Langfuse vs Vigil
What Langfuse does and what Vigil does, side by side. Every Langfuse fact below was read on Langfuse's own pages on 2026-10-09 and links to the page it came from; a feature its pages do not mention is marked “not stated”, never “no”. Where Langfuse is stronger, this page says so.
What Langfuse is, and what Vigil is
Langfuse, in its own words: “Trace, evaluate, and improve AI agents with one open platform.”langfuse.com. SDK, outside the request path: "from langfuse.openai import openai"; traces are sent asynchronously.
Vigil is a proxy for AI API calls. Change one base URL, add one header, and every call is logged with its cost, tokens, latency, errors and agent, then forwarded unchanged. Where the provider supports explicit prompt caching, Vigil can place the cache markers itself when optimisation is switched on. It stores no prompt text and caches no responses.
- Langfuse says: "No, Langfuse is not an LLM Proxy." It is "an asynchronous observability layer", to avoid "introducing a single point of failure".langfuse.com
Feature by feature
The Langfuse column was read on Langfuse's own pages on 2026-10-09; each cell links to the page. “Not stated” means the pages read do not mention it, which is not the same as “no”. The Vigil column is read from Vigil's code and docs.
| Feature | Langfuse | Vigil |
|---|---|---|
| Integration | SDK and OpenTelemetry; "Fully async requests, meaning Langfuse adds almost no latency."langfuse.com | Proxy. Change the base URL and add the X-Vigil-Key header; no SDK. |
| Open source | "MIT licensed, except for the ee folders."github.com | No. Vigil is not open source. |
| Hosting | Cloud in US, EU, Japan and a HIPAA region; self-host by Docker Compose, Helm or Terraform.langfuse.com | Cloud only: a Cloudflare Worker in front of the provider, a dashboard on Vercel, data in Supabase (Frankfurt, EU). |
| Cost per call | Yes, inferred from the model name and ingested usage: "Langfuse works them out from the generation’s model parameter"; custom models need a definition.langfuse.com | Yes: cost, tokens, latency, status and agent on every call, from a rate registry where every rate carries its source, date and provenance; a dash where no rate is published. |
| Providers | "100+ integrations". Cost needs a model definition, not a proxy route.langfuse.com | 17 providers routed by segment. |
| Prompt caching | Reports cache-read tokens from the usage your SDK sends. It is not in the request path, so it cannot create a hit.langfuse.com | Creates cache hits: Anthropic, Google Vertex, Mistral get cache markers placed by Vigil when optimisation is on; AWS Bedrock when the request carries an API key rather than a SigV4 signature. On other providers Vigil records the provider’s own caching. |
| Response caching | None for LLM responses; prompts fetched from prompt management are cached client-side.langfuse.com | No. Vigil does not cache responses. |
| Routing and fallbacks | None: "Langfuse does not sit between your application and the LLM provider’s API".langfuse.com | Model routing is in beta: Sends a conversation to a cheaper model from the same provider — one tier down — when its opening's size and tools are in a band the agent's routing level covers and the cheaper model can take the request; every call of the conversation gets the same answer. No provider fallbacks or load balancing. |
| Prompt management | Yes: "storing, versioning, and retrieving prompts", labels, rollbacks.langfuse.com | No prompt management or versioning. |
| Evals | Yes: LLM-as-a-judge, scores, annotation queues, datasets and experiments.langfuse.com | No evaluations. A score can be sent for a call through the feedback endpoint. |
| Tracing | "hierarchical traces capture every LLM call, tool invocation, and retrieval step".langfuse.com | One record per call. No spans or traces. |
| Alerts | Metric alerts to Slack, webhooks and GitHub Actions; 2, 20, 50 or 100 per plan.langfuse.com | Cost spikes, failures and findings are flagged in the dashboard. Nothing is sent by email, Slack or push. |
| Prompt injection and PII | No detection. Masking functions you write "redact sensitive information before trace data leaves your application".langfuse.com | Flags likely prompt injections and personal data in prompts, on every plan. Nothing is blocked or redacted. |
| Budgets and rate limits | None on LLM spend (not in the request path); ingestion rate limits per plan.langfuse.com | No spend caps or rate limits on your provider traffic. Optimisation can be switched off per agent. |
| Seats | 2 users on Hobby, unlimited from Core; project-level RBAC with the Teams add-on or Enterprise.langfuse.com | Seats per plan: 1 on Free, 3 on Plus, unlimited from Pro. Teammates are invited; there are no role tiers. |
| Compliance | "covered by SOC 2 Type II and ISO 27001 audits"; GDPR; a HIPAA-ready region.langfuse.com | No SOC 2 or ISO 27001 report. A DPA is published. |
| Data residency | US (us-west-2), EU (eu-west-1, Ireland), Japan (ap-northeast-1).langfuse.com | Account and telemetry stored in the EU (Supabase, Frankfurt). The proxy runs on Cloudflare’s edge. No prompt text or responses are stored. |
Where Langfuse is stronger
- Tracing with spans across every LLM call, tool invocation and retrieval step.
- Evaluation: LLM-as-a-judge, annotation queues, datasets and experiments.
- Prompt management with versioning, labels and rollbacks.
- Open source under MIT, self-hostable, with SOC 2 Type II, ISO 27001 and a choice of three regions.
- Outside the request path, so it cannot add latency or fail a request.
Where Vigil differs
- It creates prompt-cache hits rather than reporting them, on the providers whose caching lets a proxy place markers, and measures what caching would have saved on the others.
- Every rate it prices with carries its source, its retrieval date and a provenance label, and a cost it cannot stand behind is a dash, not a number.
- Monitoring is not gated by plan: every agent, every provider, on Free. Paid plans switch optimisation on.
- It does less: no response cache, no evals, no prompt management, no traces, no alerts outside the dashboard.
Pricing, side by side
Langfuselangfuse.com
- Hobby Free: 50k units a month (a unit is a trace, an observation or a score), 30 days of data access, 2 users
- Core $29/mo: 100k units, $8 per further 100k; 90 days of data access; unlimited users
- Pro $199/mo: 100k units, 3 years of data access, SOC 2 and ISO 27001 reports; Teams add-on $300/mo for SSO and RBAC
- Enterprise $2,499/mo: 100k units, audit logs, SCIM, SLAs; custom volume pricing
- Self-host Free: "all core Langfuse features for free without any limitations"; enterprise edition by contract
Vigil vigil.tools/pricing
- Free $0: every agent monitored; 10,000 logged calls a month, 7 days of history; every optimisation On during your 7-day trial, as Plus, then Shadow only
- Plus $29/mo: every optimisation On up to $100 saved a billing month, then Shadow until the next; 500,000 logged calls, 30 days
- Pro $99/mo: every optimisation On up to $500 saved a billing month, then Shadow until the next; 2,000,000 logged calls, 90 days
- Supreme $299/mo: every optimisation On up to $2,000 saved a billing month, then Shadow until the next; Unlimited logged calls, 365 days
- Enterprise Custom: by contract
Monthly, excl. VAT. Monitoring is free on every plan; what a paid plan buys is a higher limit on the money Vigil saves you per billing month before optimisations go to Shadow.
Other comparisons: Helicone vs Vigil · Portkey vs Vigil · LiteLLM vs Vigil · OpenRouter vs Vigil