AGY Observer MCP
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| AGY_OBSERVER_CLI | No | Absolute path to the official `agy` CLI executable to invoke (the desktop IDE cannot substitute for the CLI). | |
| AGY_OBSERVER_ENABLED | No | Execution gate for model runs. When set to "0" model execution is disabled; set to "1" (and restart the MCP server) to allow agy_run to invoke the Antigravity CLI after confirming costs, working directory and execution authorization. | 0 |
| AGY_OBSERVER_STATE_DIR | No | Absolute path to the private state directory where each run's status.json, events.jsonl and transcript.md are stored in plaintext. | |
| AGY_OBSERVER_WORKSPACE | No | Absolute path to the (recommended empty) workspace directory used when running the CLI. |
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
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
| experimental | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| agy_runB | Send one authorized prompt to the configured CLI model; may incur cost and allow CLI tools. No retries. |
| agy_historyA | List local history summaries and task previews. Use agy_status for full prompts and replies. |
| agy_statusB | Read one run's current or terminal status from the configured artifact directory. |
| agy_cancelA | Request cancellation of an active run owned by this server instance. |
| agy_discussion_createA | Create a local peer collaboration task with complementary duties and 1..8 attempt budget. No model call. |
| agy_discussion_readA | List task summaries when discussion_id is omitted; otherwise read roles, attributed messages, decisions, context manifests and linked runs. No model call. |
| agy_discussion_noteA | Submit a client proposal/question/answer/challenge/revision/feedback/decision/summary. An answer must reference the peer question. Decisions require reply_to and accepted/rejected/deferred; summary closes the task. Client authorship is not authenticated. No model call. |
| agy_discussion_turnB | Yield ONE authorized peer turn using bounded shared context. Instruction is optional; the peer may initiate questions or proposals, including the first turn. Failure/cancellation consumes an attempt; no automatic turns/retries. May incur cost and CLI side effects. |
| agy_discussion_recoverB | Reconcile an interrupted review with its terminal run; refuses active/nonterminal runs. Never retries or refunds attempts. |
| agy_council_createC | Create a GPT/Gemini/Claude task with sealed initial judgments and separate 1..8 AGY attempt budgets. No model call. |
| agy_council_noteC | Submit GPT/client public contribution. During collect use initial with honest context declaration (not authenticated). Later revision needs reason+references; support needs grounds+references; plan needs sources+expected version; summary archives without consensus. No model call. |
| agy_council_roundC | Explicitly call Gemini and/or Claude once each via AGY (default both concurrently). Separate attempt budgets; no retries/loop. Collect uses only original material, disallows leading focus, and keeps replies sealed. Later batch shares a pre-call snapshot. May incur cost; CLI tools are not forcibly disabled. |
| agy_council_revealB | Publish all three successful initial judgments together and enable shared discussion. No model call. Sealing is workflow isolation, not a local access-control boundary. |
| agy_infoA | Report non-secret configuration, version, and execution gate; does not inspect CLI authentication. |
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 14 tools
Most tools are clearly separated by workflow prefix (discussion vs council) and action, but agy_run, agy_discussion_turn, and agy_council_round all invoke models, and the two *_note tools could be confused at a glance. Descriptions clarify context, so misselection risk is moderate rather than high.
All names use a consistent agy_ prefix with snake_case and a domain_action structure (discussion_read, council_create, run, status). There are no mixed conventions or vague standalone verbs.
14 tools fits the well-scoped 3-15 range and supports two distinct workflows plus generic run management. Each tool maps to a concrete operation rather than padding the surface.
The discussion workflow has create/read/note/turn/recover, but the council workflow lacks an equivalent read/inspect tool and there is no explicit list/delete/archive operation for council tasks. Generic run status/history/cancel help, but agents may hit dead ends inspecting council state before reveal.