Skip to content

Roadmap

Where Embody and Envoy are headed. This is a living document -- direction, not promises -- and items move as the ecosystem moves (client support for new MCP protocol features in particular). Dates are deliberately absent. Want to influence it? Open an issue.

Landed since the last revision (2026-08-12)

Items that sat here and have since shipped -- kept one revision for the record:

  • Headless fresh-install runs. The smoke harness seeds dialog responses and runs a virgin install unattended end to end (startup verdict plus a feature phase: TDXN round-trip, checkpoint, portable export, Envoy, Convoy). What remains of the old "Next up" item is the narrower polish below.
  • Streaming SSE bridge (A-46). Tool responses stream incrementally with idle-window and absolute caps -- the shared prerequisite the elicitation item below used to be blocked on. Server-side progress notifications and elicitation now wait only on client protocol adoption.
  • Runner-forced settings can no longer bake into receipts. The transient parameter registry (v6.2.2) keeps Filecleanup/Toxdropexpr test values out of exports -- the worst consequence of invisible modal-adjacent state.
  • The whole v6.1-v6.2 line shipped and published (TDXN rename with /tdxn/ docs URLs, per-client MCP configs, the 80% smaller promoted surface, the TDXN review fixes, the skills once-over, .agents/skills for every skills-capable client).

Next up

  • Ideas adopted from td-mcp-rs (credited in the README). Five things it does that Envoy does not yet, in the order they matter:
    • OS dialog detection and dismissal (landed: list_dialogs / dismiss_dialog, Windows verified, macOS unverified). A TD modal (missing file, save-changes, crash recovery) is exactly how an Envoy call dies with a timeout; the fix is a tool that lists the popups a TD process owns and dismisses one, on Windows and macOS. Convoy remote launches gain the most.
    • Per-call instance addressing (landed: instance on every tool). An optional instance argument on every tool instead of the sticky switch_instance, which drags every peer bridge along when it changes target.
    • Stable error codes with fix hints (landed: error_code on every envelope). Every failure carries a code and a suggestion, with a test that the catalog is complete. Agents that retry act on codes better than on prose.
    • Capture for every operator family (landed: capture_op). capture_top is TOP-only; rasterizing CHOP, SOP, DAT and COMP viewers through an OP Viewer TOP gives the agent eyes on all of them.
    • Shader lint on DAT writes (landed). A GLSL text write returns the consuming operators' compile diagnostics in the same response, instead of leaving them for the next get_op_errors.
  • Structured decisions instead of invisible modals. When an AI session drives an operation that needs a human-style decision (palette black-box vs full export, file-cleanup keep/delete, dropped-tox expression warnings), the tool should return a structured needs_decision payload -- and accept the decision as an explicit parameter -- rather than opening a TD modal nobody is sitting in front of. TD dialogs remain the path for human-driven moments. (The seeded auto-response layer the smoke harness uses is the test-side half; the tool-payload half is still open.)
  • Setup Wizard suppression for automated installs (landed for the bootstrap path: embody_bootstrap.py grafts a self-deleting Execute DAT that applies the wizard's backend on first open). The seeded-response path covers the smoke harness today; a first-class headless switch (apply Auto defaults, enable Envoy per configuration, no dialogs armed) is the remaining polish.

Gated on MCP 2026-07-28 client adoption

The 2026-07-28 MCP revision (stateless core, versioned extensions) ships in SDK 2.x, which Envoy already runs. The features below need clients that speak the new revision; Claude clients currently negotiate the 2025-era revision with Envoy, so these wait on the ecosystem:

  • Protocol-revision telemetry. Log which MCP revision each session negotiates -- the tripwire that tells us when 2026-07-28 clients actually arrive in the field and the items below become real.
  • Tasks extension binding. Bind the shipped job layer -- run_tests with background=True, save_project, and the get_job_status poll, whose results park restart-proof on disk -- to MCP's Tasks extension, with capability gating and a polling fallback for older clients. The job layer was deliberately built in the Tasks shape, so this is an adapter rather than a rearchitecture.
  • MCP Apps. Interactive UI surfaces served through the protocol into the client. Natural Embody candidates, roughly in value order: a live TOP capture viewer (pick an op, watch it -- versus one-shot image embeds), the Embody manager / externalization dashboard, and eventually the Setup Wizard itself for agent-driven installs.
  • True MCP elicitation. Mid-tool questions routed to the client over the protocol. The bridge-side blocker is gone: the old read-to-EOF transport was replaced by the incremental streaming reader (A-46), so what remains is bidirectional routing plus a client that speaks the new revision. The structured-decision pattern above delivers the same workflow benefit today without protocol support.

Hardening backlog

Smaller items accepted-and-deferred from recent reviews (v6.0.162's adversarial panel, mostly). Tracked here so they cannot silently vanish:

  • Cross-instance install mutex: two TD instances bootstrapping the same project venv can race; uv's own locking and the spec-stamp self-heal mitigate, but a lock beside the venv would close it.
  • A worker-thread print() on the transport-security fallback path routes to the Textport from off-main-thread (fires only on a broken venv).
  • Import-gate edge: a 1.x MCP stack whose original import died before fastmcp registered can slip past the legacy-stack refusal (worst case is a crash-then-clean-restart, not a stuck state).
  • The import-gate warm-up poll re-arms unbounded if a cold import hangs indefinitely (e.g. antivirus wedge) -- needs a bound and a give-up message.
  • Multi-project-in-one-TD-process: the stale-interpreter refusal's wording says "upgraded on disk" when the actual cause is a second project with a different mcp version (the remedy it gives -- restart TD -- is correct).
  • Fresh-install smoke: a first-class wizard-suppression switch (see Next up) -- the settle-check race and the seeded-response path landed already.