Adaption Labs Brings Model Training and Fine-Tuning Into Hermes Agent
Adaption's new plugin lets Nous Research's Hermes Agent generate datasets and fine-tune custom models directly from its own command loop using AutoScientist.
Takeaways
- Adaption shipped a Hermes Agent plugin for in-agent dataset generation and fine-tuning.
- Three skills:
adaption-dataset,adaption-invent, andadaption-trainingfor AutoScientist. - Install via hosted MCP server with OAuth, plus plugin repo
adaptionlabs/adaption-hermes-plugin. - Headless API-key auth supported for CI and gateways with no browser login.
- AutoScientist reportedly lifted win rates from 48% to 64% versus Adaption's own researchers.
- Training and data operations consume Adaption credits; cost estimates are advisory, not enforced.
Adaption Labs has released an Adaption plugin for Nous Research’s Hermes Agent. The integration lets developers generate synthetic data, prepare datasets, launch fine-tuning jobs, and monitor training through the same agent they use for coding and research.
The plugin connects Hermes to Adaption’s hosted services through the Model Context Protocol, or MCP. MCP gives an agent a standard way to call external tools, while Hermes skills tell the agent when and how to use them. Developers can therefore describe a target capability in natural language and have Hermes coordinate the data and training workflow.
Installation has two moving parts
Hermes plugins cannot initiate OAuth for an MCP server, so authentication and skill installation happen separately. The MCP connection provides access to Adaption’s hosted tools, while the plugin repository supplies the instructions Hermes uses to operate them.
- Update Hermes to a release that supports Agent Plugins with
hermes update. - Add the hosted Adaption MCP server and complete OAuth in a browser.
- Install the skills bundle from the
adaptionlabs/adaption-hermes-pluginrepository. - Confirm the connection by asking Hermes to list the datasets associated with the account.
List my Adaption datasets.
Headless environments and CI gateways can authenticate with an API key using --auth header. Hermes stores that credential in the active profile’s .env file as MCP_ADAPTION_API_KEY, so teams should apply the same access controls and secret-rotation policies they use for other deployment credentials.
Three skills cover the workflow
The repository contains three skill definitions, with execution handled by Adaption’s hosted MCP tools.
| Skill | What it does |
|---|---|
adaption-dataset |
Imports, adapts, augments, translates, localizes, combines, and exports datasets. |
adaption-invent |
Generates synthetic datasets from a natural-language specification. |
adaption-training |
Selects models, recommends hyperparameters, launches AutoScientist jobs, and monitors training. |
The skills direct Hermes to request a cost estimate before starting an operation that consumes credits. Hermes does not enforce that instruction at the platform level, so unattended workflows need external spending limits, approval gates, or account monitoring.
AutoScientist handles the training decisions
Adaption’s AutoScientist docs describe a system that automates base-model selection, hyperparameter recommendations, training, and monitoring. Adaption Labs, co-founded by former Cohere research head Sara Hooker, positions the service as a way to reduce the specialist work required to produce a useful fine-tuned model.
According to Adaption’s internal evaluations, AutoScientist increased win rates from 48% to 64% against configurations prepared by the company’s AI researchers. That result comes from an in-house comparison, so teams still need task-specific evaluations before adopting a trained checkpoint.
AutoScientist builds on Adaptive Data, Adaption’s dataset preparation service. Together, the tools cover a loop that begins with dataset creation or import, continues through curation and training, and ends with a downloadable checkpoint that can be tested against application-specific criteria.
A shorter route from prompt to checkpoint
A developer could ask Hermes to build a dataset for extracting invoice fields and fine-tune a 7-billion-parameter model on it. Hermes can translate that request into dataset operations, retrieve a cost estimate, launch the training job, monitor its status, and report when the checkpoint is available.
This workflow reduces handoffs among notebooks, dataset tools, shell commands, and training dashboards. It also preserves the agent’s session context, including the target task and prior implementation work, while the training run proceeds.
Budget and evaluation remain external controls
- Usage costs: Dataset generation, augmentation, translation, localization, and AutoScientist training consume Adaption credits.
- Evaluation: Adaption says AutoScientist has improved win rates across several models, but benchmarks such as SWE-bench and ARC-AGI may not represent a team’s production workload. Application-specific test sets remain necessary.
- Compatibility: The integration requires a recent Hermes release with Agent Plugins support.
- Authentication: Interactive use requires browser-based OAuth, while automated environments need an API key stored in the Hermes profile.
- Data handling: Because execution uses a hosted MCP service, teams working with sensitive datasets should review Adaption’s retention, access, and compliance terms before uploading data.
Who benefits from the integration
Teams already using Hermes gain the clearest advantage because the plugin adds dataset and training operations to an established agent workflow. It is also suited to developers who need specialized smaller models but lack dedicated infrastructure for model selection, hyperparameter search, and training orchestration.
The release reflects Adaption’s broader approach: model quality can improve through tighter data and training loops, even when the underlying base model stays the same. By exposing that loop through Hermes, the company gives open-source agent users a direct path from a task description to an evaluated fine-tuned checkpoint.