GEAS · coordination platform

Who can do what,
when, and what depends on what.

A coordination platform for human developers and AI agents inside an organization. It sits above Git and agent harnesses: it doesn't replace either. It answers one atomic, verifiable question — who can claim what resource, at what time, and with what permissions — using declarative YAML trees versioned in Git and a lightweight database for runtime state.

github.com/cacelass/geas ← back to projects
01
problem

What it solves

Multiple developers (or agents) working on the same repository collide on the same files. GEAS solves the coordination problem before it happens: atomic claims, resource locks, and role-based access control — so two agents can never lock the same file, and tickets are ordered by priority (which ones unblock the most others). Built in pure Python with zero runtime dependencies.
Profiles
3 deployment profiles: Individual, Team, Enterprise
Purity
Zero runtime dependencies — stdlib-only Python
Concurrency
Atomic claims + resource locks + TTL heartbeats
02
decisions

Engineering decisions

→ Declarative tree, not imperative: organization.yml, departments/, roles/ — YAML trees versioned in Git describe the desired state. geas sync applies them idempotently to the DB.
→ Atomic ticket claims: UPDATE ... WHERE status='FREE' — two concurrent agents never claim the same ticket. Locks prevent concurrent resource edits.
→ RBAC with declarative roles: each role declares its permissions. geas role grant is idempotent. The permission catalogue is versioned and auditable.
→ Git providers via REST: GitHub, GitLab, Bitbucket — stdlib-only, no SDKs. If the token is missing, the remote API fails cleanly (never silently degrades).
→ Multi-provider harness: run a ticket with a local command or call an LLM directly (OpenAI / Anthropic via HTTP, stdlib). Token usage and cost are recorded.
→ MCP tools for AI agents: agents query tickets, permissions, and context via Model Context Protocol — no custom client needed.
→ 3 deployment profiles: Individual (SQLite, minimal), Team (+roles, RBAC, locks, web UI), Enterprise (+departments, policies, SSO/OIDC). A single codebase scales from solo to org.
Relationship GEAS is the coordination layer that MeshHarmes plugs into for multi-agent workflows — it manages tickets, locks, and permissions so harness agents don't collide.
03
architecture

Architecture

Git: declarative tree
organization.yml, departments/, roles/, repositories/, policies/ — all versioned, all synced via geas sync .
DB: runtime state
Tickets, locks, executions, event log, audit log. SQLite by default, PostgreSQL planned for Team/Enterprise.
CLI
Init, org management, tickets, resources, work, harness, MCP tools — 50+ subcommands.
Web panel
geas serve → read-only dashboard at http://127.0.0.1:8787/. Ticket detail pages with traceability.