Puppet Master
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| PUPPET_MODEL | No | Model id sent with each request to the worker server, e.g. qwen2.5-coder:14b for Ollama or a catalog model id for OpenRouter. | flint |
| PUPPET_API_KEY | No | API key for hosted OpenAI-compatible APIs. Read only from the environment and sent as 'Authorization: Bearer <key>'. Not supported in puppet-master.toml by design. | |
| PUPPET_BASE_URL | No | Server URL of the OpenAI-compatible worker model endpoint, without /v1. Set this to point Puppet Master at a different server (Ollama, LM Studio, vLLM, mlx_lm, OpenRouter, or other hosted APIs). Environment variables win over puppet-master.toml. | http://127.0.0.1:8000 |
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 |
|---|---|
| puppet_statusA | Health and budgets of the local Flint model. Call once per session before delegating. |
| puppet_serverA | Start/stop/restart the local Flint llama-server using the fixed settings in ~/.puppet-master/server.toml. Never call start first: every Flint tool autostarts it, and it stops itself after idle_stop_min. Use stop when you have finished the delegated work for the user's request (it reports the tokens Flint saved this session), when the user needs RAM back, and always before compiling or other memory-heavy commands. |
| puppet_readA | USE THIS FIRST to read, search or summarize files/logs instead of opening them yourself. Returns a short answer whose claims are mechanically verified against the source (unverified claims are removed). Before acting on a claim, read its cited lines yourself. Pass files, path:start-end ranges or globs (e.g. /repo/src/**/*.py), never directories. Anything over one chunk is searched: Flint sees only the blocks that best match the question, and |
| puppet_writeA | USE THIS for boilerplate, scaffolding, tests from a spec, docstrings, mechanical refactors. Flint writes a STAGED patch (your working tree is untouched) and tests run on it automatically. You MUST read the patch at patch_path before calling puppet_apply. To retry, call again with the same task and your feedback in constraints (max 2 attempts). A file over ~2k tokens is edited by region: Flint sees an outline and rewrites only the blocks that match the task, so name the function or class to change. Limits: ≤3 files, ≤6k tokens of editable text total. |
| puppet_applyA | Apply a staged Flint patch to the working tree. Only call after you have read the patch file yourself and agree with it. Refuses if checks failed or the patch was modified. |
| puppet_rejectB | Reject a staged patch after review; counts as a failed attempt. |
| puppet_checkA | Run an allow-listed check (names from puppet-master.toml, e.g. 'test', 'lint') and get only the failures. With patch_id, runs against that staged patch in an isolated worktree. |
| puppet_askA | Drafts and prose only (docs, emails, commit messages, rewrites, explanations of text you provide). Output is UNVERIFIED. For any question about code or files, use puppet_read instead. |
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 8 tools
The patch lifecycle tools (write/apply/reject/check) have clearly separated roles, and the read-vs-ask split is explicitly and repeatedly contrasted in the descriptions. The only mild overlap is between puppet_status (health/budgets) and puppet_server (start/stop/restart), since both concern the same Flint server process.
Every tool uses the same snake_case, puppet_-prefixed verb-noun convention (puppet_status, puppet_read, puppet_write, puppet_apply, puppet_reject, puppet_check, puppet_ask). Predictable and uniform throughout.
Eight tools is well-scoped for a delegation/patch-review workflow; each tool maps to a distinct step (status, server lifecycle, read, write, apply, reject, check, ask) with no filler.
The surface covers the full delegate-read-write-review-apply lifecycle plus server management. Minor gaps: no way to list staged patches, revert an applied patch, or roll back a bad apply, which an agent may occasionally need.