gptme's Web UI Now Shows Your Agent Working in Real Time
The gptme web UI now streams tool outputs live as they arrive and renders tool calls as structured, collapsible cards — no more waiting for long commands to finish before seeing anything.
The gptme web UI now streams tool outputs live as they arrive and renders tool calls as structured, collapsible cards — no more waiting for long commands to finish before seeing anything.
The gptme web UI now shows you what your agent is doing while it’s doing it.
Two PRs merged today give you live streaming tool output and structured tool
call cards — the web UI finally has the same real-time visibility you get from
running gptme in the terminal.
The problem
Before this week, the web UI went silent during tool execution. If your agent ran a shell command, built a file, or called a slow API, the UI showed a spinning indicator and nothing else until the tool finished. For a 30-second compilation, that’s 30 seconds of “is anything happening?”
That made the terminal strictly better than the browser for actually watching your agent work. That’s backwards for a local-first tool.
What shipped
Live streaming output (#2674):
The server now emits a tool_output SSE event for each output chunk as it
arrives from the tool executor. The web UI subscribes and renders partial output
in a terminal-style code block in real time — same experience as watching the
terminal, but in the browser.
# Before: silent for the full 5 seconds
sleep 5 && echo "done"
# After: you see a live code block appear, then "done" streams in when it's ready
Structured tool call cards (#2676):
Tool calls are now rendered as collapsible RichToolCall cards with a colored
left border by category (file ops, shell, code, browser), an icon, a one-line
summary of the key argument, and expandable args and output. Scanning through a
long conversation for “that shell command that ran the tests” is now a visual
scan, not a text search.
SSE stream recovery (#2679): If the event stream drops — server restart, network blip, browser tab going idle — the UI reconnects and replays missed events. The conversation stays consistent rather than silently falling behind.
Why this matters
gptme runs on your machine against your shell. The whole point is that you control what the agent touches. Real-time visibility is how you exercise that control — you see the tool call, you see the output, you intervene if something looks wrong. Hiding execution output behind a spinner is the wrong default for a local-first agent.
These three PRs together close the gap between “watching in the terminal” and “watching in the browser.” The web UI is now a valid interface for actually supervising agent runs, not just reading the transcript afterward.
Honest limits
The RichToolCall card styling is MVP-level. Some tool output types (binary
output, very long shell output) render as a truncated code block. The category
coloring covers the common cases (file, shell, code, browser); edge cases like
MCP tool calls fall back to a generic card. More polish is coming.
Try it
pip install --upgrade gptme
# Start the server
uv run gptme-server --cors-origin=http://localhost:5701
# Or via the packaged webui
uv run gptme serve
Open the web UI and run a task with a long-running shell command. You should see the output stream in live. If the card categories aren’t mapping correctly for a specific tool type, file an issue.