instrument on-device agents
capture latency, token usage, and silent failures from phones, robots, and vehicles
debug full execution traces
follow orchestration, inference, and tool calls end to end when behavior drifts
operate fleets safely
ship rollouts with runtime constraints, audit trails, and rollback-ready controls
This Year
0
commits
0
prs
0
issues
Most Active In
- Swarm0
- Hive0
- Wax0
- Conduit0
Projects
6 original
- Swarm
No description available.
n/a
0
- Hive
No description available.
n/a
0
- Wax
No description available.
n/a
0
- Conduit
No description available.
n/a
0
- Colony
No description available.
n/a
0
- Terra
No description available.
n/a
0
Topics
No topics detected.
In Public Repos
No language data.
Signals
0
external prs merged
Languages
Language distribution unavailable.
observability
instrument once, see every on-device turn
Observe spans capture invocation, inference, and tool execution so silent failures are diagnosable in production.
import Terra
let agent = Terra.Agent(name: "SupportAgent", id: "agent-123")
try await Terra.withAgentInvocationSpan(agent: agent) { scope in
scope.addEvent("agent.start")
}Yggdrasil
Live
P95 LATENCY
184ms
THROUGHPUT
3.1K
TOOL SUCCESS
98.7%
ERROR RATE
0.19%
Gemini Flash
Size8B
Latency96ms
Throughput2.8K/s
GPT-4o Mini
Size8B
Latency118ms
Throughput2.4K/s
Qwen 4B
Size4B
Latency84ms
Throughput3.4K/s
Sonnet 4.5 (Anthropic)
Size14B
Latency152ms
Throughput1.9K/s
router.dispatchgeminicoremltool34ms
retrieval.rankqwentool61ms
model.inferencegpt-4ocoreml218ms
response.mergesonnettool49ms
12:31:04.112INFOpipeline accepted trace batch: 12 spans
12:31:04.146DEBUGgemini route selected via confidence gate
12:31:04.329WARNvector cache miss for segment seg_3044
12:31:04.441ERRORtool timeout on policy.lookup (retry=1)
12:31:04.590INFOresponse completed with guardrails: pass
YGGDRASIL
CPU
GPU
Memory
On-Device AI Runtime
Monitor the Edge
Yggdrasil gives teams one control surface for model health, execution traces, and operational risk in prod
- Real-time visibility into latency, token usage, and silent model failures
- Trace-level debugging across orchestration, inference, and tool execution
- Fleet controls for OTA rollouts, constraints, and audit-ready operations
#buildinpublic
Yggdrasil started after repeatedly shipping on-device AI and realizing we were flying blind in production. teams had no reliable way to see silent inference failures, trace crashes, or understand regression risk before user impact.
we're building Observe in public on top of Swarm, Hive, Colony, Wax, and Conduit so teams can instrument once, ship safely, and scale from single apps to device fleets with verifiable runtime controls.
pricing
Billing period changes the Pro plan price.
STARTER
$0
instrument your first on-device agent.
- up to 10k trace events/month
- no credit card required
PRO
$499per month/per seat
for production mobile and edge teams.
- 10m trace events/month
- on-device failure diagnostics
- alerts + webhook integrations
- phone and slack support
ENTERPRISE
custom
for regulated and safety-critical fleets.
everything in pro, and:
- fleet rollouts and staged rollback controls
- policy enforcement with audit trails
- private deployment options
- priority support
on-device event packs
plans include baseline volume. add packs for launch spikes, OTA windows, and incident investigation bursts without dropping visibility.
$199 per 5M events