open-agent-connector
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| OAC_API_KEY | No | Sent as Authorization: Bearer … Leave empty for backends without auth. | |
| OAC_BASE_URL | Yes | OpenAI-compatible base URL, including /v1 (for example http://localhost:4000/v1). Required. | |
| OAC_FETCH_PATH | No | Path of the native fetch endpoint. | /web/fetch |
| OAC_JUDGE_MODEL | No | Vision-capable chat model used by review: "judge". Defaults to the first text model. | first text model |
| OAC_SEARCH_PATH | No | Path of the native search endpoint. | /search |
| OAC_TEXT_MODELS | No | Comma-separated chat model ids. ask_agent auto and page digests try them in this order and fail over to the next. If empty, the first id from GET /models that doesn't look like an image, audio or embedding model is used. | auto-pick |
| OAC_IMAGE_MODELS | No | Comma-separated image model ids. All of them draw in parallel. generate_images stays disabled until this is set. | |
| OAC_SEARCH_MODEL | No | A model that supports the OpenAI Responses API web_search tool. Used after the native providers. | |
| OAC_IMAGE_OUT_DIR | No | Where candidate images are saved. init sets .generated-images inside the project. | $TMPDIR/open-agent-connector/images |
| OAC_FETCH_PROVIDERS | No | Provider ids for a native fetch endpoint (POST {base}/web/fetch with {model, url, format}). If empty, pages are fetched locally. | |
| OAC_TEXT_TIMEOUT_MS | No | HTTP timeout for text operations. | 120000 |
| OAC_IMAGE_TIMEOUT_MS | No | HTTP timeout for image operations. | 240000 |
| OAC_MAX_RESULT_CHARS | No | Maximum characters any tool returns to Claude. | 12000 |
| OAC_SEARCH_PROVIDERS | No | Provider ids for a native search endpoint (POST {base}/search with {model, query, max_results}), tried in order. Example: 9router's antigravity. | |
| OAC_MAX_INLINE_IMAGE_BYTES | No | Larger images are never inlined. | 1500000 |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| ask_agentA | Delegate a self-contained chore to a cheaper helper model to save Claude tokens: summarising docs/logs, drafting boilerplate, translating, brainstorming, explaining an API. Give full context in |
| web_searchA | Live web search done by helpers. Returns a short grounded answer + source URLs instead of raw pages, saving Claude tokens. Use for docs, error messages, versions, news. If the result starts with "NOT A LIVE SEARCH", treat it as possibly outdated. |
| web_fetchA | Read a web page. With |
| generate_imagesA | Generate images with every configured image model in parallel (OAC_IMAGE_MODELS), save candidates to disk, and return them for you (Claude) to choose the best. review: "paths" (default; open chosen files with Read), "judge" (a helper vision model pre-ranks — cheapest), "inline" (images embedded in this result). Copy the winner into the project yourself afterwards. |
| list_helper_modelsB | Show which helper model ids are configured for each role and which ids the backend advertises. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 5 tools
Each tool has a distinct primary role: ask_agent delegates text chores, web_search returns grounded search answers, web_fetch extracts page content, generate_images handles image generation, and list_helper_models introspects configuration. There is minor overlap between ask_agent and web_search for documentation/error queries, but the descriptions clarify when to use each.
All tool names use consistent snake_case, which makes them easy to scan. The pattern is mostly verb_noun (ask_agent, generate_images, list_helper_models) but web_search and web_fetch use noun_verb, a minor deviation from full consistency.
Five tools are well-scoped for a connector that delegates to helper models and performs web/image operations. Each tool covers a distinct capability without redundancy or bloat.
The surface covers delegation, live web search, page fetching, image generation, and helper model introspection. Minor gaps exist, such as no explicit local file-reading tool or helper configuration tool, but the core workflows are supported.