headroom vs new-api
Two of the top projects in Gateways and observability, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
headroom
Compresses tool outputs, logs and history before they reach the LLM
new-api
Gateway that unifies LLM providers behind OpenAI, Anthropic and Gemini APIs
Verdict
headroom ranks higher, 92 to 91. Of the 4 score parts we measure for both, headroom leads on 2, new-api on 1 and 1 is even.
- Adoptionheadroom leads by 4 points, 97 to 93: close to even.
- FreshnessEven at 100.
- Maintenanceheadroom leads by 9 points, 95 to 86.
- Agent-readynew-api leads by 25 points, 70 to 45: a wide gap.
Pick headroom if you are
Developers cutting LLM input token costs in apps or coding agents
- Seeded, offline benchmark: 21–57% fewer input tokens across four agent scenarios
- One-command `headroom wrap` for many coding agents, reversible with `headroom unwrap`
Pick new-api if you are
Teams and operators sharing and metering access to multiple LLM providers
- Converts between OpenAI, Anthropic Messages and Gemini request formats via RelayKit
- Channel priorities, weights, retries and multiple upstream keys per channel
| What we compare | headroom | new-api |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 97, widely used | 93, widely used |
| Momentum | Not scored for this kind | Not scored for this kind |
| Freshness | 100, active | 100, active |
| Maintenance | 95, healthy | 86, healthy |
| Agent-ready | 45, minimal | 70, partly |
| Facts from GitHub and the README | ||
| Stars | 79.4k | 51.8k |
| License | Apache-2.0 (permissive) | AGPL-3.0 (copyleft) |
| Last commit | Nov 2026 | Nov 2026 |
| Last release | Nov 2026 | Nov 2026 |
| Language | Python | Go |
| Docker | Yes | Yes |
| GPU | Not needed | Not needed |
| arm64 or Apple Silicon | Mentioned | Mentioned |
headroom
Headroom is a Python library, local proxy and CLI that shrinks agent context before it is sent to an LLM provider. It routes JSON, source code and text to separate compressors and keeps originals locally so the model can retrieve them. `headroom wrap` configures coding agents such as Claude Code, Codex, Aider and Cline to go through the proxy.
Who it is for: Developers cutting LLM input token costs in apps or coding agents
Strengths
- Seeded, offline benchmark: 21–57% fewer input tokens across four agent scenarios
- One-command `headroom wrap` for many coding agents, reversible with `headroom unwrap`
- Compression runs locally; reported 0.21 ms p50 on a 10K-token JSON result
- Originals cached locally and retrievable by the model via `headroom_retrieve`
Weaknesses
- Little savings on prose or already-dense output; gains depend on repetitive payloads
- Cursor and ZCode need manual base-URL setup; Cortex Code is library-only
- Third-party OpenAI-compatible providers need an extra flag or the key goes to OpenAI
- Output token reduction is off by default and reported as an estimate
- no GPU
- Docker + Compose
- Needs Python 3.13 (for uv tool install), Serena (installed by headroom wrap), Kompress-v2-base model from Hugging Face
- Models: OpenAI, Anthropic, Bedrock, Gemini, OpenAI-compatible providers
- port 8787
new-api
New API is a Go and Gin gateway that connects upstream providers such as OpenAI, Anthropic, Gemini, Azure OpenAI, AWS Bedrock, Vertex AI, DeepSeek and Qwen. It exposes them through OpenAI Chat/Responses, Anthropic Messages and Gemini endpoints. A web console handles channel routing, users, quotas, pricing and usage logs. It runs on SQLite by default, or on MySQL or PostgreSQL with optional Redis.
Who it is for: Teams and operators sharing and metering access to multiple LLM providers
Strengths
- Converts between OpenAI, Anthropic Messages and Gemini request formats via RelayKit
- Channel priorities, weights, retries and multiple upstream keys per channel
- Per-user quotas, subscriptions, usage logs and expression-based pricing
- Starts on SQLite with one container; Compose file adds PostgreSQL and Redis
Weaknesses
- AGPL-3.0 with extra terms requiring attribution and a visible project link in modified UIs
- Protocol conversion is not exact; tools and fields may not map across protocols
- Multi-node setups need shared database, secrets and Redis, or rate limits count per node
- Public or resale use brings filing and compliance duties the README leaves to the operator
- no GPU
- Docker + Compose
- Needs SQLite, MySQL >= 5.7.8 or PostgreSQL >= 9.6, Redis (optional), ClickHouse (optional, log database)
- Models: OpenAI, Anthropic, Google Gemini, Azure OpenAI, AWS Bedrock
- port 3000
More in Gateways and observability
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.