Tuner MCP
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| HF_TOKEN | No | Token for accessing gated Hugging Face repositories. | |
| TINKER_API_KEY | Yes | API key for Tinker. Required for live sampling, evaluation, and training. | |
| TUNER_TASK_URL | No | URL for task queue. Defaults to 'memory://' for local MVP. Can point to a supported Redis/Valkey URL for remote deployment. | memory:// |
| TUNER_AUTH_TOKEN | No | Authentication token for HTTP transport. Required when using --transport http; HTTP mode fails closed if absent. | |
| TUNER_MAX_SAMPLES | No | Maximum number of samples. | |
| TUNER_ALLOWED_ROOTS | No | Semicolon-separated on Windows, colon-separated elsewhere. Paths are restricted to these roots. Defaults to the current directory. | |
| TUNER_MAX_BATCH_SIZE | No | Maximum batch size for training admission. | |
| TUNER_MAX_INPUT_TOKENS | No | Maximum input tokens. | |
| TUNER_MAX_PROMPT_BYTES | No | Bounds direct sampling/logprob request payloads. | |
| TUNER_MAX_DATASET_BYTES | No | Maximum dataset size in bytes. | |
| TUNER_MAX_ARTIFACT_BYTES | No | Caps each artifact read size. | |
| TUNER_MAX_TRAINING_STEPS | No | Maximum number of training steps. | |
| TUNER_MAX_CONCURRENT_RUNS | No | Maximum concurrent runs. | |
| TUNER_MAX_GENERATION_TOKENS | No | Maximum generation tokens. | |
| TUNER_MAX_TOTAL_GENERATION_TOKENS | No | Caps aggregate generated tokens per operation. |
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": true
} |
| logging | {} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| recipes_listC | Discover recipe requirements, availability and verification status. |
| recipe_getB | Inspect a recipe's typed configuration and execution requirements. |
| dataset_prepareB | Stage validated local or pinned Hugging Face data as a persistent dataset ID. |
| dataset_fetch_hfC | Fetch a pinned Hugging Face split, map rows to a Tuner schema, stage a dataset ID. |
| dataset_search_hfC | Search public Hugging Face datasets by popularity; returns repo IDs and SHAs. |
| dataset_probe_hfC | Probe configs and row mappings for a pinned Hugging Face dataset split. |
| experiment_autoplanC | Select live models, Cookbook recipes, and pinned HF candidates without training. |
| training_planA | Prepare an immutable training plan and report blockers; does not launch training. |
| recipe_planB | Validate and freeze an exact config for one allowlisted Cookbook recipe. |
| training_startA | Submit a validated plan to Docket and return its run ID promptly. Spends credits. |
| recipe_startA | Submit a reviewed official Cookbook recipe. This may spend Tinker credits. |
| training_stopB | Stop local orchestration. Already submitted remote work may continue. |
| capabilities_getB | Describe Tuner's tool surface and optionally fetch live Tinker server capabilities. |
| models_listA | List Cookbook-known models; live mode returns authoritative server-supported models. |
| dataset_validateB | Validate a local JSON/JSONL dataset, reporting exact malformed record indexes. |
| dataset_inspectC | Validate and return a small preview of a local dataset. |
| training_listB | List remote Tinker runs or persistent local Tuner workflow records. |
| training_getC | Get a Tuner workflow record or a Tinker training run by ID. |
| training_metricsA | Read latest metrics, or page from a byte cursor (start at zero) while training runs. |
| training_logsC | Read bounded recursive logs, or an artifact path with a byte cursor. |
| checkpoint_listB | List training and sampler checkpoints for a Tinker training run. |
| checkpoint_getA | Get checkpoint weight metadata using an exact tinker:// path. |
| sampleC | Sample a base model or checkpoint through the model-recommended Cookbook renderer. |
| compute_logprobsB | Compute prompt token log probabilities for a base model or checkpoint. |
| train_sftC | Run Cookbook supervised fine-tuning. This operation spends credits. |
| evaluateC | Run a Cookbook benchmark and persist evaluation artifacts. This spends credits. |
| compare_runsC | Compare compatible evaluations using an optional explicit metric policy. |
| training_resumeB | Resume SFT with total max_steps or additional_steps after checkpoint. Spends credits. |
| train_dpoC | Train on chosen/rejected pairs using Cookbook DPO. Spends credits. |
| train_rlC | Run an allowlisted arithmetic/math group-rollout RL recipe. Spends credits. |
| train_distillC | Run on-policy or off-policy teacher/student distillation. Spends credits. |
| objects_listC | Recover saved datasets and plans after reconnecting; returns bounded summaries. |
| object_getB | Retrieve a saved dataset or resolved plan by its persistent ID. |
| dataset_render_previewB | Preview Cookbook rendering and loss masks without submitting training. |
| experiment_artifactsB | List a run's files or read one artifact using a bounded byte cursor. |
| experiment_rolloutsC | Read a bounded page of persisted rollout or evaluation trajectories. |
| benchmarks_listA | List benchmark names discovered from installed Cookbook source. |
| evaluation_getB | Inspect an evaluation's persistent status, benchmark scores and artifacts. |
| evaluation_failuresB | Read evaluation trajectories with errors or non-positive rewards. |
| usage_getA | Get account usage for a half-open YYYY-MM-DD date range; preserve upstream units. |
| sessions_listC | List sessions owned by the server's Tinker account. |
| session_getA | Inspect one owned Tinker session and its remote run identifiers. |
| session_trace_exportB | Save an owned session trace as a local artifact for bounded inspection. |
| checkpoint_exportB | Queue a checkpoint export, or wait when background=false. |
| checkpoint_set_ttlB | Set checkpoint retention; null requests indefinite retention. |
| checkpoint_deleteB | Permanently delete the checkpoint at this exact Tinker path. |
| checkpoint_publishB | Make this exact checkpoint publicly accessible on Tinker. |
| checkpoint_unpublishB | Remove public access to this exact checkpoint on Tinker. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
| capabilities_resource | |
| recipes_resource | |
| models_resource |
TDQS
Scored across 48 tools
Most tools are grouped by resource and action, but the large set has several close planning/staging tools (experiment_autoplan/training_plan/recipe_plan, dataset_fetch_hf/dataset_probe_hf/dataset_prepare, dataset_validate/dataset_inspect) that could cause misselection. Descriptions help, but boundaries are not always obvious.
Names consistently use snake_case and resource prefixes, but action position varies: list/get tools are noun_verb (recipes_list, checkpoint_get) while training actions are verb_noun (train_sft, compute_logprobs) and some are bare verbs (evaluate, sample). The prefix grouping keeps it readable, but the mixed conventions are not fully predictable.
48 tools is far beyond the 25+ threshold and would overwhelm an agent even though the domain spans training, datasets, checkpoints, and evals. Several narrow operations (dataset_validate vs dataset_inspect, checkpoint_* variants) could be consolidated.
The surface covers the full training lifecycle: datasets, recipes, plan/start/resume/stop training, multiple RL/DPO/SFT/distill methods, checkpoint management, evaluation, sessions, and usage. Minor gaps exist (no evaluation_list, no training/cancel/delete for remote runs, no dataset deletion), but agents can complete core workflows.