Dashboard

Developer views#

Eight tabs for when the top-level pages have told you that something is wrong and you need to know why.

Flow#

An animated architecture diagram that lights up as messages move through the system: channel → gateway → model → tools → back. Colour-coded by type, with errors in red.

It is the fastest way to build a mental model of where time and money go in a turn, and the fastest way to spot a stage that is silently failing — a channel that never lights up is a channel that is not delivering.

curl -s localhost:8900/api/flow | jq
curl -s localhost:8900/api/component/gateway | jq
curl -s localhost:8900/api/component/brain | jq

Clicking any component opens its detail: the tool's call count and latency percentiles, the runtime's configuration, the machine's resources, the gateway's connection state.

Models#

Per-model behaviour rather than per-model cost. Which models you actually use, their latency distribution, their error rates, and how usage has shifted over time.

The routing question — "should this task be on a cheaper model?" — starts here and ends in Reducing spend.

LLM Context#

What the model actually saw on each turn.

This is the tab that resolves arguments. An agent that "ignored the instructions" usually did not receive them — because a compaction dropped them, or a file read returned less than expected, or a system prompt was overwritten. The context view shows the assembled input rather than what you assume it was.

curl -s localhost:8900/api/context-anatomy | jq

Agent Graph#

Cross-session spawn topology, built from span data. Which agent spawned which, how deep the tree went, and what each branch cost.

On a fleet running orchestrators — QM, Hermes, OpenClaw with sub-agents — this is the only view where the real shape of the work is visible. A three-level delegation tree looks like eight unrelated sessions everywhere else.

curl -s 'localhost:8900/api/local/agent-graph?limit=500' | jq '.nodes | length'

Tools#

Every tool the agent uses, by provenance, with call count and p50/p95 latency.

Provenance matters: a built-in tool, an MCP server's tool and a skill-provided tool behave differently and fail differently, and a catalogue that merges them hides the pattern. The MCP server list is here too, with per-server statistics.

curl -s localhost:8900/api/tool-catalog | jq '.tools[:10]'
curl -s 'localhost:8900/api/tool-catalog/<name>/calls' | jq
curl -s localhost:8900/api/mcp-servers | jq
curl -s localhost:8900/api/mcp-stats | jq

A tool with a p95 far above its p50 is usually the reason a session felt slow.

Context usage#

Context-window utilisation over time, compaction triggers, and tokens reclaimed.

Compaction is expensive and lossy, and a session that compacts repeatedly is usually solvable — by trimming what gets loaded rather than by upgrading the model. This tab shows you the pattern.

curl -s localhost:8900/api/context-economics | jq
curl -s localhost:8900/api/usage/compression | jq

Runtime extras#

What the selected runtime uniquely exposes, beyond the generic tabs. The tab is hidden unless the current runtime has extras.

Examples: Grok's manifest of the codebase archives it staged for upload; Cline's rejected tool calls and its schedule executions; Copilot's credit ledger; Kimi's context-window occupancy; Devin's abandoned branch count.

curl -s localhost:8900/api/harness/data | jq

Ask#

Plain-English questions about your own agent usage, answered from the store.

curl -s localhost:8900/api/advisor/ask -X POST \
  -H 'content-type: application/json' \
  -d '{"question": "which runtime grew fastest in cost last month?"}' | jq

Saved questions become dives you can re-run. If you would rather your coding agent asked these questions as part of its own work, that is the MCP server.