Skip to content

Roadmap

v0.3.1 (next patch) — Fixes

  • Channel / connector-registry unification (#31). Channels passed as Agent(channels=[...]) are now registered into the ConnectorRegistry, so workflow steps using channels.send_message (e.g. alarm_to_workorder.notify_technician) dispatch through them in live mode. sandbox=True now also gates channel.connect() / channel.disconnect() — no SMTP/Slack/Telegram logins in sandbox.

v0.2.1 — Consolidation

A focused hardening release between v0.2.0 and v0.3. No new features; the goal was an honest, stable base with its loose ends tightened:

  • Loud stub for machina.mcp.MCPServer. import machina.mcp stays importable across the v0.2 → v0.3 transition; instantiating MCPServer() now raises NotImplementedError with a pointer back here instead of silently handing back an empty namespace.
  • EmailConnector available as a communication connector. See docs/connectors/email.md for setup.
  • LiteLLM contract test against the real litellm.get_llm_provider parser. Pins the provider:model → provider/model normalization that produced reactive fix b48f649, and anchors the inverse (colon form must keep being rejected) so a future LiteLLM relaxation is noisy, not silent.
  • Extended example validator. tests/validate_examples.py now imports every runnable examples/*/agent.py and verifies module-level Agent(...) construction actually runs. Catches the "imports fine but blows up at first call" class of bug.
  • Per-module coverage floors in CI for the core modules (agent, config, llm, observability, workflows). Each floor sits below the measured baseline with a ~5% buffer, so normal refactor churn has headroom but a silent regression trips CI.

v0.3 — Next

Ordered by what moves adoption the most:

  1. MCP server layer. Expose every connector's declared capabilities as Model Context Protocol tools so Claude Desktop, Cursor, Continue, and any MCP-compatible client can talk to Machina connectors with no agent code. This is the biggest adoption multiplier on the roadmap and the reason machina.mcp has been reserved as a stable import path.
  2. More CMMS connectors — MaintainX, Limble, Fiix. Same BaseConnector / capability-declaration pattern as SAP PM, Maximo, UpKeep.
  3. Multi-agent orchestration (AgentTeam). Deferred to v0.3.1.
  4. Anomaly detection & RUL estimation on top of the IoT connector stream.
  5. Plugin system for community-contributed connectors without forking the core package.
  6. WhatsApp and Teams communication connectors.

MCP direction (standing position)

MCP is transport, not a replacement for connectors. The connector layer is Machina's normalization layer — it maps vendor payloads onto the canonical maintenance domain — and that, with the write-path invariants and typed capabilities on top, is the moat. MCP carries normalized data; it does not produce it.

  • The internal flip is rejected. We do not rewire Machina's own runtime↔connectors boundary to speak MCP, and we do not replace connectors with a bag of MCP tools. Internal boundaries stay native Python; MCP lives only at the edge. (See MACHINA_SPEC.md §17 for the full argument.)
  • The transport/mapper split already future-proofs against vendor MCPs. When a CMMS vendor ships its own MCP server, that becomes a new transport feeding the existing per-vendor mappers (connectors/cmms/mappers/) — a new fetch path, not a re-normalization. The durable work (mapping) is insulated from transport.
  • Outbound MCP (Machina as an MCP server) already exists — every connector's capabilities can be exposed as MCP tools (item 1 under v0.3 above).

Gated: inbound MCP-client connector

A generic inbound connector — Machina as an MCP client, consuming a vendor's MCP server into the domain model (the inverse of the outbound server). It would be built as transport (a generic MCP client) plus per-vendor mappers, reusing the same transport/mapper split. Gated behind the trigger "first real vendor CMMS MCP" — not built now. Until a CMMS vendor actually ships an MCP server worth consuming, a generic MCP-client adapter would be speculative surface with nothing to validate it against.

What's deferred beyond v0.3

  • Non-Python SDKs (Go / TypeScript clients).
  • Hosted control plane. Machina stays a framework, not a product.

How to steer the roadmap

  • New connector or integration idea → open an issue labelled connector describing the system, its API, and a minimal capability set.
  • Framework bug or papercut → issue with a failing test case when possible.
  • Strategic disagreement ("this should be a higher priority") → open a discussion; the ordering above is the maintainer's current best guess, not a commitment.