The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.
description: >- What is being worked on next, and the known gaps, tracked in public issues. icon: map
Roadmap
This page is a reading of the open issue tracker, not a commitment. It has no dates. Each item links to the issue that owns it, so the issue stays the source of truth if this page goes stale.
Taking off the early-beta badge
This checklist defines done. The badge comes off when its boxes are closed, the memory smoke test passes on macOS, Windows and Linux, and the release workflow can publish without hand-holding. A ticked box is not always a passing test, so read each box against the text beside it.
| Area | State |
|---|---|
| Memory end to end (store, restart, recall, and the agent actually uses it) | Two of three items ticked (one, two). The third is the end-to-end smoke test, which the checklist records as not passing yet. |
| Reply persistence: a completed reply never vanishes | Done |
| Clean-install sign-in on all three platforms | Open |
Windows: native modules are refused when %TEMP% grants Modify to a non-owner group, which leaves memory and connectors unavailable |
Open |
The release-pipeline and Windows path-length blockers on the checklist have since been closed, though their boxes are still unticked.
Larger efforts
- Agent to agent. This umbrella issue covers five phases: an agent-card endpoint and discovery, accepting tasks, delegating to an external agent as a tool, streaming and push notifications, and multi-instance coordination.
- Pluggable memory adapters. This issue adds an adapters directory and the dependency rule. Supermemory, mem0, agentmemory and cognee would follow. Today memory has two engines and no adapter directory.
- External agent runtimes. This issue adds a layer for driving Cursor, Windsurf and Codex, with a shared ACP transport for more and per-session model and effort selection. Session import from one of them is the active part.
- Local and self-hosted. One issue would bring back a managed local runtime with model downloads and a compute-backend selector. Today local inference means an endpoint you run yourself. Another is a no-UI Linux build for a server or a Raspberry Pi.
- Observability. This issue adds scores to traces, so user feedback and automated quality signals sit next to the run they describe.
Built but not working yet
Two surfaces exist in the build and do nothing:
- Follow-up suggestion chips are built but nothing produces them, so the row is always empty (issue).
- SearXNG is configurable and reachable over the MCP server, but
searxng_searchis not registered as an agent tool, so the agent in chat cannot use it (issue).
The app also lists its own maturity. Settings → About shows every capability as stable, beta or coming soon. It is a better answer than this page to "is X finished".
Recently removed
These features were once documented and are gone:
- The live Google Meet agent. It joined a call through the old embedded Chromium webview and spoke back as a camera stream. See Voice.
- The embedded Chromium (CEF) runtime. The shell now uses Tauri's native webview, and browser control drives a real Chrome over the DevTools protocol. See Browser and computer control and the historical CEF notes.
- In-app model downloads and runtime management for local models. Point OpenHuman at a server you run.
- The skills sandbox. Skills are now a catalogue you browse and install, plus a run surface, not code running in a JavaScript sandbox in the app.
- The memory tree, graph view, people list and
MEMORY.md, replaced by the per-turn memory pack. Memory has the full list.
See also
- Release policy: how a release is cut and what the version gate does.
- Platform and availability: what runs where today.