What works, what has limits, and what is not here.
Termyte has a working local product. It is still an alpha. This page is part of the product: you should know what the system can do before you trust what it tells an agent.
Available now
Local activity capture, task state, evidence, checkpoints, resume packets, cross-agent handoffs, background processing, searchable project experience, context delivery, feedback, and a local Viewer.
Works with limits
OpenCode uses a generated local plugin and does not support automatic context injection. Ranking calibration is incomplete, redaction is heuristic, and processing quality depends on the selected provider.
Not provided
Company-source connectors, a hosted team service, Viewer task-management UI, import or bulk deletion, sandbox isolation, guaranteed agent safety, or proof of productivity gains.
Evidence is not the same as proof
Termyte links memories and context decisions to captured traces, observations, files, commands, and task evidence. That makes the work inspectable. It does not prove that a memory caused an agent to succeed, fail, or make a particular change.
If a task passes after a briefing is delivered, several things may have caused that result. The current alpha records the path so you can inspect it. It does not claim to measure cause and effect.
Capture has a boundary
Only activity that reaches a configured integration becomes a Termyte record. Redaction reduces exposure before processing, but it is heuristic and should not be treated as comprehensive secret removal.
The Viewer is for inspection
The local Viewer exposes session, task, context, memory, provenance, and runtime data, but it is not a full task-management application. Task operations are performed through the CLI or MCP tools.