Lookspan

Local-first observability dashboard for AI agents. MCP-native. See every span your agents emit.

npx lookspan          # → http://127.0.0.1:3100

▶ Watch the full 75-second presentation

Agent (MCP · LangGraph · CrewAI · OpenTelemetry · HTTP)  →  POST /api/ingest  →  SQLite  →  real-time dashboard

🇪🇸 ¿Prefieres español? Lee el README en español.

⭐ If Lookspan saves you a debugging session, give it a star — it's the #1 way to help it grow. Got a use case, a bug, or a framework you'd like supported? Open an issue or say hi in Discussions.

The problem

When an AI agent misbehaves — fails, stalls, or quietly burns more tokens than expected — there's no native way to see what happened step by step. Existing observability tools are cloud-first: they want accounts, API keys, and shipping your production data to someone else's servers.

Lookspan takes the opposite approach: everything runs on your machine, data never leaves it, and infra cost is zero. Instrument your agent with an adapter (or just POST JSON) and open the dashboard in your browser.

Quick start

npx lookspan              # → http://127.0.0.1:3100, no install, no cloud

Send your first span from any language:

curl -X POST http://127.0.0.1:3100/api/ingest \
  -H "Content-Type: application/json" \
  -d '{"spans":[{"traceId":"t1","spanId":"s1","parentSpanId":null,"type":"llm_call","name":"agent.run","startedAt":"2026-06-02T10:00:00Z","endedAt":"2026-06-02T10:00:01Z","status":"ok","framework":"custom","model":"gpt-4o","provider":"openai","usage":{"inputTokens":1000,"outputTokens":500,"costUsd":0}}]}'

Open http://127.0.0.1:3100 and watch the trace appear — with its cost computed server-side.

Open it like any other app

If you'd rather click an icon than type a command:

npm install -g lookspan
lookspan install-desktop          # bake in flags if you want: --port 3200 --retention 7d

That adds a Lookspan entry to your Desktop and Start Menu on Windows, to ~/Applications on macOS, and to your app menu on Linux — with the Lookspan icon on Windows and Linux. Clicking it starts the server and opens the dashboard. lookspan uninstall-desktop removes it.

It needs a real install — a shortcut into the npx cache breaks the first time npm cleans it, so install-desktop refuses to create one.

Features

• HTTP span ingest — POST /api/ingest accepts JSON batches of spans. Works with any agent that can make an HTTP request.
• MCP-native — the @lookspan/mcp TypeScript SDK wraps any McpClient and emits a span per MCP tool call, without changing your agent code.
• Python SDKs — lookspan (generic client) plus adapters for LangGraph/LangChain (lookspan-langgraph) and CrewAI (lookspan-crewai).
• OpenTelemetry — an OTLP/HTTP receiver at POST /v1/traces; point any OTel exporter at it with no Lookspan SDK. gen_ai.* attributes map to provider/model/tokens.
• Real-time streaming — SSE endpoint GET /api/stream pushes span.ingested, trace.updated and alert.triggered to the dashboard, no polling.
• React dashboard — recent traces with a health strip + per-row latency/cost mini-bars; trace detail with a timeline (waterfall) or tree view and a conversation transcript of the prompt/response; replay diffs and A/B run comparison; costs & overview (error rate, latency p50/p95/p99, cost per day); alerts history.
• Cost tracking — aggregates input/output/cached/reasoning tokens and computes cost_usd per span and per trace from a model pricing table, overridable with --pricing.
• Export & audit — download the trace set as CSV (spreadsheet-friendly, UTF-8 BOM, formula-injection safe), JSON (metadata-only by default; ?raw=1 to include attributes), or a self-contained printable HTML audit report (format=html) with provenance, summary cards and hand-drawn SVG charts. GET /api/export/traces?format=csv|json|html; honours the active framework/status/session filters. Every response carries provenance/integrity (X-Lookspan-Export-Sha256, -Count, -Truncated).
• Alerts — get notified when a trace fails or exceeds a cost/token/duration threshold (toast + desktop notification + CLI + persisted history).
• Evaluation scores — attach metrics to a trace (POST /api/traces/:id/scores) from an LLM judge, an assertion, or by hand.
• Replay & LLM-as-judge — re-run a trace's captured prompt against the same or a different model and diff cost/latency/output, or have a judge model score the response 0–1. Needs a provider key (env, in-memory only).
• Datasets & experiments — collect prompts into a test set (seed from a trace or add by hand), run the whole set against a model in batch and score each output with the judge — aggregate cost/latency/score per run.
• Local SQLite (default) or external Postgres — versioned migrations. SQLite file at ~/.lookspan/lookspan.db by default. Passing a postgres://… URL to --db / LOOKSPAN_DB connects to that PostgreSQL server and persists data across restarts. See docs/CONFIGURATION.md → Postgres. Optional retention with --retention.
• Security — binds to 127.0.0.1 by default and only answers requests addressed to it (so a rebound domain gets nothing); non-loopback binds require --token; no cross-origin access unless you grant it with --cors-origin; server-side redaction of credential-looking attributes before storage.
• One-line CLI — npx lookspan starts the server and the dashboard with no global install.

Integrating your agents

OpenAI SDK (drop-in)

Wrap your client in one line — every model call is traced (no OTel, no proxy):

npm install @lookspan/openai

import OpenAI from 'openai';
import { observeOpenAI } from '@lookspan/openai';

const openai = observeOpenAI(new OpenAI());
await openai.chat.completions.create({ model: 'gpt-4o', messages });

Anthropic SDK (drop-in)

npm install @lookspan/anthropic

import Anthropic from '@anthropic-ai/sdk';
import { observeAnthropic } from '@lookspan/anthropic';

const anthropic = observeAnthropic(new Anthropic());
await anthropic.messages.create({ model: 'claude-sonnet-4-6', max_tokens: 1024, messages });

TypeScript / MCP

npm install @lookspan/mcp

import { wrapMcpClient, HttpSpanExporter } from '@lookspan/mcp';

const exporter = new HttpSpanExporter({ endpoint: 'http://127.0.0.1:3100/api/ingest' });
const { client } = wrapMcpClient(mcpClient, { exporter, agentId: 'my-agent' });

// Use it exactly as before — every callTool emits a tool_call span.
await client.callTool({ name: 'read_file', arguments: { path: '/tmp/foo.txt' } });
await exporter.flush();

Python (generic, LangGraph, CrewAI)

pip install lookspan            # + lookspan-langgraph / lookspan-crewai

from lookspan import LookspanClient
from lookspan_langgraph import LookspanCallbackHandler

client = LookspanClient(endpoint="http://127.0.0.1:3100/api/ingest")
handler = LookspanCallbackHandler(client=client, agent_id="my-agent")

result = graph.invoke({"messages": []}, config={"callbacks": [handler]})
client.flush()

OpenTelemetry (no SDK)

Point any OTel exporter at the standard OTLP endpoint:

export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=http://127.0.0.1:3100/v1/traces
# protobuf (the OTel default) and JSON are both accepted

More runnable examples in examples/.

Evaluate & replay

The drop-in SDKs can capture each call's prompt and reply (captureContent, off by default; secrets are scrubbed server-side). With that, Lookspan can close the loop from observe to improve — open a trace and use the Replay & judge panel, or call the API directly:

# Provider keys live in memory only — never written to the DB or logged.
LOOKSPAN_OPENAI_API_KEY=sk-... npx lookspan
#   ...or LOOKSPAN_ANTHROPIC_API_KEY / --openai-key / --anthropic-key

# Replay the captured prompt against another model and diff cost/latency/output
curl -X POST localhost:3100/api/traces/<id>/replay -H 'content-type: application/json' \
  -d '{"model":"gpt-4o-mini"}'   # omit "model" to re-run the original

# Score the response 0–1 with an LLM judge (stored as an "llm-judge" score)
curl -X POST localhost:3100/api/traces/<id>/judge -H 'content-type: application/json' \
  -d '{"metric":"correctness"}'

To enable replay/judge, explicitly pass { captureContent: true } to observeOpenAI / observeAnthropic only when prompts/outputs are safe to persist.

Datasets & experiments

Scale evaluation from one trace to a whole test set. Build a dataset (seed items from real traces or add them by hand), then run it against a model — each item is replayed and, optionally, scored by the judge, with aggregate cost/latency/score per run. Manage it all under Datasets in the dashboard, or:

# Create a dataset and add the captured prompt of a trace as an item
DS=$(curl -s -X POST localhost:3100/api/datasets -d '{"name":"regressions"}' -H 'content-type: application/json' | jq -r .dataset.id)
curl -X POST localhost:3100/api/datasets/$DS/items/from-trace -H 'content-type: application/json' -d '{"traceId":"<id>"}'

# Run the whole set against a model, judging each output
curl -X POST localhost:3100/api/datasets/$DS/run -H 'content-type: application/json' \
  -d '{"model":"gpt-4o-mini","judge":true,"metric":"correctness"}'

HTTP API

CLI options

npx lookspan [options]

Commands:
  install-desktop [options]  Add a one-click desktop launcher (options are baked in)
  uninstall-desktop          Remove it again

Options:
  -p, --port <port>        Port to listen on            (default: 3100)
      --host <host>        Host to bind to              (default: 127.0.0.1)
      --db <path|url>      SQLite path or postgres:// URL (default: ~/.lookspan/lookspan.db)
      --retention <dur>    Prune traces older than e.g. 7d, 24h, 30m
      --token <token>      Require Authorization: Bearer <token> on the API
      --pricing <file>     Custom model pricing table (JSON)
      --alert-error                Alert when a trace fails
      --alert-cost <usd>           Alert when a trace costs more than <usd>
      --alert-tokens <n>           Alert when a trace exceeds <n> tokens
      --alert-duration <ms>        Alert when a trace takes longer than <ms>
      --open               Open the dashboard in your browser
  -h, --help               Show help
  -v, --version            Show version

Every flag has a LOOKSPAN_* environment-variable equivalent (LOOKSPAN_PORT, LOOKSPAN_TOKEN, LOOKSPAN_PRICING, LOOKSPAN_ALERT_*, …). Replay & LLM-as-judge read LOOKSPAN_OPENAI_API_KEY / LOOKSPAN_ANTHROPIC_API_KEY (or --openai-key / --anthropic-key); these stay in memory and are never persisted.

See docs/CONFIGURATION.md for the complete flag + environment-variable reference, defaults, and examples.

How it compares

Lookspan isn't trying to be a full platform. It bets on being the zero-setup observability layer for the TypeScript/MCP agent stack, with the best first-five-minutes experience. See the ROADMAP.

Security

Lookspan binds to 127.0.0.1 (loopback) and requires no auth by default — right for local use. If you expose it (--host 0.0.0.0), protect it with a token:

LOOKSPAN_TOKEN=my-token npx lookspan --host 0.0.0.0
# /api/* and /v1/* then require Authorization: Bearer my-token (/api/health is exempt).

The collector also redacts before persisting, so telemetry does not drag secrets into the database:

• values of credential-looking keys (authorization, api_key, token, secret, password, cookie…) in input and attributes;
• secret-looking values wherever they appear — including in output — for provider keys, GitHub and Slack tokens, AWS access key ids and JWTs, so a key pasted into a prompt is caught even under an innocent key name.

The scan walks twelve levels deep, which covers the nesting real agent payloads reach (an OpenAI tool call puts its arguments six levels down). Anything deeper is replaced wholesale rather than stored unscanned: if the scan cannot prove a subtree is clean, it does not keep it.

Development

This is an npm-workspaces monorepo. packages/ holds internal libraries, apps/ the dashboard, python/ the standalone Python SDKs.

git clone https://github.com/JoniMartin27/lookspan.git
cd lookspan
npm install
npm run dev        # API on :3100, dashboard with hot-reload on :5173
npm run ci         # typecheck + lint + test + build + smoke-test the published bundle

The Python SDKs are separate packages with their own test suites (CI runs all three):

cd python/lookspan-core && pip install pytest . && pytest -q

lookspan-crewai installs with --no-deps for tests — CrewAI itself is only imported inside attach(), so the span-building paths are covered with stand-in events instead of a heavy dependency tree.

Contributions welcome — see .github/CONTRIBUTING.md. Release process in docs/PUBLISHING.md. Security policy: SECURITY.md.

License

MIT — Copyright (c) 2026 Jonathan Martin. See LICENSE.

Lookspan is part of Fervon, the studio behind a portfolio of open-source developer tools (Trace, InferBench, ClaudeScope, Launchpad and more). The Fervon forge identity — carbon surfaces, ember/amber accents — is applied across the landing, the dashboard and the docs site.