microsoft · Salil Das (sadlilas) · Aug 18, 2026 · Aug 18, 2026 · Aug 18, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -18,6 +18,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   `~/.amplifier/openai-chatgpt-oauth.json`, refreshing them itself. Requires "Sign in with
   device code" enabled in the account's ChatGPT Security settings. Like `github-copilot`,
   `auth set openai-chatgpt` is refused — there is no static key to store.
+- **Chat Completions provider.** `provider.module: "chat-completions"` is now a valid
+  host-config value, backed by `amplifier-module-provider-chat-completions`. It integrates
+  any server speaking the OpenAI Chat Completions wire format
+  (`/v1/chat/completions`) — llama.cpp, vLLM, LM Studio, LocalAI, SGLang, TGI, and other
+  OpenAI-compatible endpoints — distinct from `openai`, which uses the OpenAI Responses
+  API. Its credential is an endpoint, not a key: `CHAT_COMPLETIONS_BASE_URL` (required)
+  selects the server, and `CHAT_COMPLETIONS_API_KEY` (optional) is sent only when set,
+  since local servers commonly need none. Both are environment-only; the persisted
+  credentials file is not consulted for this provider. Default model is `default`.
 
 ## [0.12.0] — 2026-07-29
 

diff --git a/README.md b/README.md
@@ -43,7 +43,7 @@ Public integrations run opencode, paperclip, and NanoClaw on it: see [who has in
 
 `amplifier-agent` ships with:
 
-- Six providers behind one interface: Anthropic, OpenAI, Azure OpenAI, Ollama, GitHub Copilot, and ChatGPT (a Plus/Pro/Team subscription via OAuth device-code, no API key), with credentials read from the environment or a cached OAuth session
+- Seven providers behind one interface: Anthropic, OpenAI, Azure OpenAI, Ollama, GitHub Copilot, ChatGPT (a Plus/Pro/Team subscription via OAuth device-code, no API key), and Chat Completions (any OpenAI Chat Completions-compatible endpoint, e.g. llama.cpp, vLLM, LM Studio), with credentials read from the environment or a cached OAuth session
 - Role-based model routing, so a sub-agent gets a model matched to its job rather than the frontier model for everything, re-matched when you switch providers
 - Context management that keeps long sessions running, compacting history before it overruns the window
 - Tools for filesystem, bash, web, search, todo, and MCP
@@ -83,7 +83,7 @@ amplifier-agent auth status              # diagnose env-vs-file precedence per p
 amplifier-agent models list              # enumerate available models from providers
 ```
 
-Full precedence rules, GitHub Copilot's environment-only caveat, and the host config file schema are in [`docs/CONFIGURATION.md`](docs/CONFIGURATION.md).
+Full precedence rules, GitHub Copilot's environment-only caveat, the `chat-completions` provider's required `CHAT_COMPLETIONS_BASE_URL`, and the host config file schema are in [`docs/CONFIGURATION.md`](docs/CONFIGURATION.md).
 
 ## Use it from your code
 

diff --git a/docs/CONFIGURATION.md b/docs/CONFIGURATION.md
@@ -11,11 +11,14 @@ Provider is auto-detected from environment variables in this precedence:
 3. `AZURE_OPENAI_API_KEY` + `AZURE_OPENAI_ENDPOINT`
 4. `OLLAMA_HOST` (defaults to `http://localhost:11434`)
 
-`github-copilot` and `openai-chatgpt` are excluded from this auto-detect chain -- neither resolves
-from a single API-key environment variable. `github-copilot` reads its own token chain from the
-environment (see below). `openai-chatgpt` has no credential env var at all: it authenticates via
-OAuth device-code, caching tokens to `~/.amplifier/openai-chatgpt-oauth.json`. Both must be
-selected explicitly with `provider.module` in a host config file.
+`github-copilot`, `openai-chatgpt`, and `chat-completions` are excluded from this auto-detect chain
+-- none of them resolves from a single API-key environment variable. `github-copilot` reads its own
+token chain from the environment (see below). `openai-chatgpt` has no credential env var at all: it
+authenticates via OAuth device-code, caching tokens to `~/.amplifier/openai-chatgpt-oauth.json`.
+`chat-completions` needs an endpoint rather than a key (`CHAT_COMPLETIONS_BASE_URL`, plus optional
+`CHAT_COMPLETIONS_API_KEY`) and has no implicit default endpoint to fall back to. All three must be
+selected explicitly with `provider.module` in a host config file, rather than silently winning a
+"first match" race.
 
 Override by passing `--config <path>` at a host config file that names a provider explicitly.
 
@@ -73,6 +76,13 @@ This matters for hosts that spawn `amplifier-agent` as a subprocess: once you ha
 >
 > An existing `gh` or VS Code login may already authenticate it through the SDK's cached OAuth, so try it before exporting anything. This is a temporary limitation: the real fix is in the provider module, whose token resolver needs to read the agent-delivered credential from its config before falling back to the environment. `auth set` support returns once that lands.
 
+> **`chat-completions` is environment-only too, but for a different reason.** Its "credential" is the target `base_url`, not an API key, so it is read purely from the environment: `CHAT_COMPLETIONS_BASE_URL` (required — the server to talk to) and `CHAT_COMPLETIONS_API_KEY` (optional; local servers like llama.cpp, vLLM, LM Studio, and LocalAI commonly need none). The persisted credentials file is not consulted for this provider.
+>
+> ```bash
+> export CHAT_COMPLETIONS_BASE_URL=http://localhost:8000/v1
+> # export CHAT_COMPLETIONS_API_KEY=...   # only if your server requires one
+> ```
+
 The file format is a versioned JSON envelope:
 
 ```jsonc

diff --git a/docs/INTEGRATION.md b/docs/INTEGRATION.md
@@ -8,7 +8,7 @@ The engine runs **one turn per invocation** and exits. Continuity across turns c
 
 `amplifier-agent` is a standalone binary. You do not need the Amplifier CLI, bundles, or any other repository in the `microsoft/amplifier*` family, and none of them is a substitute for it here.
 
-Use it when your software needs to run an agent: a loop with tools, file access, sub-agents, and/or multi-turn state. It also works for plain LLM calls, where you get routing across six providers behind one interface.
+Use it when your software needs to run an agent: a loop with tools, file access, sub-agents, and/or multi-turn state. It also works for plain LLM calls, where you get routing across seven providers behind one interface.
 
 Then pick a surface below, install the engine ([INSTALL.md](INSTALL.md)), and finish with the [checklist](#checklist-for-a-new-integration).
 

diff --git a/docs/LAYERS_AND_RELEASES.md b/docs/LAYERS_AND_RELEASES.md
@@ -115,7 +115,7 @@ The engine ships with `bundle.md` baked into the wheel. It declares which module
 
 **Pre-wired modules:**
 
-- **Providers:** `provider-anthropic`, `provider-openai`, `provider-azure-openai`, `provider-ollama`, `provider-github-copilot`, `provider-openai-chatgpt`
+- **Providers:** `provider-anthropic`, `provider-openai`, `provider-azure-openai`, `provider-ollama`, `provider-github-copilot`, `provider-openai-chatgpt`, `provider-chat-completions`
 - **Orchestrator:** `loop-streaming` (with `extended_thinking: true`)
 - **Context:** `context-simple` (300K tokens, auto-compact at 80%)
 - **Tools:** `tool-filesystem`, `tool-bash`, `tool-web`, `tool-search`, `tool-todo`, `tool-apply-patch`, `tool-delegate`, `tool-mcp`, `tool-skills`, `tool-mode`, `tool-recipes`

diff --git a/docs/architecture/architecture.dot b/docs/architecture/architecture.dot
@@ -84,7 +84,7 @@ digraph amplifier_agent {
 
     node [fillcolor="#d9d9d9"];
     kernel    [label="amplifier-foundation\nbundle / session kernel"];
-    providers [label="LLM providers\nanthropic · openai · azure\nollama · copilot · chatgpt"];
+    providers [label="LLM providers\nanthropic · openai · azure\nollama · copilot · chatgpt\nchat-completions"];
     mcp       [label="MCP servers"];
   }
 

diff --git a/docs/spec/bundle-and-cache.md b/docs/spec/bundle-and-cache.md
@@ -30,7 +30,8 @@ default_provider: anthropic                REQUIRED, engine-level, top-level key
 
 providers:                                 install-only stubs, no config and no credentials
   provider-anthropic, provider-openai, provider-azure-openai,
-  provider-ollama, provider-github-copilot, provider-openai-chatgpt
+  provider-ollama, provider-github-copilot, provider-openai-chatgpt,
+  provider-chat-completions
 
 session.orchestrator: loop-streaming       extended_thinking: true
 session.context:      context-simple       max_tokens 300000, auto_compact

diff --git a/docs/spec/host-config.md b/docs/spec/host-config.md
@@ -81,7 +81,8 @@ approval.patterns  must be a list of strings
 `provider` selects the provider module and carries its config.
 
 ```
-provider.module   one of: anthropic, openai, azure-openai, ollama, github-copilot, openai-chatgpt
+provider.module   one of: anthropic, openai, azure-openai, ollama, github-copilot,
+                  openai-chatgpt, chat-completions
 provider.config   free-form; belongs to the provider module
 ```
 

diff --git a/docs/spec/providers-and-models.md b/docs/spec/providers-and-models.md
@@ -9,21 +9,22 @@ routes a wire `model` field to a provider (see `http-face.md`).
 
 ## Supported providers
 
-Six providers are supported, and only six. The provider name is the value used in configuration,
+Seven providers are supported, and only seven. The provider name is the value used in configuration,
 in `auth` subcommands, and in `models list --provider`.
 
 ```
-anthropic       provider-anthropic
-openai          provider-openai
-azure-openai    provider-azure-openai
-ollama          provider-ollama
-github-copilot  provider-github-copilot
-openai-chatgpt  provider-openai-chatgpt
+anthropic          provider-anthropic
+openai             provider-openai
+azure-openai       provider-azure-openai
+ollama             provider-ollama
+github-copilot     provider-github-copilot
+openai-chatgpt     provider-openai-chatgpt
+chat-completions   provider-chat-completions
 ```
 
 Each module is installed from `git+https://github.com/microsoft/amplifier-module-<module>@main`.
-All six are declared by the shipped bundle as install-only, so preparing the bundle makes every
-provider importable before any session exists.
+All seven are declared by the shipped bundle (`bundle.md`'s top-level `providers:` stub list) as
+install-only, so preparing the bundle makes every provider importable before any session exists.
 
 The agent holds no static table of default models, credential field shapes, or display names. Those
 come from the provider module at runtime, so they cannot drift from provider truth.
@@ -44,12 +45,13 @@ configuration.
 Primary variables, in the order consulted:
 
 ```
-anthropic       ANTHROPIC_API_KEY
-openai          OPENAI_API_KEY
-azure-openai    AZURE_OPENAI_API_KEY, then AZURE_OPENAI_KEY
-ollama          OLLAMA_HOST, then OLLAMA_BASE_URL
-github-copilot  GITHUB_TOKEN
-openai-chatgpt  (none -- OAuth device-code)
+anthropic          ANTHROPIC_API_KEY
+openai             OPENAI_API_KEY
+azure-openai       AZURE_OPENAI_API_KEY, then AZURE_OPENAI_KEY
+ollama             OLLAMA_HOST, then OLLAMA_BASE_URL
+github-copilot     GITHUB_TOKEN
+openai-chatgpt     (none -- OAuth device-code)
+chat-completions   CHAT_COMPLETIONS_BASE_URL, plus optional CHAT_COMPLETIONS_API_KEY
 ```
 
 `AZURE_OPENAI_KEY` is the only deprecated alias. Consulting it emits a one-time warning on stderr.
@@ -68,13 +70,25 @@ openai-chatgpt has no environment variable at all. It resolves from a cached OAu
 flow (`login_on_mount`) and refreshed automatically thereafter. Its resolution reports source
 `"file"` when a token is cached and `"none"` otherwise -- never `"env"`.
 
+chat-completions has its own dedicated resolution branch, distinct from the generic
+env-then-file chain above: its primary variable's value lands in `fields["base_url"]`, not
+`fields["api_key"]`, because the thing it needs to know is *which server to talk to*, not a
+secret. `CHAT_COMPLETIONS_API_KEY` is consulted only when set (local servers such as llama.cpp,
+vLLM, LM Studio, and LocalAI commonly need none) and lands in `fields["api_key"]` alongside it.
+Unlike every other provider, the persisted credentials file is never consulted for
+chat-completions -- with no `CHAT_COMPLETIONS_BASE_URL` in the environment it resolves
+unconditionally to `source == "none"`, with no file fallback and no usable default to fall back
+to (unlike ollama's built-in localhost).
+
 A resolution reports the provider, whether it resolved, the source (`env`, `file`, `default`, or
 `none`), the variable consulted, and the resolved fields. Ollama backed only by the built-in default
 host reports unresolved on purpose, so auto-enrollment does not enlist a local daemon that may not
 be running.
 
-Requesting credentials for a key-based provider that resolves to `none` is an error. Ollama and
-unrecognized provider names never raise.
+Requesting credentials for a key-based provider that resolves to `none` is an error. Ollama,
+chat-completions, and unrecognized provider names never raise -- chat-completions is not one of
+the key-based providers this rule applies to, so a missing `CHAT_COMPLETIONS_BASE_URL` is left for
+the provider module itself to reject at call time, not raised here.
 
 ## The credentials file
 
@@ -115,6 +129,11 @@ to `~/.amplifier/openai-chatgpt-oauth.json`, refreshed by the provider module it
 are enumerated in the same `_CONFIG_CREDENTIAL_UNSUPPORTED` gate; this is temporary and specific to
 these two providers.
 
+`auth set chat-completions` is accepted (unlike github-copilot, it is not refused), but as noted
+above the chat-completions resolution branch never reads the credentials file -- only
+`CHAT_COMPLETIONS_BASE_URL` / `CHAT_COMPLETIONS_API_KEY` in the environment are consulted. Set
+those instead of relying on `auth set` for this provider.
+
 `auth clear` without `--force` exits 2.
 
 ## Provider selection at boot
@@ -125,7 +144,7 @@ these two providers.
 3. no further fallback: a bundle declaring neither is a hard error at boot
 ```
 
-`provider.module` is closed to the six supported names. Any other value fails validation with
+`provider.module` is closed to the seven supported names. Any other value fails validation with
 error code `config_invalid_provider_module`. `"auto"` is not a valid value.
 
 There is no `--provider` flag and no environment-based provider auto-detection. See Non-goals.

diff --git a/skills/amplifier-agent/SKILL.md b/skills/amplifier-agent/SKILL.md
@@ -22,7 +22,7 @@ metadata:
 
 `amplifier-agent` is an agent engine that other software runs on. Give it a prompt and it runs the full loop, with tools, sub-agents, skills, and MCP, then returns a result. Anything that can spawn a subprocess can use it; Python hosts can embed the engine library in-process instead.
 
-Reach for it when the project needs an *agent* (a tool loop, file access, sub-agents, multi-turn state) rather than a single completion. You can also use it for plain LLM calls, with routing across six providers behind one interface.
+Reach for it when the project needs an *agent* (a tool loop, file access, sub-agents, multi-turn state) rather than a single completion. You can also use it for plain LLM calls, with routing across seven providers behind one interface.
 
 **The engine runs one turn per invocation and exits.** Continuity across turns comes from a session id, not from a long-lived process. Every surface below is a different way of delivering a prompt to that same engine.
 
@@ -54,7 +54,7 @@ The installer needs `uv` and `curl` and will not bootstrap them silently; it tel
 
 Install as **the same user that runs the host process**; a host spawning a subprocess inherits that user's `PATH`. `amplifier-agent doctor` is the check that the install actually works, so run it before writing any integration code.
 
-Credentials are read from the environment, first match wins: `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `AZURE_OPENAI_API_KEY` plus `AZURE_OPENAI_ENDPOINT`, `OLLAMA_HOST`. GitHub Copilot is environment-only (`COPILOT_AGENT_TOKEN`, `COPILOT_GITHUB_TOKEN`, `GH_TOKEN`, `GITHUB_TOKEN`). ChatGPT (`openai-chatgpt`) has no credential env var at all: it authenticates via OAuth device-code, caching tokens to `~/.amplifier/openai-chatgpt-oauth.json`. Or store a static key with `amplifier-agent auth set anthropic sk-ant-...` (not supported for github-copilot or openai-chatgpt).
+Credentials are read from the environment, first match wins: `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `AZURE_OPENAI_API_KEY` plus `AZURE_OPENAI_ENDPOINT`, `OLLAMA_HOST`. GitHub Copilot is environment-only (`COPILOT_AGENT_TOKEN`, `COPILOT_GITHUB_TOKEN`, `GH_TOKEN`, `GITHUB_TOKEN`). ChatGPT (`openai-chatgpt`) has no credential env var at all: it authenticates via OAuth device-code, caching tokens to `~/.amplifier/openai-chatgpt-oauth.json`. The `chat-completions` provider is environment-only too, for any OpenAI Chat Completions-compatible endpoint (llama.cpp, vLLM, LM Studio, LocalAI, and similar): `CHAT_COMPLETIONS_BASE_URL` (required) plus optional `CHAT_COMPLETIONS_API_KEY`. Or store a static key with `amplifier-agent auth set anthropic sk-ant-...` (not supported for github-copilot or openai-chatgpt).
 
 ## Pick a surface
 
@@ -191,7 +191,7 @@ A per-instance config file looks like this:
 | `config_unreadable`, `config_malformed_json` | The `--config` file could not be opened, or is not a JSON object | Check the path the host wrote, and that it serialized an object |
 | `config_unknown_key` | Unrecognized **top-level** config key | The top level is closed: `approval`, `provider`, `providers`, `mcp`, `skills`, `debug`, `allowProtocolSkew` |
 | `config_invalid_type` | A known key has the wrong type, or an unknown sub-key in a closed inner shape | `skills.*` and `debug.*` are closed and raise this rather than `config_unknown_key`, which is reserved for the top level and `providers.<id>` entries |
-| `config_invalid_provider_module` | `provider.module` is not a known provider | One of `anthropic`, `openai`, `azure-openai`, `ollama`, `github-copilot`, `openai-chatgpt`. `"auto"` is not valid |
+| `config_invalid_provider_module` | `provider.module` is not a known provider | One of `anthropic`, `openai`, `azure-openai`, `ollama`, `github-copilot`, `openai-chatgpt`, `chat-completions`. `"auto"` is not valid |
 | `protocol_version_mismatch` | Wrapper and engine protocol versions differ | Update the lagging side. `allowProtocolSkew` is an unblock, not a fix |
 | `lifecycle_unsupported` | `submit()` called twice on one handle | New handle per turn, same `sessionId` with `resume` |
 | `env_injection_rejected` | The wrapper refused the environment you asked it to inject | Check the key against the wrapper's allowlist and blocked-key list |