ANcpLua.Agents.Hosting.BitNet 5.0.0

ANcpLua.Agents.Hosting.BitNet

Consumer toolkit for Microsoft Agent Framework — local-LLM hosting via Microsoft's BitNet b1.58 over the OpenAI-compatible /v1 surface.

Alpha-channel package. Keep isolated from stable/preview consumers unless explicitly intended.

• Compatible with: Microsoft.Agents.AI 1.8.x
• Tested against: Microsoft.Agents.AI 1.8.0 + Microsoft.Extensions.AI 10.6.x
• Capability tested against: BitNet b1.58 2B-4T weights served by Microsoft's prebuilt bitnet.cpp Docker image

Naming: * = consumer-facing facade / entry-point, bare = primitive consumers may compose with. See the convention in ANcpLua.Agents.

Install

dotnet add package ANcpLua.Agents.Hosting.BitNet

Or add the PackageReference by hand:

<PackageReference Include="ANcpLua.Agents.Hosting.BitNet" />

Under Central Package Management (recommended; required by ANcpLua.NET.Sdk), pin the version in your Directory.Packages.props and leave the Version= off the PackageReference itself.

Standing up a BitNet server

The hosting package only speaks HTTP to an OpenAI-compatible endpoint — it never builds, downloads, or spawns the binary. You can satisfy that contract any way you like; Microsoft's prebuilt Docker image is the easiest path:

make bitnet-up                    # idempotent — stops any prior container first
export BITNET_URL=http://localhost:11434

make bitnet-up wraps scripts/bitnet-docker.sh start. The script pins the image by digest (sha256:9d5f7f4e...cd243a as of 2026-05-12), so byte-identical runs are guaranteed until you intentionally re-resolve.

What it bundles: bitnet.cpp, the b1.58-2B-4T GGUF weights, the patched llama-server, all under /v1/chat/completions on port 11434. No Python, cmake, LUT codegen, or git clone involved.

If your environment cannot pull this image (air-gapped hosts, vendor mirrors, custom builds), point BITNET_URL at any other OpenAI-compatible /v1/chat/completions endpoint — LM Studio, vLLM with a BitNet build, your own llama-server build, a private inference gateway. The hosting package does not care how the server got there. See Other OpenAI-compatible servers below.

Health check:

curl -fsS http://localhost:11434/health && echo ok

Stop:

make bitnet-down

Pinning by digest (production / supply-chain)

make bitnet-up already runs digest-pinned. If you want to re-resolve to a newer Microsoft build, the recipe is:

docker buildx imagetools inspect \
  mcr.microsoft.com/appsvc/docs/sidecars/sample-experiment:bitnet-b1.58-2b-4t-gguf \
  --format '{{json .Manifest.Digest}}'

Edit the IMAGE= constant at the top of scripts/bitnet-docker.sh with the new digest, commit, done.

Other OpenAI-compatible servers

LM Studio, vLLM with a BitNet build, a private inference gateway, your own fork — all work. Point BITNET_URL at them.

Note on Ollama: as of mid-2026 Ollama does not run BitNet b1.58 (upstream llama.cpp lacks the i2_s tensor type that BitNet requires; Ollama inherits that gap — see ollama/ollama#10334).

Why a dedicated hosting package

Stock Microsoft.Extensions.AI.OpenAI works against any OpenAI-compatible endpoint — but llama-server builds older than ggml-org/llama.cpp#19831 (merged 2026-02-23) silently ignore the SDK-emitted max_completion_tokens field and generate to the context limit. This package promotes the LegacyMaxTokensPolicy shim out of test-only territory and ships it as part of the runtime path. Once Microsoft's BitNet fork picks up the upstream merge, the policy becomes a no-op self-deleting decorator.

Usage — four modes

Mode 0 — zero config (defaults to BITNET_URL)

builder.AddBitNetChatClient();   // connection name defaults to "bitnet"

Resolves:

public sealed class MyAgent([FromKeyedServices("bitnet")] IChatClient chat) { ... }

Mode 1 — Aspire-style named connection

builder.AddBitNetChatClient("bitnet");
// reads ConnectionStrings:bitnet = "http://localhost:11434"

Mode 2 — programmatic configuration (keyed multi-endpoint)

builder.AddBitNetChatClient("local", o =>
{
    o.Endpoint = new Uri("http://localhost:11434");
    o.Model    = "bitnet-b1.58-2B-4T";
    o.ApiPath  = "/v1";
});

Mode 3 — assembly attribute + bundled source generator

[assembly: BitNetEndpoint("bitnet", "http://localhost:11434", Model = "bitnet-b1.58-2B-4T")]

// Program.cs — one call wires every declared endpoint
builder.AddDiscoveredBitNetClients();

What gets registered

• IChatClient keyed by the connection name (singleton)
• IOptions<BitNetClientOptions> bound to BitNet:<name>:* config section
• Health check named bitnet:<name> against <endpoint>/health
• OpenTelemetry decoration via Microsoft.Extensions.AI.UseOpenTelemetry() — emits standard gen_ai.* spans/meters

Environment overrides

For test scenarios and the BitNetFixture contract (ANcpLua.Agents.Testing.BitNet.BitNetFixture):

• BITNET_URL — overrides Endpoint. When unset, the fixture auto-starts the Docker image itself (unless BITNET_FIXTURE_NO_DOCKER is truthy).
• BITNET_API_PATH — overrides ApiPath (default /v1)
• BITNET_MODEL — overrides Model (default bitnet-b1.58-2B-4T)
• BITNET_FIXTURE_NO_DOCKER — set truthy to opt the fixture out of auto-Docker

Env vars win over config so existing fixture consumers keep working.

Smoke-test the auto-Docker path with make smoke (sets BITNET_SMOKE_TEST=1 and uses the Microsoft.Testing.Platform-correct --filter-method flag — global.json pins MTP as the runner, so legacy VSTest flags like --filter / --logger are rejected).

Troubleshooting