For AI agents → AGENTS.md

The seven core projects are Arbora, Overstory, Medulla, Meristem, Scion SDK, Tuig, and Axil. Tuig and Axil are further operator surfaces on the same Arbora boundary. Overstory provides the browser operator surface through Arbora. MCP, REST, and A2A clients use the same governed Arbora boundary. Arbora composes the runtime and serves every operation. Meristem is the in-run harness. It applies decisions it receives and holds no store. Medulla commits durable knowledge, evidence, provenance, and reasoning results. Source systems connect through the governed Scion SDK contract.

Arbora is the governed runtime entry point. With Python 3.12+ and uv installed, launch the zero-infrastructure local profile:

uvx --from arbora setup-config generate --profile tiny
uvx --from arbora arbora --transport stdio

The generated profile starts with local, bounded defaults. Continue with the Arbora quick start for provider configuration, Overstory hosting, messaging, and authenticated network transports.

Every repository documents its own contract and links back to this same runtime map. The boundaries are deliberate: Arbora is the door, Meristem runs agents in-run, Medulla owns durable knowledge and reasoning, Overstory presents the browser surface, and Scion SDK governs source integration.

Build and release evidence lives with the component that owns it. Use each repository's Actions page for current CI and its Pages Status or Capabilities section for shipped behavior. Public issues and pull requests are tracked in the owning repository so operational and design discussions retain their architectural context.

The Arbora Architecture Explorer shows every spec row of the core repositories, the connectors and the desktop client with its status, placed on the part of the architecture it builds, and is rebuilt on every merge to a core repository's main branch, and at least every 30 minutes. SPECIFIED rows are open work waiting to be built; delivery and acceptance are separate states. The contribution guide links the shared arbora-development and spec skills.

Concept-proof phases P1 to P5 shipped as release bundles. The architecture explorer shows the releases. Rust serde types are the source of truth for the Medulla wire format. Clients are hand-written and tested on round-trip vectors, and the repositories commit no generated API contracts. Spec status is computed at read time. Spec IDs use the prefixes MEDULLA-, ARBORA-, MERISTEM-, SCION-, OVERSTORY-, TUIG-, AXIL-, CONN--, and PIPE-.

The path is spec-driven, from constitution to release. Each step names its skill and the files it writes:

1. Constitution. Read the constitution. Every plan checks against it.
2. Route. Pick the owner repository with the ecosystem boundaries and arbora-development. Choose an open (SPECIFIED) row in Canopy.
3. Specify. speckit-specify and speckit-clarify write specs/<id>/spec.md: user stories, Given/When/Then scenarios, and FR- and SC- labels mapped to row IDs.
4. Clarify. Resolve every [NEEDS CLARIFICATION] marker in spec.md before you plan.
5. Plan. speckit-plan writes plan.md. Add contracts/ for cross-repository seams, data-model.md, quickstart.md, and research.md when unknowns remain.
6. Tasks. speckit-tasks writes tasks.md. Each line reads - [ ] T001 [P] [US1] <ROW-ID>.
7. Analyze and checklist. speckit-analyze checks consistency and writes checklists/requirements.md.
8. Implement. sdd-implementer builds each task. Run the repository gates.
9. Land. Open a pull request whose commits carry a Spec: <ROW-ID> trailer and a bound test. Do not edit status.json by hand.
10. Release. Landed rows ship in the next release bundle.

Standards: constitution, spec standard, spec templates, and Canopy. Install the skills with pip install universal-skills, then install-skills --all-detected --symlink. The step-by-step guide covers local setup, reuse, tests, and pull requests.

All seven core projects are released under the MIT License.