> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tracelane.dev/llms.txt
> Use this file to discover all available pages before exploring further.
# How Tracelane compares
> A sourced feature matrix against LiteLLM, Portkey, Bifrost, Helicone, Langfuse, LangSmith, Braintrust and Arize Phoenix — sourced from each vendor's own pages.
Four tables — gateway, observability, evals, trust — built the same way as our
[gateway-overhead benchmark](/benchmarks): every number and every claim is sourced, and a
`—` is not a loss, it's a "we didn't find it," which is not the same thing.
## Method
* **Every competitor cell that asserts a capability links to a numbered source**, fetched
from the vendor's own page — docs site, pricing page, or public GitHub repository — on
**2026-09-07**. No third-party blogs, no re-published "comparison" pages from a fourth
party. A bare "Not documented" or "—" is a cell where we found nothing to cite.
* **Every Tracelane cell cites a code path, a public doc** ([`/pricing`](/pricing),
[`/benchmarks`](/benchmarks)) **or a dated production proof** — never a spec or a
marketing draft.
* **A cell marked "Not documented" means we did not find that capability documented on
the vendor's own site as of the fetch date — it does not mean the vendor lacks it.**
Vendors change their products continuously; if something below is stale, [open an
issue](https://github.com/tracelane/tracelane/issues) and we'll fix it.
* **"Yes" for Tracelane always means shipped and either deployed to production or
reproducible from the public repo** — not planned, not "in the design doc." Where a
capability is partial or gated by plan, the cell says so and the FAQ or a footnote
explains the boundary.
See also: [Tracelane vs LiteLLM](/comparisons/vs-litellm) ·
[Tracelane vs Portkey](/comparisons/vs-portkey) ·
[Tracelane vs Helicone](/comparisons/vs-helicone) ·
[Tracelane vs LangSmith Engine](/comparisons/vs-langsmith-engine) ·
[the gateway-overhead benchmark](/benchmarks).
## Gateway
| Capability | Tracelane | LiteLLM | Portkey | Bifrost | Helicone | Langfuse | LangSmith | Braintrust | Arize Phoenix |
| - | - | - | - | - | - | - | - | - | - |
| Hot-path runtime | Rust \[T1] | Python proxy (the build measured in [/benchmarks](/benchmarks)); a "Rust core" is announced on GitHub with no public build as of 2026-09-07 \[1] \[6] | TypeScript/Node.js \[2] | Go \[3] | Rust (`Helicone/ai-gateway` repository) \[22] | N/A — no gateway | N/A — no self-hostable gateway runtime documented \[5] | N/A — no gateway | N/A — no gateway |
| Measured gateway overhead (p50) | **2 ms** — AIGatewayBench, CCX23, 2026-09-07 \[T2] | Python proxy: 14–15 ms p50, same bench run \[T2]; the announced Rust gateway has no public build, and its self-reported "8 ms P95 at 1k RPS" is not independently measured \[6] | 3–4 ms p50, same bench run \[T2] | 2 ms p50, same bench run \[T2] | Not published \[4] | — | — | — | — |
| Providers supported | 191 \[T3] | 140+ upstream vendors / \~1,892 models (its own count) \[7] | Claims 1,600+ LLMs \[8]; source code lists \~75 provider modules \[9] — see "Portkey's provider count" below | 23+ upstream vendors (its own count) \[3] | Multiple — count not documented \[4] | — | — | — | — |
| BYOK, 0% token markup | Yes \[T3] | Yes — per-key `api_key` in config \[10] | Yes — virtual keys with your own provider keys \[8] | Yes — virtual keys \[3] | Yes — routes through your keys \[4] | — | Partial — lists an "LLM Gateway" to "control agent model calls"; BYOK not documented on that page \[5] | — | — |
| Response caching | Exact-match caching in production. Embedding-similarity caching is built but not enabled on the hosted product \[T4] | Yes — exact + semantic, 8+ backends \[11] | Yes — simple + semantic \[12] | Yes — semantic caching \[3] | Partial — caching listed as a feature; exact vs. semantic not documented \[4] | Not documented | Not documented | Not documented | Not documented |
| Rate limits + budgets per key | Yes — per-key RPM and budget, 429 on overage \[T3] | Yes — per key/team/user RPM + TPM \[10] | Yes — hourly/daily/per-minute on requests or tokens \[12] | Yes — budget management \[3] | Yes — rate limiting feature \[4] | Not documented | Not documented | Not documented | Not documented |
| Guardrails on the hot path | Yes — **R1–R8**, deterministic rule-based rails inline at the gateway (cost/budget, secrets+PII redaction, tool pinning, lethal-trifecta, format, system-prompt-leak, topic/competitor, prompt-injection); every verdict is written to the audit ledger. R1/R3/R4/R8 free on every deployment incl. OSS self-host; R2/R5/R6/R7 (data-governance + quality) on Team and above \[T5] | Yes — PII masking, prompt-injection detection, secret detection \[1] | Yes — "50+ AI Guardrails" \[2] | Partial — enterprise only, custom plugin \[3] | Not documented \[4] | Not documented | Not documented | Not documented | Not documented |
| Failover / routing | Partial — model→provider map fails closed (400 `unroutable_model`, no silent default); multi-key routing built \[T6] | Yes — fallback chains, context-window fallback, cooldown \[13] | Yes — automatic fallbacks, circuit breaker \[2] | Yes — automatic failover, load balancing \[3] | Not documented \[4] | Not documented | Not documented | Not documented | Not documented |
| OpenAI-compatible endpoint | Yes — `/v1/chat/completions` \[T1] | Yes \[1] | Yes \[2] | Yes \[3] | Yes \[4] | — | Partial — gateway is managed \[5] | — | — |
| Anthropic-native `/v1/messages` | Yes — proven on production with a real Claude Code session \[T7] | Partial — routes Anthropic models, but via OpenAI-shaped format \[10] | Not documented \[2] | Not documented \[3] | Yes — direct integration listed \[4] | — | — | — | — |
## Observability
| Capability | Tracelane | LiteLLM | Portkey | Bifrost | Helicone | Langfuse | LangSmith | Braintrust | Arize Phoenix |
| - | - | - | - | - | - | - | - | - | - |
| OTel-native ingest (OTLP) | Yes — the OTLP decode path preserves `parent_span_id` by design \[T8] | Partial — emits OTel spans for guardrail violations \[1]; native ingest not confirmed | Partial — OTel emission, not primary ingest \[8] | Partial — Prometheus + distributed tracing \[3] | Not documented \[4] | Yes — OTel-native, primary wire format \[14] | Partial — proprietary format primary \[5] | Partial — proprietary format primary \[15] | Yes — native OTel \[16] |
| Full-fidelity capture (no default sampling) | Yes — no default sampling; capture on in the published benchmark \[T2] | Not documented | Not documented | Not documented | Not documented | Yes — all LLM and non-LLM calls \[14] | Yes — "full visibility" \[5] | Not documented | Not documented |
| Trace tree + tool spans | Yes \[T9] | Not documented | Not documented | Not documented | Not documented | Yes — traces include tool calls \[14] | Yes — step-by-step agent view \[5] | Yes — tracing with spans \[15] | Yes — every step incl. tool calls \[16] |
| Sessions / threads | Yes — `/sessions`, keyed on `gen_ai.conversation.id` \[T10] | Not documented | Not documented | Not documented | Yes — sessions listed \[4] | Yes — multi-turn sessions \[14] | Yes \[5] | Yes — "Loop thread" \[15] | Not documented |
| Multi-agent swimlanes / agent graph | Yes \[T9] | Not documented | Not documented | Not documented | Not documented | Yes — agent graphs \[14] | Yes — agent step view \[5] | Not documented | Not documented |
| Full-text trace search | Yes \[T9] | Not documented | Not documented | Not documented | Partial — HQL query language \[4] | Not documented \[14] | Yes — production monitoring + search \[5] | Not documented | Not documented |
| Cost per request / model | Yes \[T11] | Yes — per key/team/user tracking \[1] | Yes — usage analytics \[2] | Yes — budget management \[3] | Not documented | Yes \[14] | Yes \[5] | Yes \[15] | Not documented |
| In-app AI assistant over your traces | Yes — "Ask Tara", a natural-language query surface over your own traces, deployed to production \[T12] | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| MCP server for trace access | Yes — `@tracelanedev/mcp` (published on npm, `0.3.0`), read-only and tenant-scoped: it reads through the gateway's own `/v1/*` routes with the caller's own bearer token \[T13] | Not documented | Not documented | Yes — MCP for external tool integration \[3] | Not documented | Not documented | Not documented | Not documented | Not documented |
| Claude Code / coding-agent integration | Yes — two paths: an OTLP flight recorder that decodes Claude Code's native span shape, and an Anthropic-native `/v1/messages` proxy, both proven on production with real Claude Code sessions \[T14] | Not documented — routing Claude models is not a coding-agent integration | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Shareable trace links | Yes — a public `/s/` page carrying the audit-ledger badge \[T15] | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
## Evals
| Capability | Tracelane | LiteLLM | Portkey | Bifrost | Helicone | Langfuse | LangSmith | Braintrust | Arize Phoenix |
| - | - | - | - | - | - | - | - | - | - |
| Datasets | Yes — proven on production \[T16] | Not documented | Not documented | Not documented | Yes — listed \[4] | Yes \[14] | Yes \[5] | Yes \[15] | Yes \[16] |
| Experiments | Yes — proven on production \[T16] | Not documented | Not documented | Not documented | Not documented | Yes \[14] | Yes \[5] | Yes \[15] | Yes \[16] |
| LLM-as-judge online evals with sampling | Yes — 11.11% sampling against a 10% configured rate, proven on production \[T17] | Not documented | Not documented | Not documented | Not documented | Yes — LLM-as-judge evaluators \[14] | Yes — online evaluations \[5] | Yes \[15] | Yes — LLM-as-judge \[16] |
| Annotation / review queues | Yes — proven on production \[T18] | Not documented | Not documented | Not documented | Not documented | Yes — annotation queues \[14] | Yes — annotation queues \[5] | Not documented | Yes — human review \[16] |
| CI eval gate (GitHub Action / CLI) | Yes — `@tracelanedev/cli` (published on npm, `0.3.0`), proven on production \[T19] | Not documented | Not documented | Not documented | Not documented | Partial — webhooks + GitHub Actions for alerts \[14] | Not documented | Not documented | Not documented |
| Prompt management + playground | Yes \[T20] | Not documented | Yes — prompt templates across tiers \[12] | Not documented | Yes — prompts listed as a feature \[4] | Yes — versioning + deployment without code changes \[14] | Yes \[5] | Yes — playground \[15] | Yes — Prompt IDE \[16] |
## Trust
| Capability | Tracelane | LiteLLM | Portkey | Bifrost | Helicone | Langfuse | LangSmith | Braintrust | Arize Phoenix |
| - | - | - | - | - | - | - | - | - | - |
| Tamper-evident / hash-chained audit ledger | Yes — Merkle-batched hash chain + per-tenant Ed25519 signing (keys held by Tracelane; the Rekor anchor is the part we cannot forge), included on every tier \[T21] | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Third-party verifiable offline | Yes — Sigstore Rekor v2 inclusion proof on anchored batches, best-effort; `tlane verify --tenant-pubkey` needs no call back to us \[T21] | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented | Not documented |
| Self-host in one binary / Compose | Yes — single-node Docker Compose, ClickHouse-backed \[T3] | Yes — Docker Compose \[1] | Yes — `npx @portkey-ai/gateway` or Docker \[2] | Yes — Docker/self-host \[3] | Yes — hosted or self-host \[22] | Yes — Docker Compose + Kubernetes \[14] | Partial — Enterprise BYOC/self-hosted SKU \[5] | Partial — Enterprise on-premise \[15] | Yes — Docker, Kubernetes \[16] |
| Open-source license | Apache 2.0 \[T3] | MIT \[1] | MIT \[2] | Apache 2.0 \[3] | Apache-2.0 (`Helicone/helicone`); GPL-3.0 (`Helicone/ai-gateway`) \[22] | MIT \[14] | None for OSS; Enterprise self-hosted SKU is proprietary \[5] | Not confirmed | ELv2 (not Apache/MIT) \[16] |
| SOC 2 / ISO 27001 (published report) | **No** \[T22] | Not documented \[7] | Yes — SOC2 + ISO27001 + GDPR + HIPAA; reports via paid plans \[17] | Not documented \[3] | Yes — SOC2 Type II at Team tier (\$799/mo) \[18] | Yes — SOC2 Type II + ISO27001 at Pro tier ($199/mo){/* pricing-guard: allow $199 competitor-price \*/} \[19] | Partial — Trust Center referenced, no detail \[5] | Yes — SOC2 Type II, all tiers \[20] | Not documented \[16] |
| Data residency / regions | No customer-configurable data residency is documented on our pricing page as of 2026-09-07 \[T22] | Self-host anywhere; no hosted multi-region \[1] | VPC hosting for Enterprise \[17] | Not documented \[3] | Not documented \[4] | US, EU, JP (vendor abbreviation) \[19] | US or EU cloud; BYOC for custom \[21] | Not documented \[20] | Not documented \[16] |
| Free tier | 1 GB ingest / 0.25 GB hot / 100 series / 10 scan-units per mo, 60 RPM, 1 seat, no commercial use, rule-based, observe-first predictive layer \[T3] | Free core; cloud pricing not public \[7] | 10K logs/mo, 3-day retention, 1 user \[12] | Not documented \[3] | 10K requests/mo, 1 GB, 1 seat \[18] | 50K units/mo, 30-day data access, 2 users \[19] | 5K traces/mo, 14-day retention, 1 seat \[21] | $0 + $10 credits/mo, 1 GB data, 14-day retention \[20] | 2 Phoenix Cloud instances free \[16] |
| Pricing model summary | Six usage meters (ingest/hot/series/query/cold/evals) + tier price; $29 Builder / $229 Team / $799 Business / $2,499+ Enterprise \[T3] | Free core; Enterprise sales-only \[7] | $49/mo Production; Enterprise custom{/* pricing-guard: allow $49 competitor-price \*/} \[12] | Not published \[3] | $79/mo Pro; $799/mo Team \[18] | $29/mo Core; $199/mo Pro; $2,499/mo Enterprise{/* pricing-guard: allow $199 competitor-price \*/} \[19] | $39/seat/mo Plus; Enterprise custom{/* pricing-guard: allow $39 competitor-price \*/} \[21] | $249/mo Pro; Enterprise custom{/* pricing-guard: allow $249 competitor-price \*/} \[20] | Cloud pricing not public \[16] |
## Portkey's provider count
Portkey's homepage says **"1,600+ LLMs."** Its own GitHub repository's `src/providers`
directory, read the same day, lists roughly **75 provider modules** — one module (e.g.
`openrouter`) can route to many models, which partly explains the gap between a provider
count and a model count. But Portkey's own pages are not internally consistent: across
different pages we found **three different figures — 250+, 1,600+, and 3,000+** — for what
should be one number. We're not asserting which is "the" real count; we're reporting that
the vendor's own marketing disagrees with itself. Read their repository yourself before
treating any of the three as ground truth: `github.com/portkey-ai/gateway/tree/main/src/providers`.
For our own "205 providers" claim: it is the count of `crates/gateway/providers.tsv`
rows (185) plus the 6 native adapters, held to that file by a guard in CI — count it
yourself.
## Honest gaps
Things competitors publish that we do not have. In rough order of how likely they are to
block a deal on day one:
1. **SOC 2 Type II report (published).** Portkey, Helicone (Team+), Langfuse (Pro+),
Braintrust (all tiers) and LangSmith (Trust Center) all have or reference a published
SOC 2. **Not yet.**
2. **Data residency / multi-region.** Langfuse offers US, EU, JP (its own abbreviation); LangSmith offers US
or EU; Portkey and Helicone offer VPC hosting on Enterprise. Tracelane is single-region
and residency is not customer-configurable. **Not yet.**
3. **HIPAA / BAA.** Documented by Portkey, Langfuse, Helicone, Braintrust and LangSmith,
each on their higher tiers. **Not yet.**
4. **SAML SSO, documented as a self-serve feature.** Tracelane authenticates through
WorkOS, but SAML is not documented as an available feature on any public Tracelane
page today. **Not yet.**
5. **Contractual SLA / uptime guarantee.** `apps/docs/pricing.mdx` states plainly:
*"Tracelane does not offer a contractual uptime SLA or service credits."* Langfuse and
Braintrust document one on their Enterprise tiers. **Not yet.**
6. **Slack / PagerDuty alert integrations, documented.** LangSmith documents webhooks +
PagerDuty; Langfuse documents Slack/GitHub Actions/webhooks; Helicone documents alerts.
Tracelane's burn-rate alerts on paid plans are not documented as integrating with a
specific paging tool. **Not yet.**
None of these are silent gaps — every one is stated as a limit in a public Tracelane doc
(mostly [pricing](/pricing)) rather than only showing up when you ask.
## Sources
### Tracelane (code-verified)
* **\[T1]** `crates/gateway/src/admission.rs` + `crates/gateway/src/server/chat.rs` — the Rust gateway hot path
* **\[T2]** [`/benchmarks`](/benchmarks) — AIGatewayBench CCX23, 2026-09-07: Tracelane 2/4/5 ms, Portkey 3–4/5–7/8–9 ms, Bifrost 2/2/3 ms, LiteLLM Python 14–15/34–36/47–52 ms
* **\[T3]** [`/pricing`](/pricing) — 205 providers, BYOK 0%, plan limits, Apache 2.0, Docker Compose, free-tier limits
* **\[T4]** exact-match caching built and proven; the semantic tier is built but not enabled on the hosted product
* **\[T5]** `crates/gateway/src/guardrail/rails/` (`r1_cost.rs` … `r8_injection.rs`) + [`/pricing`](/pricing) FAQ ("What inline guardrails run on every request?") for the free-vs-Team+ split
* **\[T6]** `crates/gateway/src/providers/mod.rs` — model→provider routing fails closed to `unroutable_model`
* **\[T7]** Anthropic-native `POST /v1/messages`, deployed and proven on production 2026-09-06 with a real Claude Code session (3 proxied spans, tokens and cost present)
* **\[T8]** `crates/shared/src/otlp/decode.rs` — the `parent_span_id must survive` comment and the OTLP decode path
* **\[T9]** multi-agent swimlanes and full-text trace search, both shipped and deployed 2026-09-06
* **\[T10]** `/v1/sessions` and `/v1/sessions/{id}/traces` handlers in `crates/gateway/src/server.rs`
* **\[T11]** `crates/gateway/src/spend.rs` — cost-per-request tracking
* **\[T12]** "Ask Tara," the in-app natural-language query surface, shipped and deployed 2026-09-06; distinct from the MCP server (below)
* **\[T13]** `apps/mcp/src/reader.ts`, `GatewayReader` reads `/v1/*` with the caller's own bearer, tenant isolation proven against production ClickHouse; `@tracelanedev/mcp@0.3.0` published to npm 2026-09-07, confirmed by `npm view @tracelanedev/mcp version`
* **\[T14]** the Claude Code flight recorder (OTLP decoder aliases + cost backfill, deployed and proven 2026-09-06) and the Anthropic-native proxy (see \[T7])
* **\[T15]** the public `/s/` page, built and deployed 2026-09-06
* **\[T16]** `crates/gateway/src/dataset_routes.rs` and `crates/gateway/src/experiment_routes.rs` — datasets and experiments, proven on production
* **\[T17]** online evals with sampling, proven on production 2026-08-29, achieved 11.11% against a configured 10% rate
* **\[T18]** annotation queues, proven on production 2026-08-29
* **\[T19]** the CI eval gate and `@tracelanedev/cli`, proven on production 2026-08-30; `@tracelanedev/cli@0.3.0` published to npm 2026-09-07
* **\[T20]** prompt management, playground and trace search, shipped 2026-09-06
* **\[T21]** [`/pricing`](/pricing) audit-ledger description and [`/audit-ledger`](/audit-ledger) — Merkle-batched hash chain, Ed25519 per-tenant signing (live), best-effort Sigstore Rekor v2 inclusion proof on anchored batches
* **\[T22]** [`/pricing`](/pricing) — no SOC 2 / ISO 27001 mentioned, no SLA clause; single-region; retention is set by plan and is not customer-configurable
### Competitors (all fetched 2026-09-07 from vendors' own pages)
1. [github.com/BerriAI/litellm](https://github.com/BerriAI/litellm) — "Rust core with Python SDK. Call 100+ LLM APIs in OpenAI (or native) format"; MIT license
2. [github.com/portkey-ai/gateway](https://github.com/portkey-ai/gateway) — "routes to 1,600+ LLMs across 45+ \[vendors]"; "50+ AI Guardrails"; MIT license
3. [github.com/maximhq/bifrost](https://github.com/maximhq/bifrost) — "high-performance AI gateway unifying access to 23+ \[vendors]"; Apache 2.0
4. [helicone.ai](https://helicone.ai) — "10,000 free requests, 1 GB storage, 1 seat, 1 organization"
5. [langchain.com/langsmith](https://www.langchain.com/langsmith) — "Control agent model calls" under the LLM Gateway heading
6. [github.com/BerriAI/litellm](https://github.com/BerriAI/litellm) — "8ms P95 latency at 1k RPS" (self-reported, not independently measured)
7. [litellm.ai](https://www.litellm.ai) — "One OpenAI-compatible API to 140+ \[vendors] and 1,800+ models"
8. [portkey.ai](https://portkey.ai) — "Portkey lets you access 1,600+ LLMs via a unified API"
9. [github.com/portkey-ai/gateway/tree/main/src/providers](https://github.com/portkey-ai/gateway/tree/main/src/providers) — directory listing, \~75 provider-module entries counted by hand; see "Portkey's provider count" above
10. [docs.litellm.ai/docs/proxy/configs](https://docs.litellm.ai/docs/proxy/configs) — BYOK via `litellm_params.api_key`; per-deployment RPM/TPM
11. [docs.litellm.ai/docs/caching/all\_caches](https://docs.litellm.ai/docs/caching/all_caches) — "similarity threshold for cache hits, 0 == no similarity, 1 = exact matches"; 8+ backend options
12. [portkey.ai/pricing](https://portkey.ai/pricing) — "$49/month Production"{/* pricing-guard: allow $49 competitor-price \*/}; "100k logs/month"; semantic caching; "10k logs free / 3-day retention"; SOC2 + ISO27001
13. [docs.litellm.ai/docs/proxy/reliability](https://docs.litellm.ai/docs/proxy/reliability) — "If a call fails after num\_retries, LiteLLM falls back to another model group"
14. [langfuse.com](https://langfuse.com) — OTel-native; "US, EU, or Japan"; free 50K units / 2 users / 30-day; annotation queues; agent graphs
15. [braintrust.dev](https://braintrust.dev) — "Tracing, evaluation, and storage infrastructure"; SOC2 Type II all tiers; "$249/mo Pro"{/* pricing-guard: allow $249 competitor-price \*/}; "14-day retention Starter"
16. [arize.com/phoenix](https://arize.com/phoenix) — "ELv2 licensed"; "native OpenTelemetry support"; "2 Phoenix Cloud instances for free"; Docker/Kubernetes self-host
17. [portkey.ai/pricing](https://portkey.ai/pricing) — "SOC2, ISO27001, GDPR, HIPAA Compliance Certificates"; VPC hosting Enterprise
18. [helicone.ai/pricing](https://helicone.ai/pricing) — "SOC-2 Type II at Team tier (\$799/mo)"; "10,000 free requests, 1 seat"
19. [langfuse.com/pricing](https://langfuse.com/pricing) — "SOC2 Type II + ISO27001 reports at Pro ($199/mo)"{/* pricing-guard: allow $199 competitor-price \*/}; "US, EU, or Japan"; "Hobby: 50K units/30-day/2 users"
20. [braintrust.dev/pricing](https://www.braintrust.dev/pricing) — "SOC 2 Type II" (all tiers); "$249/mo Pro"{/* pricing-guard: allow $249 $10 competitor-price */}; "14-day retention Starter"; "$0 + \$10 credits/mo free"
21. [langchain.com/pricing](https://langchain.com/pricing) — "$39/seat/mo Plus"{/* pricing-guard: allow $39 competitor-price \*/}; "5K traces/mo free"; "14-day retention base"; "US or EU cloud"
22. [github.com/Helicone/helicone](https://github.com/Helicone/helicone) and [github.com/Helicone/ai-gateway](https://github.com/Helicone/ai-gateway) — per-repository `LICENSE` files; self-hosting documented in the README
***
*Matrix compiled 2026-09-07. Every competitor source above was fetched that day from an
official vendor page — a docs site, a pricing page, or a public GitHub repository. If
anything here is inaccurate or out of date, please [open an
issue](https://github.com/tracelane/tracelane/issues) and we'll correct it.*
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.