AGIW MCP
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| TYPESAFE_API_KEY | No | TypeSafe API key for the optional Jev tools. If not set, agiw-mcp can read it from the macOS Keychain or Linux Secret Service. | |
| AGIW_MCP_AUTHOR_MODEL | No | Model ID of the author model. Can also be set as author_model in config.json. | |
| AGIW_MCP_MODEL_BASE_URL | No | Base URL of the local OpenAI-compatible model server, e.g. http://127.0.0.1:1234/v1. Can also be set as model_base_url in config.json. | |
| AGIW_MCP_REVIEWER_MODEL | No | Model ID of the reviewer model, which must differ from the author model. Can also be set as reviewer_model in config.json. |
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 |
|---|---|
| agiw_readinessA | The "initiate online code mode" check. Only checks, never starts work. Returns READY, or the FIRST missing prerequisite with an exact fix, in this order: Python, configuration, Node.js 22+, vendored Nisi hashes, local OpenAI-compatible model server listing two distinct configured models, consent recorded, TypeSafe API key present (macOS: presence checked without reading it; Linux: |
| agiw_statusB | Versions, configuration (home directory shown as ~) and which prerequisites are met. Contains no secrets and starts no work. |
| agiw_consentA | Show or revoke consent to send text to TypeSafe for the Jev checks. "show" returns exactly what is sent. Consent cannot be granted through this tool: only the user can grant it, by running |
| agiw_jev_preA | Send a task description (at most 12 KB) to Jev at TypeSafe with fixed questions about risk, blockers, missing detail, irreversible actions and reliance on old reports. Returns Jev's answers and which are flagged. Advisory, not a verdict. Refused unless consent is recorded and a key is available. Never retried automatically. |
| agiw_jev_postA | Send a result or report (truncated to 12 KB) to Jev at TypeSafe with fixed questions about self-contradiction, unbacked claims, risk of acting on it and misclassified findings. Advisory, not a verdict. Refused unless consent is recorded and a key is available. Never retried automatically. |
| agiw_nisi_runA | Run one Nisi v0.2.0 workflow (draft, syntax check, tests, independent review, bounded repair) on file CONTENTS you pass in, using the configured local author and reviewer models. Runs in a private temporary directory that is removed afterwards. Never reads or writes project files and never executes generated code; returns proposed contents and Nisi's report for you to review and apply yourself. Tests are the built-in checks (some file changed, no empty file) plus optional contains/not_contains assertions. |
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 6 tools
Each tool targets a distinct phase: readiness gate, informational status, consent management, isolated workflow execution, and pre/post Jev review. The main overlap is agiw_readiness vs agiw_status, both of which report on prerequisites, though descriptions differentiate them (ordered gate with a fix vs. a snapshot of versions/config).
All names share a consistent agiw_ prefix, snake_case, and a domain_action shape (agiw_jev_pre, agiw_jev_post, agiw_nisi_run). Minor deviation: agiw_readiness, agiw_status, and agiw_consent are noun-only with no action verb, but they remain readable and predictable.
Six tools is well-scoped for a gate-and-review server: two diagnostic tools, one execution tool, one consent tool, and a pre/post review pair. Every tool has a clear purpose with no filler.
The lifecycle (check readiness, run workflow, review before/after via Jev) is fully covered, and read-only-by-design behavior is explicit. The notable gap is that consent can only be granted out-of-band in a terminal, so an agent cannot complete that step itself; there is also no tool to list or inspect prior Jev/nisi results.