Your agents are production systems. Now they have traces.

Ground Station is open-source observability for AI agents: it captures every model call, tool call, and token of an agent’s run, reconstructs it as a trajectory, and tells you why it was slow, what it cost, and what to change.

Get early access View on GitHub

trajectories/tr_01JB7QK4XEFix the flaky checkout tests

claude-codeacme/storefront running

Built for

Claude Code
Codex
OpenCode
Pi
Hermes
OpenClaw
browser agents
custom frameworks

02 trajectory the primitive

The trajectory is the trace of autonomous software.

One complete agent execution, from prompt to completion, captured as a single structured object. Every turn, model call, tool call, observation and subagent, in order, with time, tokens and cost attached.

distributed tracingrequesttracespans

agent observabilityprompttrajectory turns model calls tool calls observations subagents

tr_01JB9M2R7QRefactor authentication to use OAuth

claude-code1m 50s210k tok$0.33 ok

03 data → information → knowledge → action

From events to answers.

Capture what happened. Understand why. Know what to change. The same four steps are the Ground Station roadmap, and each one only works because the one before it is exact.

V0 · Data beta
What happened?
Every event, captured at the source: model calls, tool calls, file reads and writes, shell commands, token counts, context size, exit codes, cost. Timestamped, structured, complete.
- agent.started
- model.completed
- tool.started
- file.read
- shell.executed
- agent.failed
events.ndjson · 268 events · streamingtr_01JBA3KX2M · Fix the flaky checkout tests · 24m 06spatterns · 3 detected in this trajectoryrecommendation · 1 opportunity
model
read_file
grep · glob
edit_file
shell--workspace ×1814m 42s · 61%
git · ls
projected-p checkout ×17 + full suite ×13m 09s · −79%
0m5m10m15m20m24m 06s
14:31:13.205model.completed14:31:22.240tool.started14:31:24.653tool.started14:31:27.207tool.started14:31:29.052tool.started14:31:31.339file.read14:31:38.443tool.completed14:31:42.468shell.executed14:31:44.875model.completed14:31:52.711context.updated14:32:03.882file.read14:32:08.961file.read14:32:11.710tool.started14:32:14.442file.write14:32:17.771tool.completed14:32:20.618process.exited14:33:22.546tool.started14:33:25.048tool.completed14:33:27.937file.read14:33:31.093tool.completed14:33:34.120tool.completed14:33:37.406model.completed14:34:35.289model.completed14:34:44.856tool.completed14:34:47.343tool.completed14:34:50.633file.write14:34:52.890shell.executed14:35:51.167tokens.counted14:35:56.822tool.started14:35:58.858tool.completed14:36:01.349tokens.counted14:36:06.351tool.started14:37:03.982tool.completed14:37:06.839tool.completed14:37:09.463tool.completed14:37:12.871tool.completed14:37:16.722shell.executed14:37:20.521context.updated14:38:18.182context.updated14:38:28.276tool.started14:38:31.235tool.completed14:38:34.920tool.started14:38:38.749tool.completed14:39:31.688model.started14:39:43.030tool.completed14:39:46.638file.read14:39:48.867tool.started14:39:51.532tool.completed14:39:58.735tool.started14:40:52.015model.completed14:41:02.839tool.completed14:41:06.530tool.completed14:41:09.991tool.completed14:41:12.043process.exited14:42:08.903tool.started14:42:11.818tool.started14:42:15.145file.write14:42:18.645shell.executed14:42:22.858tool.started14:43:24.686file.read14:43:28.363shell.executed14:43:32.257tool.started14:44:25.483tokens.counted14:44:32.116file.read14:44:38.665file.write14:44:41.982tool.completed14:44:44.716shell.executed14:45:39.927tokens.counted14:45:48.725tool.completed14:45:51.232tool.completed14:45:57.043model.completed14:46:05.337shell.executed14:46:54.819context.updated14:47:04.838file.read14:47:08.078tool.completed14:47:15.470tool.completed14:47:19.244tool.completed14:47:22.247model.completed14:47:29.226process.exited14:48:20.101tokens.counted14:48:31.738tool.completed14:48:33.777model.completed14:48:44.142tool.started14:49:37.807tokens.counted14:49:44.642tool.started14:49:48.040file.read14:49:55.173tool.started14:49:57.392tool.completed14:50:00.675shell.executed14:50:03.797model.completed14:50:13.505process.exited14:51:13.552file.read14:51:16.279file.write14:51:18.792tool.completed14:51:21.429process.exited14:52:11.707tokens.counted14:52:22.343tool.started14:52:25.333shell.executed14:52:28.524process.exited14:53:22.264model.started14:53:31.840tool.completed14:53:33.899file.read14:53:37.765shell.executed14:53:41.867model.completed14:53:47.054tool.started14:54:40.705tool.started14:54:43.444file.read14:54:48.326tool.completed14:54:51.967tool.started14:54:53.845tool.started14:54:56.447context.updated14:55:05.969shell.executed
duration
24m 06s
model calls
30
tool calls
103
tokens
2.41M
est. cost
$2.54
longest op
cargo test · 52.4s
total test runtime
14m 42s
1. ①cargo test --workspace×18 · 14m 42s61% of trajectory
2. ②read_file tests/checkout_test.rs×31 · 4 versions27 identical reads
3. ③context window32k → 178k96k from test output
optimization opportunitybased on this run + 212 similar trajectories
Run affected-package tests while iterating.
61% of this trajectory was spent in cargo test --workspace. The agent ran the full suite 18 times, although every edit touched a single crate. Iterate with cargo test -p checkout and run the full workspace once before completion.
edits outside crates/checkout
0 of 9
full-suite failures outside checkout
0 of 18 runs
cargo test -p checkout
≈ 8.2s vs 49.0s
runtime
−48%
24m 06s → 12m 33s
tool calls
−17
103 → 86
compute
−43%
CI-minutes equivalent
+1 Context: truncate passing test output to its last 20 lines → context −53%, model cost −31%.
V1 · Information beta
What is happening?
Hundreds of disconnected events become one trajectory. Not a log file: a timeline with durations, tokens and cost that you can read in seconds.
events.ndjson · 268 events · streamingtr_01JBA3KX2M · Fix the flaky checkout tests · 24m 06spatterns · 3 detected in this trajectoryrecommendation · 1 opportunity
model
read_file
grep · glob
edit_file
shell--workspace ×1814m 42s · 61%
git · ls
projected-p checkout ×17 + full suite ×13m 09s · −79%
0m5m10m15m20m24m 06s
duration
24m 06s
model calls
30
tool calls
103
tokens
2.41M
est. cost
$2.54
longest op
cargo test · 52.4s
total test runtime
14m 42s
1. ①cargo test --workspace×18 · 14m 42s61% of trajectory
2. ②read_file tests/checkout_test.rs×31 · 4 versions27 identical reads
3. ③context window32k → 178k96k from test output
optimization opportunitybased on this run + 212 similar trajectories
Run affected-package tests while iterating.
61% of this trajectory was spent in cargo test --workspace. The agent ran the full suite 18 times, although every edit touched a single crate. Iterate with cargo test -p checkout and run the full workspace once before completion.
edits outside crates/checkout
0 of 9
full-suite failures outside checkout
0 of 18 runs
cargo test -p checkout
≈ 8.2s vs 49.0s
runtime
−48%
24m 06s → 12m 33s
tool calls
−17
103 → 86
compute
−43%
CI-minutes equivalent
+1 Context: truncate passing test output to its last 20 lines → context −53%, model cost −31%.
V2 · Knowledge next
Why is it happening?
Patterns surface across the trajectory: repeated commands, duplicate reads, runaway context. Ground Station explains where the time and the tokens actually went.
“This run was slow because the agent re-ran the full test suite 18 times.”
events.ndjson · 268 events · streamingtr_01JBA3KX2M · Fix the flaky checkout tests · 24m 06spatterns · 3 detected in this trajectoryrecommendation · 1 opportunity
model
②read_file
grep · glob
edit_file
①shell--workspace ×1814m 42s · 61%
git · ls
projected-p checkout ×17 + full suite ×13m 09s · −79%
0m5m10m15m20m24m 06s
duration
24m 06s
model calls
30
tool calls
103
tokens
2.41M
est. cost
$2.54
longest op
cargo test · 52.4s
total test runtime
14m 42s
1. ①cargo test --workspace×18 · 14m 42s61% of trajectory
2. ②read_file tests/checkout_test.rs×31 · 4 versions27 identical reads
3. ③context window32k → 178k96k from test output
optimization opportunitybased on this run + 212 similar trajectories
Run affected-package tests while iterating.
61% of this trajectory was spent in cargo test --workspace. The agent ran the full suite 18 times, although every edit touched a single crate. Iterate with cargo test -p checkout and run the full workspace once before completion.
edits outside crates/checkout
0 of 9
full-suite failures outside checkout
0 of 18 runs
cargo test -p checkout
≈ 8.2s vs 49.0s
runtime
−48%
24m 06s → 12m 33s
tool calls
−17
103 → 86
compute
−43%
CI-minutes equivalent
+1 Context: truncate passing test output to its last 20 lines → context −53%, model cost −31%.
V3 · Action later
What should change?
Findings become specific recommendations with an estimated impact, before you change a single prompt, tool or model.
events.ndjson · 268 events · streamingtr_01JBA3KX2M · Fix the flaky checkout tests · 24m 06spatterns · 3 detected in this trajectoryrecommendation · 1 opportunity
model
②read_file
grep · glob
edit_file
①cargo test--workspace ×1814m 42s · 61%
git · ls
projected-p checkout ×17 + full suite ×13m 09s · −79%
0m5m10m15m20m24m 06s
duration
24m 06s
model calls
30
tool calls
103
tokens
2.41M
est. cost
$2.54
longest op
cargo test · 52.4s
total test runtime
14m 42s
1. ①cargo test --workspace×18 · 14m 42s61% of trajectory
2. ②read_file tests/checkout_test.rs×31 · 4 versions27 identical reads
3. ③context window32k → 178k96k from test output
optimization opportunitybased on this run + 212 similar trajectories
Run affected-package tests while iterating.
61% of this trajectory was spent in cargo test --workspace. The agent ran the full suite 18 times, although every edit touched a single crate. Iterate with cargo test -p checkout and run the full workspace once before completion.
edits outside crates/checkout
0 of 9
full-suite failures outside checkout
0 of 18 runs
cargo test -p checkout
≈ 8.2s vs 49.0s
runtime
−48%
24m 06s → 12m 33s
tool calls
−17
103 → 86
compute
−43%
CI-minutes equivalent
+1 Context: truncate passing test output to its last 20 lines → context −53%, model cost −31%.

04 fleet.live 1 → 1,000 agents

One agent or one thousand. See what every one is doing right now.

Ground Station is not just a trace viewer for runs that already finished. It is a live operations view across every agent in your organization, so you notice the one that has been stuck for eighteen minutes before it finishes burning the budget.

trajectories · 24h: 1,284
success: 92.4%
median runtime: 3m 41s
p95 runtime: 14m 22s
tokens: 84.2M
est. cost: $742
tool errors: 1,421

select a row to open its trajectory

model	duration	tools	tokens	context	status
sonnet-5.5	03:21	14	842k	38%	running
gpt-5-codex	00:41	7	231k	21%	running
sonnet-5.5	18:43	114	3.12M	92%	slow · 3.0× typical
sonnet-5.5	04:11	22	670k	44%	running
opus-5.5	07:52	38	1.20M	57%	awaiting approval
haiku-4.5	00:12	3	18k	9%	running
gpt-5-codex	01:58	11	190k	26%	running
sonnet-5.5	02:37	9	104k	17%	running

05 detections knowledge, applied

Find the loop before it burns another million tokens.

Ground Station watches trajectories for the behavior that costs you: slow runs, tool loops, cost regressions, runaway context. Every finding comes with its evidence and a cause, not just an alert.

findings4 open

highslow trajectoryFix the flaky checkout tests · claude-code

24m 06s against a typical 6m 18s

evidence

p50 · 6m 18sthis run · 24m 06s

0m5m10m15m20m25m

where the 24m 06s went

cargo test 61%model 22%

other tools 4%harness overhead 13%

explanation

The trajectory was slow because the agent re-ran the full workspace test suite after every edit.

primary contributor

cargo test --workspace18 executions · 14m 42s cumulative · 61% of runtime

06 compare v41 ↔ v42

Know when an agent got worse.

Success rate barely moved. Tokens, runtime and cost did. Compare any two versions, models or prompts on the behavior that actually changed, not just on whether the task passed.

Diff behavior across

model upgrades
prompt changes
tool changes
agent releases
repository changes
infrastructure changes

v41→v42acme-coder

n = 1,904 / 1,877 trajectories

changed in v42 system prompt +212 tokens · new tool repo_search · model unchanged

Agent v41 versus v42, per-task metrics
metric	v42	change
tokens / task	46.2k	83.4k	+80%
runtime	3m 41s	5m 25s	+47%
tool calls / task	27	44	+63%
repo searches / task	4.1	11.8	+188%
est. cost / task	$0.41	$0.74	+80%
success rate	92.1%	92.4%	+0.3pp

finding v42 performs 2.9× more repository searches before its first edit. Outcomes are unchanged; cost per task is up 80%.

07 relay open-source collector

One daemon. Every agent.

The relay is a small Rust collector that runs on the same machine as your agents. Adapters stream events into it; it normalizes, redacts, buffers and ships them. Your agents never block on telemetry.

Claude Codehooks adapter
Codexadapter
OpenCodeplugin
Custom agentsSDK · TS · Python · Rust
Any GenAI appOTLP in

relayrust · local up 14d 3h

1ingestunix socket · OTLP/gRPC
2normalizeground schema v1
3redactsecrets · paths · env · regex
4sampleper agent · per tool
5bufferdisk-backed · survives restarts
6batchzstd · 512 events
7exportretry · backoff · auth

events/s: 1,284
redacted: 1,904
dropped: 0
p99 overhead: 0.4ms

Ground Stationself-hosted or cloud
OTLP exporterany OpenTelemetry backend
Local storeSQLite · local-only mode

Open source, down to the schema.

Telemetry from autonomous software is too sensitive for a black box. The collector, adapters, SDKs and event schema are open, so you can read exactly what is captured and extend it for your own agents.

relay: The collector. A single Rust binary that runs next to your agents.
adapters: Claude Code, Codex and OpenCode, with more on the way.
sdk: Instrument any agent loop in TypeScript, Python or Rust.
schema: A versioned, documented event schema aligned with OpenTelemetry GenAI conventions.

quickstart~/src/storefront

$curl -fsSL https://groundstation.sh/install.sh | sh✓ installed ground + relay → ~/.ground/bin$ground login✓ device registered · org acme$ground connect claude-code✓ hooks written → ~/.claude/settings.json✓ relay running · unix:///tmp/ground-relay.sock$claude "fix the flaky checkout tests"→ trajectory tr_01JBA3KX2M → relay  2 model calls · 5 tool calls · 43,140 tokens · live

08 redaction before the network

Redaction happens before anything leaves the machine.

Agent telemetry contains source code, credentials, customer data and proprietary prompts. The relay applies your rules locally, so privacy is a property of the architecture, not a checkbox in a dashboard. Toggle the rules to see what would be uploaded.

your machine

raw eventfrom claude-code adapter

{
  "event": "tool.completed",
  "trajectory": "tr_01JBA3KX2M",
  "tool": "shell",
  "cwd": "/Users/dana/src/acme-billing",
  "command": "psql $DATABASE_URL -f refunds.sql",
  "env": { "DATABASE_URL": "postgres://billing:hunter2@10.0.3.12:5432/prod" },
  "stdout": "maria.lopez@example.com | 4 refunds\njwu@example.org | 2 refunds\n(5 rows)",
  "prompt": "Why are refunds failing for enterprise customers?",
  "user": "dana@acme.dev",
  "duration_ms": 184,
  "exit_code": 0
}

relay rules~/.ground/relay.toml

Also: sampling per agent and tool, local persistence, retention windows.

uploaded event5 rules applied

{
  "event": "tool.completed",
  "trajectory": "tr_01JBA3KX2M",
  "tool": "shell",
  "cwd": "path:7c1e9a42",
  "command": "psql $DATABASE_URL -f refunds.sql",
  "env": { "DATABASE_URL": "[SECRET:postgres_url]" },
  "stdout": "[EMAIL] | 4 refunds\[EMAIL] | 2 refunds\n(5 rows)",
  "prompt": "[EXCLUDED · 11 tokens]",
  "user": "user:3f9a0c1e",
  "duration_ms": 184,
  "exit_code": 0
}

09 otlp.export fits your stack

Speaks OpenTelemetry. Adds agent semantics.

Ground Station doesn’t replace your tracing. Trajectories export as OTLP traces, so the agent run sits inside the same trace as the request that started it, and the 18 minutes stop being one opaque span.

How Ground Station concepts map to OpenTelemetry
ground station	otel	name
trajectory	→ trace	`ground.trajectory.*`
model invocation	→ span	`gen_ai.chat`
tool invocation	→ span	`gen_ai.execute_tool`
subagent	→ child span	`gen_ai.invoke_agent`
events	→ span events	`file.read · shell.exit`
measurements	→ metrics	`gen_ai.client.token.usage`

trace 4bf92f3577b34da6a3ce929d0e0e4736

POST /api/tasksweb
tasks.createapi
queue.enqueueapi
ground.trajectoryRefactor payments module
gen_ai.chatsonnet-5.5
gen_ai.execute_toolread_file
gen_ai.chatsonnet-5.5
gen_ai.execute_toolshell · cargo test
gen_ai.execute_tooldatabase · psql
db.querypostgres
gen_ai.invoke_agentsubagent · explore
tasks.update_statusapi
+ 106 more agent spans · 2 span events each

span attributes · gen_ai.execute_tool · cargo test

gen_ai.operation.name: execute_tool
gen_ai.tool.name: shell
ground.trajectory.id: tr_01JB8W4ZQD
ground.tool.command: cargo test --workspace
ground.tool.exit_code: 101
ground.context.tokens: 184,112

Every agent leaves a trajectory. Make it observable.

Ground Station is in private beta. The collector, adapters and schema are open source. Bring one agent, and see its first trajectory in minutes.

Get early access Star on GitHub

The observability layer for autonomous software

V0Datacapture everythingbeta
V1Informationreadable trajectoriesbetayou are here
V2Knowledgepatterns & explanationsnext
V3Actionrecommendationslater