE_ID and AUTH_OIDC_ISSUER are blank, make dev uses local login: the web app only accepts connections from this computer, you click Open, and you are the owner. To invite your team, or to run it on a server or with make docker, set up Google or SSO sign-in by filling in the AUTH_* block (see docs/auth.md for the Google Cloud Console and SSO steps).

All configuration lives in that repo-root .env — make dev and make docker both load it for the API and the UI (Auth.js needs AUTH_SECRET / AUTH_GOOGLE_ID / AUTH_GOOGLE_SECRET at runtime). A packages/ui/.env.local is also read for UI-only keys, but for keys present in both files the root .env takes precedence.

Open http://localhost:3000 to start chatting with your executive. The API runs on port 8000 and the UI on 3000.

To check everything is working, open Settings → Setup status. It tests the AI key and each chat app's token, then shows a green, amber or red light for each part of your setup, with what to do about anything that isn't green.

First run: requires Python 3.11+ and Node 22.6+. The initial uv sync pulls heavy ML dependencies (ChromaDB + sentence-transformers/PyTorch), and the first boot downloads a small embedding model (~90 MB) to build the local vector index — so the first make dev takes a few minutes before the app is ready. Subsequent starts are fast.

On Windows: run make from Git Bash or WSL, not PowerShell or cmd. The recipes are POSIX shell (if [ -f .env ]; …), and GNU Make falls back to cmd.exe when no sh is on PATH — which fails with -f was unexpected at this time. If Make still picks the wrong shell, point it at one: make dev SHELL="C:/Program Files/Git/bin/sh.exe". Note that make stop uses lsof and has no Windows equivalent; stop the two dev servers from their own terminals instead.

For contributors not using make:

cd packages/core
uv sync
uv run uvicorn openexecutive.api.main:app --reload --port 8000

# In a second terminal
cd packages/ui
npm install
npm run dev

uv run executes inside the project's virtualenv without activating it, so these commands are the same on macOS, Linux and Windows. (Activating manually works too, but the path differs per platform: .venv/bin/activate on macOS/Linux, .venv\Scripts\Activate.ps1 on Windows.) The packages/ui commands are listed one per line rather than chained with && so that they run in every shell too: Windows PowerShell 5.1, the version that ships with Windows, has no && operator and rejects the chained form with The token '&&' is not a valid statement separator in this version.

Set Up Integrations

Every integration is optional and off until you configure it. Each one has a setup guide:

IntegrationWhat it doesSetup
SlackAnswers mentions and DMsSet SLACK_BOT_TOKEN and SLACK_APP_TOKEN (socket mode), or SLACK_SIGNING_SECRET to take events over HTTP at /webhook/slack
DiscordAnswers DMs, mentions and slash commandsRun the Discord Bot
TelegramAnswers messages to your botdocs/telegram_setup.md
Google ChatAnswers mentions in a spacedocs/google_chat_setup.md
Email and calendar (Google Workspace)Reads and sends the Executive's Gmail, books on its calendarGoogle Workspace credentials
Email and calendar (Microsoft 365)The same with Outlook mail and calendarMicrosoft 365 credentials
As youDrafts replies in each person's own Gmail or Outlook, sent as their Approvals allowYour own Gmail or Outlook
Google Drive foldersSyncs shared folders into the knowledge base (read-only)docs/drive_sync_setup.md
OneDrive foldersSyncs chosen OneDrive or SharePoint folders into the knowledge base, as the Executive's Microsoft 365 sign-indocs/onedrive_sync_setup.md
OneDrive in chatSearch, read, upload, move, copy and share OneDrive files (no delete, no anonymous links)Set MS365_MCP_ONEDRIVE=true and sign in again (docs/onedrive_sync_setup.md)
ConfluenceSyncs chosen spaces into the knowledge base (read-only, Cloud or Server/Data Center)docs/confluence_sync_setup.md
NotionSyncs the pages shared with an internal integration into the knowledge baseSet NOTION_SYNC_ENABLED=true and NOTION_API_KEY (see .env.example)
MCP toolsGives the Executive tools from other MCP servers, with deny rules for destructive onesCopy packages/core/mcp_servers.json.example to company/mcp_servers.json and list your servers under mcpServers; MCP turns on when the file exists (Docker keeps it on /data, see docs/deployment.md)
MCP clientsLets any MCP client talk to the ExecutiveConnect as an MCP Server
BrowserUses each person's own Chrome to do tasks on websites, through the Open Executive extension (on by default)docs/browser_extension_setup.md

Synced Drive, OneDrive, Confluence and Notion content is kept apart from your uploaded documents and labelled as unreviewed, so it ranks below them.

Web search for any model

ENABLE_WEB_SEARCH uses Anthropic's server-side web_search tool, so only Claude models get it. Any model that can use tools, local ones included, can search the web through a search MCP server under MCP tools instead. For example, Keenable needs no account or API key. Add it under mcpServers in company/mcp_servers.json (Docker: /data/company/mcp_servers.json):

"keenable": {
  "url": "https://api.keenable.ai/mcp?keenable_title=openexecutive"
}

Restart the API. The Executive finds keenable__search_web_pages and keenable__fetch_page_content with search_tools and runs them with call_tool, like every other MCP tool.

Run the Discord Bot

  1. Create a Discord application at https://discord.com/developers/applications
  2. Enable the Message Content privileged intent (Bot → Privileged Gateway Intents)
  3. Invite the bot with bot + applications.commands scopes
  4. Set env vars in .env: DISCORD_BOT_TOKEN, DISCORD_APP_ID, DISCORD_GUILD_IDS
  5. Run the API normally — the bot starts as part of the FastAPI lifespan when DISCORD_BOT_TOKEN is set:
make dev

The bot is embedded in the API process (alongside the email poller, scheduler, and resumer) so it shares the same SQLite database and ChromaDB vector store under /data in production. Skip the token to disable.

For iterating on bot-only code without restarting the API, make discord runs the bot as a standalone process against the same local DB. Use it instead of the embedded bot, not alongside it: stop the API, or start it without DISCORD_BOT_TOKEN, first — two processes on one token each open a gateway connection and every message gets answered twice.

Users can DM the bot, @mention it in a channel (replies in a thread), or use /ask and /today slash commands. Slash commands sync to DISCORD_GUILD_IDS instantly on startup; leave blank for global registration (up to 1-hour propagation delay).

Deploying to production

The bot runs inside the existing API process — no extra service. Set these on the API and restart it:

DISCORD_BOT_TOKEN=...
DISCORD_APP_ID=...
DISCORD_GUILD_IDS=...

Discord user access is managed via the /people UI — add a Person row with discord_user_id set.

The bot starts on the next lifespan boot. To disable it, unset DISCORD_BOT_TOKEN and restart.

Onboarding Your Company

The first time you visit the app, you'll be guided through a wizard to set up your company profile:

  • Company basics (name, industry, stage, team size)
  • Business model and revenue
  • Competitive landscape
  • Strategic priorities
  • Culture and values
  • Optional: financial position, document upload

After onboarding, the Executive will reference your specific company context in every response.

The built-in knowledge base is trusted by default — the Executive can use it straight away, and the Review queue starts empty rather than asking you to sign off on ~100 shipped documents. Items reach the queue when you upload a document, edit an existing file, or deliberately send a domain for review from the Review page. While a domain is being reviewed its content is withheld from the Executive, so the UI confirms the count before it starts.

Interfaces

InterfaceHow to Use
Web UIhttp://localhost:3000
SlackMention @OpenExecutive or DM the app
EmailCC or email the configured address (Gmail via the Google Workspace MCP, or Outlook via the Microsoft 365 MCP)
TelegramMessage the configured bot
Google ChatMention the app in a space
DiscordDM the bot, @mention it in a channel, or use /ask / /today slash commands
BrowserAdd the Open Executive extension to Chrome, connect it, then ask in the web chat (setup)
CLIopenexecutive chat
MCPPoint any MCP client at http://localhost:8000/mcp (see Connect as an MCP Server)

Connect as an MCP Server

Open Executive exposes its company context and specialist council to external MCP clients (Claude Code, Claude Desktop, Cursor, or any other agent), so they can ground themselves in your company without being re-briefed. There is no separate server process to start: the server is mounted into the FastAPI app at /mcp over Streamable HTTP, so running the API is running the MCP server.

The endpoint is http://localhost:8000/mcp, or https://<your-host>/mcp once deployed. Authentication is the same shared-secret gate as every other route, so clients send x-api-key: $BACKEND_SHARED_SECRET. With that variable unset locally the gate is off and the header can be omitted.

On any internet-reachable instance set both BACKEND_SHARED_SECRET and OE_PUBLIC_DEPLOYMENT=1, which makes the API refuse to start without a secret rather than serving /mcp unauthenticated. See Deployment.

{
  "mcpServers": {
    "open-executive": {
      "type": "http",
      "url": "http://localhost:8000/mcp",
      "headers": {
        "x-api-key": "YOUR_BACKEND_SHARED_SECRET"
      }
    }
  }
}

Claude Code can write that entry for you:

claude mcp add --transport http open-executive http://localhost:8000/mcp \
  --header "x-api-key: YOUR_BACKEND_SHARED_SECRET"

Connected clients get read-only resources (company profile, today's brief, Needs you cards, Done for you, tasks, watches, Approvals, people, departments, and episodic memory for decisions, initiatives and advice) and six tools: consult_specialist (domain analysis from any of the ten specialists, each grounded in your company's knowledge base), search_knowledge, list_workflows, run_workflow, answer_card (Approve, Skip or reply to a Needs you card) and ask_executive. The full inventory is on the MCP Server section of the /architecture page.

/mcp answers as the caller the API verified, by the web app's rules. With no x-caller-email (open mode) the client is the owner, as the CLI is. A request that carries a signed-in person (a deployment that puts its own sign-in in front of /mcp and stamps the caller as the web UI's proxy does; see signed callers) gets that person's answers. The resources read through the same routes as the web app, so each one shows what that person sees there. With signed callers on, a client holding only the shared secret is a service: never the owner, and its turns ask as no one.

ask_executive, answer_card and run_workflow are the person's own turn, the same as typing in the web chat. Every step goes through the same three checks as every other channel: Approvals (Careful, Balanced or Hands-off), Your rules, and In training. A step that has to wait is left on a Needs you card. Every action is recorded as coming from the AI app. On a server without signed callers (CALLER_ASSERTION_PUBLIC_KEYS), whoever holds the shared secret says who is asking.

This is the inverse of the MCP_ENABLED and MCP_SERVERS_CONFIG_PATH settings under Configuration, which configure the MCP gateway: Open Executive consuming other servers' tools. The server side needs no configuration beyond BACKEND_SHARED_SECRET. If a client errors on connect, try /mcp/ with the trailing slash, since the endpoint 307-redirects and not every client follows the redirect.

Document Upload

Upload your pitch deck, financial model, strategy docs, or any company documents via the web UI or API. The Executive will reference them when relevant.

# Via CLI
openexecutive upload deck.pdf model.xlsx strategy.md

# Via API
curl -X POST http://localhost:8000/documents \
  -F "[email protected]" \
  -F "domain=strategy"

Deployment

Two containers — the FastAPI backend and the Next.js UI — plus one persistent volume at /data. docker/docker-compose.yml is the reference topology and also what make docker runs locally, except that the compose UI runs as a dev server rather than the production image.

Versioned images are published to GitHub Container Registry on every release, so a deployment can pull instead of building:

docker pull ghcr.io/foreactlabs/openexecutive-api:<version>
docker pull ghcr.io/foreactlabs/openexecutive-ui:<version>

Tags: X.Y.Z and X.Y for a release, latest for the most recently published release, and main for the current head of main. The available versions are listed on each package's page under the repository's Packages. See docs/deployment.md for how releases are cut.

Upgrading. Settings → About shows the running version and says when a newer release is out (UPDATE_CHECK_ENABLED=false turns the check off). To upgrade, back up, then rebuild or pull the new tag; see docs/deployment.md.

⚠️ Single-instance only: the scheduler claims rows via UPDATE … RETURNING, which is not safe across processes. A second API replica double-fires every scheduled action. Pin the API to one instance. The UI is stateless.

Set ANTHROPIC_API_KEY, BACKEND_SHARED_SECRET, BACKEND_ALLOWED_ORIGINS and OE_PUBLIC_DEPLOYMENT=1 on any internet-reachable instance. See docs/deployment.md for the full guide — persistent state, health-check timing, resource sizing, operations, and common failure modes.

Access control

The deployed UI is gated behind Google sign-in or SSO sign-in (any OpenID Connect provider, such as Keycloak, Okta or Entra ID) with an email allow-list, and the public API is protected by a shared-secret header between the UI proxy and the FastAPI backend. See docs/auth.md for the full setup (Google Cloud Console and SSO steps, required environment variables, adding/removing users, rotating secrets, and a debugging table).

Configuration

All settings via environment variables. Minimum required: ANTHROPIC_API_KEY — unless you configure a local or OpenRouter backend instead (see Running on Local Models). At least one provider must be set or the app refuses to start.

VariableRequiredDefaultDescription
ANTHROPIC_API_KEYYes¹—Anthropic API key
ANTHROPIC_WORKSPACE_IDNo—Required only for an organisation-scoped Anthropic key; sent as the anthropic-workspace-id header. Workspace-scoped keys need no value
DEFAULT_MODELNoclaude-sonnet-5-5Executive + most specialists
DEEP_REASONING_MODELNoclaude-sonnet-5-5CSO, CFO, GC, Board (deep reasoning on)
VECTOR_STORE_PATHNo./chroma_dbChromaDB directory
EPISODIC_DB_PATHNo./episodic_memory.dbSQLite for episodic memory
COMPANY_PROFILE_PATHNo./company/profile.yamlCompany profile
ENABLE_CACHINGNotrueAnthropic prompt caching
ROUTING_MODELNoclaude-haiku-5-5Model for intent routing
SLACK_BOT_TOKENNo—Slack bot OAuth token
SLACK_APP_TOKENNo—Slack socket mode token
SLACK_SIGNING_SECRETNo—Slack signing secret: hear Slack over HTTP at /webhook/slack instead of socket mode
EXEC_EMAIL_ADDRESSNo—The Executive's own mailbox (Gmail via Google Workspace MCP, or Outlook via Microsoft 365 MCP)
EMAIL_POLL_INTERVAL_SECONDSNo60How often to poll for new email
TELEGRAM_BOT_TOKENNo—Telegram bot token (from @BotFather)
TELEGRAM_WEBHOOK_SECRETNo—Random string for webhook validation
BROWSER_EXTENSION_URLNo—Where Add to Chrome goes (your own listing, or the published one); unset shows no button: load it from the repo (setup)
OE_BROWSER_OFFNo—1 turns browser use off
DISCORD_BOT_TOKENNo—Discord bot token (Developer Portal → Bot tab)
DISCORD_APP_IDNo—Discord application ID (General Information tab)
DISCORD_GUILD_IDSNo—Comma-separated guild IDs for dev slash-command registration
DISCORD_NOTIFY_CHANNEL_IDNo—Default channel ID for outbound notifications
GOOGLE_CHAT_PROJECT_NUMBERNo—GCP project number for Google Chat
GOOGLE_CHAT_SERVICE_ACCOUNT_FILENo—Path to service account JSON key
GOOGLE_OAUTH_CLIENT_IDNo—Google OAuth client ID (Gmail MCP)
GOOGLE_OAUTH_CLIENT_SECRETNo—Google OAuth client secret (Gmail MCP)
MS365_MCP_CLIENT_IDNo—Entra app registration client ID (Microsoft 365 MCP: Outlook mail + calendar)
MS365_MCP_TENANT_IDNocommonEntra tenant ID for the Microsoft 365 MCP sign-in
MS365_MCP_CLIENT_SECRETNo—Entra client secret (only for a confidential app registration)
MS365_MCP_EXPECTED_USERNAMENo—Pin the Microsoft 365 sign-in to this mailbox UPN
MS365_MCP_ONEDRIVENofalseAdd the OneDrive tools (and the file permissions they need) to the Microsoft 365 sign-in; run --login again after turning it on
MCP_ENABLEDNoon when the config file existsMCP gateway (the Executive using other servers' tools); false keeps it off with the file in place
MCP_SERVERS_CONFIG_PATHNo./company/mcp_servers.jsonMCP gateway server list and tool filters (see mcp_servers.json.example)
EMAIL_PROVIDERNogoogleBackend for the inbound mailbox poller + alert email: google or microsoft
CALENDAR_PROVIDERNogoogleBackend for the typed calendar booking tools: google or microsoft
DELEGATION_GOOGLE_CREDENTIALS_DIRNo—Where the owner's own Gmail credential for As you lives (see docs/deployment.md)
DRIVE_SYNC_ENABLEDNofalseSync shared Google Drive folders into the knowledge base
DRIVE_SYNC_SERVICE_ACCOUNT_FILENo—Service account key with drive.readonly
DRIVE_SYNC_FOLDER_IDSNo—Comma-separated Drive folder IDs to sync
ONEDRIVE_SYNC_ENABLEDNofalseSync OneDrive or SharePoint folders into the knowledge base, as the Microsoft 365 sign-in
ONEDRIVE_SYNC_FOLDERSNo—Comma-separated <drive id>/<item id> entries (openexecutive onedrive-folder <sharing link> prints one)
CONFLUENCE_SYNC_ENABLEDNofalseSync Confluence spaces into the knowledge base
CONFLUENCE_URLNo—Confluence base URL (Cloud or Server/Data Center)
CONFLUENCE_USERNAME / CONFLUENCE_API_TOKENNo—Confluence Cloud sign-in (email + API token)
CONFLUENCE_PERSONAL_TOKENNo—Confluence Server/Data Center personal access token
CONFLUENCE_SYNC_SPACE_KEYSNo—Comma-separated space keys to sync
NOTION_SYNC_ENABLEDNofalseSync the Notion pages shared with your integration into the knowledge base
NOTION_API_KEYNo—Notion internal integration token
OPENROUTER_ENABLEDNofalseRoute Claude calls through OpenRouter and unlock non-Anthropic models per-agent in the Council UI
OPENROUTER_API_KEYNo—Required when OPENROUTER_ENABLED=true
OPENROUTER_CATALOG_ENABLEDNotrueFetch OpenRouter's live /models catalog at startup to populate the Council dropdown; falls back to a built-in list on failure
OPENROUTER_CATALOG_PROVIDERSNoopenai,google,anthropic,meta-llama,deepseek,x-aiVendor prefixes surfaced from the live catalog
OPENROUTER_CATALOG_PER_PROVIDERNo6Newest tool-capable paid models per vendor (0 = no cap)
OPENROUTER_CATALOG_REFRESH_SNo21600Background re-fetch cadence in seconds (0 = startup only)
LOCAL_MODELS_ENABLEDNofalseRoute selected slugs to a local OpenAI-compatible server (Ollama, LM Studio, vLLM, llama.cpp)
LOCAL_BASE_URLNo—Local server URL incl. version path, e.g. http://localhost:11434/v1. Required when LOCAL_MODELS_ENABLED=true
LOCAL_API_KEYNo—Optional bearer token (vLLM / gateways); Ollama & LM Studio need none
LOCAL_MODELSNo—Comma-separated local model slugs to surface in the Council UI and route locally, e.g. llama3.3,qwen2.5
LOCAL_TIMEOUT_SNo300Per-call timeout for local generation, in seconds
LOCAL_REASONING_EFFORTNo—reasoning_effort sent on every local request: none, minimal, low, medium or high. Set it for thinking-only models (GLM on Fireworks) that otherwise spend the whole token budget reasoning
HONCHO_ENABLEDNofalsePer-person memory layer (honcho.dev) — a peer card shared across all channels; on the Memories page each person sees only their own
HONCHO_API_KEYNo—Required when HONCHO_ENABLED=true
HONCHO_BASE_URLNo—Self-hosted Honcho endpoint
ENABLE_WEB_SEARCHNotrue²Let the Executive and specialists answer with live web results (news, market data, competitor moves) alongside your uploaded documents
WEB_SEARCH_MAX_USESNo2Max billed searches per agent per turn

See .env.example for the full list.

¹ ANTHROPIC_API_KEY is required only when you serve Claude models directly. It can be omitted entirely if you run on local models (LOCAL_MODELS_ENABLED) or route through OpenRouter (OPENROUTER_ENABLED).

² The application default is on, but .env.example ships ENABLE_WEB_SEARCH=false so a fresh setup incurs no per-search charges — if the agents tell you they can't search the web or read the news, flip it to true in your .env and restart. Uses Anthropic's server-side web_search tool, so it applies to Claude models (local models can't use it). WEB_SEARCH_ALLOWED_DOMAINS / WEB_SEARCH_BLOCKED_DOMAINS scope where it may look (set at most one).

Running on Local Models

Open Executive can run against any OpenAI-compatible local server — Ollama, LM Studio, vLLM, or llama.cpp — instead of (or alongside) the Anthropic API. Local model slugs route to your server through the same provider abstraction the hosted models use; no agent or orchestrator code changes.

# 1. Pull a capable, tool-use-friendly model (example: Ollama)
ollama pull llama3.3

# 2. In .env — point at the local server and list the slugs to expose
LOCAL_MODELS_ENABLED=true
LOCAL_BASE_URL=http://localhost:11434/v1   # Ollama default
LOCAL_MODELS=llama3.3

# 3. (Optional) run with NO Anthropic key — make local the default everywhere
DEFAULT_MODEL=llama3.3
DEEP_REASONING_MODEL=llama3.3
ROUTING_MODEL=llama3.3
# ...and leave ANTHROPIC_API_KEY unset

The listed slugs appear in the Council UI model dropdown, so you can also run a hybrid setup — keep the Executive on Claude while flipping individual specialists to a local model per-agent.

Caveats. Server-side web search (ENABLE_WEB_SEARCH) and Anthropic prompt caching / extended thinking have no local equivalent and are automatically disabled for local models. Multi-agent routing leans heavily on tool use, so pick a model that's strong at it (e.g. Llama 3.3 70B, Qwen2.5) — small models may route poorly. LOCAL_API_KEY is only needed if your server (vLLM, or a gateway) requires a bearer token; Ollama and LM Studio need none.

Using a hosted OpenAI-compatible gateway

The LOCAL_* settings are not limited to localhost — the same recipe works with any hosted OpenAI-compatible endpoint (an aggregator or inference gateway):

LOCAL_MODELS_ENABLED=true
LOCAL_BASE_URL=https://gateway.example.com/v1
LOCAL_API_KEY=your-gateway-key
LOCAL_MODELS=vendor/model-a,vendor/model-b

The listed slugs are sent to the gateway verbatim and appear in the Council UI dropdown, exactly like local slugs. Gateways that speak OpenRouter's request format and model namespace can alternatively be used through the OpenRouter path (OPENROUTER_ENABLED=true + OPENROUTER_API_KEY + OPENROUTER_BASE_URL=https://gateway.example.com/v1), which keeps that path's Claude-name translation and feature handling.

The local-model caveats above apply to the LOCAL_* path unchanged, plus one that matters more for hosted endpoints: everything the Executive processes — company profile, documents, conversations — is sent to whichever endpoint you configure here. Point these settings only at a provider you trust with that data.

Adding a New Specialist Agent

  1. Create packages/core/openexecutive/agents/your_agent.py extending BaseAgent
  2. Add a system prompt constant in packages/core/openexecutive/prompts/domain_prompts.py
  3. Register in packages/core/openexecutive/orchestrator/router.py — add to SPECIALIST_REGISTRY and the specialist enum in SPECIALIST_TOOLS
  4. Add domain alias to DOMAIN_ALIASES in packages/core/openexecutive/knowledge/retriever.py, and the specialist's area in plain words to _AREAS in packages/core/openexecutive/orchestrator/answer_sources.py
  5. Add knowledge docs to packages/core/openexecutive/knowledge/builtin/your_domain/
  6. Add at least 2 eval scenarios to packages/core/openexecutive/evals/_scenarios/
  7. Submit a PR — CI requires all of the above

Development

make dev          # Start FastAPI + Next.js
make test         # Run Python tests
make eval         # Run eval suite
make lint         # Run ruff + mypy
make docker       # Build and run Docker stack

# Unit tests only (no API calls required)
pytest packages/core/tests/unit/ -v

Evaluation System

packages/core/openexecutive/evals/_scenarios/ holds over 100 scenarios, covering every specialist and features such as As you, Approvals, departments and MCP tools. make eval runs them against a running API and scores each answer with an LLM-as-judge (Opus by default, or EVAL_JUDGE_MODEL). Each scenario defines a query, simulated company context, expected topics, required specialist routing, and a rubric. CI checks that a new agent or prompt change comes with its scenarios; the scored run is make eval, which writes its results to evals/results/.

Privacy

Everything in company/ is gitignored — the profile YAML, uploaded documents, and the ChromaDB vector store. None of this leaves your local machine (or your own volume in cloud deployments) except as part of prompts sent to the Anthropic API. Anthropic does not train on API data.

Inside the app, what the Executive keeps about a person is theirs: Always in the loop notes and the peer memory card are shown only to that person, the owner included.

Contributing

See .github/CONTRIBUTING.md. All PRs must include:

  • Working implementation (no stubs)
  • Tests for new behavior
  • Eval scenarios for new agents or prompt changes

License

Apache 2.0 — free to use commercially, requires attribution.