- Rust 98.9%
- Nix 1.1%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
~larandar
71d46eb7b2
`docs/tower-adapter-layer.md` was written before any of it was built. Four things it did not anticipate are now recorded there, three of which cost something: `provider-openai` had to exist before step 2 was possible at all; the pivot's ENDPOINT TYPES then had to follow, which is the step the plan was actually missing, since no adapter crate can build a `TypedEndpointConverter` it cannot name the source marker of; the orphan rule bit once on `Anthropic <-> OpenAI`, exactly where `adapter-extraction.md` said it would; and the hazard that bit was the opposite of the one named -- not two adapters claiming a request, but a terminal layer rejecting one its own stack had already translated. What it bought is measured rather than asserted: fourteen registrations to two adapters, any `Named` provider now translating without a Rust edit, and `ANTHROPIC_COMPATIBLE_PROVIDERS` left with no registry to drift from. `docs/adapter-extraction.md` gets a supersession note rather than an edit. Its registry section describes a file that no longer exists; the rest of it -- the orphan-rule analysis, the member-versus-excluded rule, the two classes of provider -- still holds, and a plan that was executed is worth keeping as what it was. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> |
||
| ai-gateway | ||
| crates | ||
| docs | ||
| nix | ||
| .env.template | ||
| .gitignore | ||
| Cargo.lock | ||
| Cargo.toml | ||
| flake.lock | ||
| flake.nix | ||
| LICENSE | ||
| NOTICE | ||
| README.md | ||
| rustfmt.toml | ||
modron-rs
The Ting estate's model router: one OpenAI-compatible and Anthropic-compatible door in front of several providers, with load balancing, rate limiting and OpenTelemetry built in.
modron is the estate's name for the capability — how a machine reaches a
model. This is the program that answers it.
What this is a fork of, and why
modron-rs is a fork of Helicone/ai-gateway,
taken at v0.2.0-beta.30 (081f310be779f27122e6038e81e2359c44db4ecb,
2025-07-20) — the last release published under the Apache License, Version 2.0.
Upstream relicensed to GPL-3.0 on 2025-11-21, in a commit that was also the last commit it ever received; the documentation section for the project was taken down, and of fifty-nine forks, none became a maintained continuation. Apache-2.0 §2 grants an irrevocable licence, so the tree at the fork point remains Apache-2.0 permanently. This repository diverges before the relicence and carries no code from it.
The fork point is a release, not merely the last Apache commit. The forty-four commits between them are almost entirely Helicone's own hosted buildout — ECR pushes, Terraform, Valkey, multi-region deployment, a Postgres listener feeding their control plane, prompt-input logging to their backend. That is coupling a self-hosted router wants less of, so the tag is both the cleaner provenance anchor and the smaller surface.
See NOTICE for attribution and the statement of modification required by Apache-2.0 §4(b).
What was removed
Everything that was Helicone's project furniture rather than the program:
their CI, container and deployment definitions (.github/, Dockerfile,
infrastructure/, render.yaml), release automation (cliff.toml,
dist-workspace.toml), editor and agent configuration, benchmarks, examples,
the test-script workspace members, and the project documents that the estate
keeps at the organisation level instead.
What remains is the gateway, its five crates, the lock file, and four files of paperwork.
Configuration
Two surfaces, and neither is a dashboard or a management API — which is most of why this fork exists.
Provider credentials come from the environment, one variable per provider, so a secret manager renders one file and nothing else needs to know:
OPENAI_API_KEY ANTHROPIC_API_KEY GEMINI_API_KEY
AWS_ACCESS_KEY AWS_SECRET_KEY
Routing is YAML, declared and diffable:
routers:
my-router:
load-balance:
chat:
strategy: model-latency
models:
- anthropic/claude-3-7-sonnet
- bedrock/anthropic.claude-3-7-sonnet-v1:0
Run it with -c <config.yaml>. The helicone: block in the shipped examples
configures an optional connection to the upstream vendor's hosted control
plane; features: none leaves it off, which is how the fork-point release
ships and the only mode the estate uses.
Building
A Cargo workspace on edition 2024, with Cargo.lock committed and no git
dependencies at all — every one of its 577 packages comes from crates.io.
That was not true at the fork point, and the difference is most of the work
this tree has had done to it. Upstream carried sixteen git-sourced packages
across two forked repositories, two of them on contributors' personal
accounts. async-openai turned out to be the published 0.29.0 plus four
lines and went back to the registry. stubr was replaced by wiremock-rs,
the crate it had vendored a copy of. aws-sdk-bedrockruntime — one line that
made cargo clone 2.3 GB, against 50 MB for the whole of crates.io — left with
the Bedrock adapter, which is excluded from the workspace rather than merely
feature-gated, because nix vendors whatever a lock names whether a target
uses it or not.
rustfmt.toml uses nightly-only options; cargo fmt on stable ignores them
and will reformat the tree.