Bunkhouse is an open-source AI employee platform for small and medium-sized main-street businesses, released by braedonsaunders on GitHub. Its core philosophy differs from most agent products: most products start from a chat window and treat prompts, models, and tools as configuration items; Bunkhouse starts from an employee record and a company inbox, so what users manage is an "employee," not a prompt.
Purpose Bunkhouse is used to "hire" AI employees for small and medium-sized businesses. Each agent has a real email address under the company domain, and coworkers do not need to install new software; they can collaborate with it simply by sending email, with email threads becoming the primary audit anchor. Agents are given a role, responsibilities, personality, governed processes, working hours, and a conservative initial level of autonomy, and are placed into an org chart shared with humans, with reporting lines, responsibilities, delegation, and escalation paths.
Main capabilities
- Real email: supports connecting to Google Workspace, Microsoft 365, or IMAP/SMTP, with incremental sync, threaded email, replies, and attachments.
- Roles rather than prompts: can be hired from role packs, including responsibilities, personality, governed processes, working hours, and a conservative initial autonomy posture.
- Self-evident processes: processes are versioned, assigned, referenced in outputs, and pinned to the revision actually followed in a given run.
- Runtime-enforced autonomy: email, record changes, money-related work, file writes, calls, shell, desktop, and remote computer access each have enforceable tuning knobs.
- Salary as budget: each agent has a monthly token budget, uses the company's own model provider keys, and overages are visible as "overtime"; Bunkhouse does not add any markup.
- Readable memory and real files: memory is human-readable and editable; deliverables are real DOCX, XLSX, and PDF files stored in connected S3-compatible storage. Users can attach work files directly in chat, and each file is accounted for, read into the request, and copied to the employee's persistent Linux home directory for use by the desktop or shell.
- Hybrid org chart: humans and agents share reporting lines, responsibilities, delegation, and escalation.
- Evidence rather than performance: computer use, commands, calls, approvals, tool effects, and run results are all persistent records bound to the employee, run, trigger, and tenant.
- Unified workbench: chat, calls, and observable work share the same workspace. Starting a call no longer opens a separateProduct, New → Call creates a conversation and replaces the central record with an animated live call stage and controls, while the right-hand workbench remains available before and after the call.
- Approval handoff: when a governed action needs a decision, the agent ends with an explicit handoff, rendering the exact request inline in the conversation, with optional notes and approve/reject controls; approval automatically resumes the paused run.
- Remote computers: configured under Library → Computers, governed by company feature switches and the agent autonomy knob. The default Docker stack includes Bunkhouse's own gateway and guacd, with no need for a Steward service or an external Guacamole server. Credentials are sealed and are only unsealed inside the server adapter, never entering prompts, browser payloads, or event records.
How to use Docker is the only prerequisite. After a fresh checkout, a single command pulls the release images and starts PostgreSQL, Redis, the web app, the Worker, the remote computer gateway, and its Guacamole sidecar. Open http://localhost:4810，以 owner@example.com / bunkhouse-first-boot to log in. You can set ADMIN_EMAIL and ADMIN_PASSWORD in the shell or in a .env file to choose the first owner credentials, and rotate any evaluation credentials before sharing the instance. Quick start migrates on startup and uses named volumes, so the database, object storage, and each agent's microVM disk overlay persist across container restarts.
Bunkhouse can run without an agent desktop. On a Linux host with /dev/kvm, the full per-agent microVM desktop can be started as part of the same stack. This desktop configuration cannot run through Docker Desktop on macOS or Windows, because those products do not pass /dev/kvm into Linux containers. If no reachable Desk host is available, email, calls, governed documents, connectors, and pure computation remain available; the browser, terminal, employee machine files, and desktop fail closed together because they belong to the same microVM.
For development from source, see Contributing; before exposing a deployment, review Security.
Use cases Suitable for small and medium-sized businesses that want to operate by "hiring AI employees," especially scenarios that require real email collaboration, governed processes, auditable evidence, and hybrid human-machine organizational structures. Also suitable for teams that need to combine AI work with the company's own model keys, budget controls, and approval mechanisms.
Limitations Bunkhouse is alpha software and has not yet completed an independent security audit, accessibility audit, or a published real-business pilot. It is recommended to evaluate with synthetic or parallel data, keep autonomy conservative, review external actions, and test recovery before relying on it. Tenant-owned rows are protected by PostgreSQL row-level security and server authorization; important ledgers reject updates and deletions at the database boundary; approvals and execution use resumable leases; sensitive configuration is sealed with AES-GCM via AppKit and never enters source control. The core is licensed under the GNU Affero General Public License v3.0, and packages/roles and future skill packs are licensed under MIT, so that roles, processes, and capabilities remain portable.
