Nostraxis is a local dashboard to observe, launch and compare work done with general-purpose agents. It brings ChatGPT/Codex, Claude Code, GitHub Copilot CLI and observed Hermes Agent sessions into a single view, without sending your history to a Nostraxis service and without replacing each provider's own authentication.

It is designed for two ways of working:

• Reviewing, from a single place, the sessions you already opened with your usual tools.
• Using it as your local work surface: pick a repository, create a session with New session and run the chosen agent from that repository.

✨ What you get

📚 Contents

• Quick install
• Connecting ChatGPT/Codex, Claude, Copilot and Hermes
• Working with repositories and New session
• What Sessions means
• Comparing prompts, agents and sessions
• Usage, credits and costs
• Privacy and data boundaries
• Advanced configuration
• Troubleshooting
• Local API
• Contributing
• Security
• License

Quick install

Requirements

• Node.js 22.5 or later. Local storage uses node:sqlite.
• One or more agent CLIs installed and authenticated if you want to launch sessions or discover histories: Codex, Claude Code, GitHub Copilot CLI or Hermes Agent.
• Git is recommended so each repository's branch and commit can be recorded; a readable local folder can also be registered.

Starting the dashboard

From the repository root:

npm install
npm run dev

Open http://localhost:4173. On macOS you can also open start-dashboard.command; it starts in its own Terminal window and stays available as long as that window is open.

Useful commands:

npm test
npm run build
npm start
npm run check

The dashboard starts with no demo data. For isolated visual development you can use:

NOSTRAXIS_SEED=1 npm run dev

Connecting ChatGPT/Codex, Claude, Copilot and Hermes

There are no passwords or subscription keys to configure inside the interface. Install and sign in to each provider's CLI through its own official flow, and open the dashboard as the same macOS/Linux/Windows user. On start, Nostraxis detects the available executables and the local histories.

Open Settings to check the detected sources and the status of each adapter. In Sessions, use the sync button to force an immediate re-read of the histories. The watcher also refreshes local sources automatically every few seconds.

Copilot being connected only confirms that its executable or account is available; it does not mean every local conversation uses the same store. Nostraxis labels GitHub Copilot CLI / Agent and VS Code Copilot Chat separately. The VS Code adapter rebuilds the editor's local chatSessions journals, imports only visible user/assistant messages and reported tokens or AI credits, and associates them with the folder in the adjacent workspace.json. Histories from another computer are not downloaded from GitHub. Remote SSH, Dev Container, custom --user-data-dir, VSCodium or other installations can store data elsewhere; configure their local workspaceStorage roots with NOSTRAXIS_VSCODE_CHAT_ROOTS_JSON.

If the executable is not on PATH, point to it for the dashboard process only:

export NOSTRAXIS_CODEX_BIN='/absolute/path/to/codex'
export NOSTRAXIS_CLAUDE_BIN='/absolute/path/to/claude'
export NOSTRAXIS_COPILOT_BIN='/absolute/path/to/copilot'
export NOSTRAXIS_HERMES_BIN='/absolute/path/to/hermes'
npm run dev

Hermes is intentionally observation-only in the initial integration. Its native source is preserved (desktop, cli, cron or a messaging channel) and classified as Interactive, Automation or Messaging. This is a Hermes-specific distinction: its filter and grouping control appear only after selecting Hermes as the provider. By default every native source is imported; NOSTRAXIS_HERMES_SOURCES can restrict that set explicitly.

Copilot: account quota

The Copilot card can read the plan, monthly limit, usage, balance and reset date that GitHub exposes to the active GitHub CLI account. It does not reuse or store the gh credential. If that session does not exist on the machine, you can provide a short-lived token to the dashboard process only:

export NOSTRAXIS_COPILOT_TOKEN='github-token-with-copilot-access'
npm run dev

The token is not persisted. In Business or Enterprise organizations, the value shown is the personal budget when GitHub exposes it; organization-wide billing reports still require their own permissions.

Working with repositories and New session

You can use Nostraxis as your local work surface for any of the three agents. The dashboard runs the selected CLI inside the repository folder: it does not clone the code or move the project elsewhere.

1. Go to Repos.
2. Paste the absolute path of your local folder, for example /Users/ana/code/my-api, and press Register repository. If it is a Git repository, the current branch and HEAD are stored as well.
3. Go back to Sessions and press + New session.
4. Choose the registered repository, the provider and, where applicable, the model.
5. Give it a name, write the goal and decide whether to allow commands and file changes.
6. Press Create session. The session runs from that folder and is traced in the dashboard.

This lets you work with Codex, Claude or Copilot while keeping a single panel for context, output, commands, touched files and whatever metrics the provider reported. The write-permission option only affects sessions created from Nostraxis; review it before starting a task that will modify your checkout.

What Sessions means

Sessions is the dashboard's operational history. Each row represents a detected run or conversation, labelled by project, provider, model, status and origin.

Use the tabs to switch between active, recent and all sessions; filters narrow by project, provider, model, status, origin, update time, cost and cache. Grouping by project or status makes it easier to follow several open tasks at once. Selecting Hermes also enables its provider-specific type filter and grouping.

When you select a session, the central panel shows its timeline and, when the source exposes it, tokens, cost, credits, duration and events. Duration is the active time: the intervals where the agent was reporting work, measured from the real event timestamps. The time before each user action and any long silence are excluded, so a conversation resumed the next day is not read as a day of execution. The conversation span and the last turn appear beside it, so the latest interaction can still be measured on its own. The side inspector keeps the repository, branch, commit, prompt, tools and related files. A Not reported value means the provider did not deliver it: it never equals zero and is never silently estimated.

Comparing prompts, agents and sessions

The Compare view contrasts up to four real sessions. It is useful both for reviewing different prompts and for running the same prompt several times to evaluate different agents, models or permissions.

Recommended flow for a controlled test

1. Register the same repository and fix a stable branch or commit.
2. Create one session per variant in New session. To compare agents, use the same goal on Codex, Claude and/or Copilot. To compare prompts, change only the text you want to evaluate.
3. Avoid changes to the starting files between runs, or record the difference explicitly.
4. Go to Compare, search sessions by project, model or date, and select them.
5. Read the results together with the Context digest, the output and the files/tools used; a difference in context or task can invalidate a cost or speed comparison.

The matrix shows model, input/output tokens, estimated cost, provider credits, duration, cache, reasoning, evaluation, tools, files, context digest and final answer when present. From each column you can open the session detail. Fields the provider did not supply stay as Not reported.

For usage reports, Analytics aggregates sessions by repository, provider, model and date range. It includes a per-model breakdown, cost/token ratio, a time series and access to the detail of each run. This is the right view to answer, for example, which agent consumed most in a repository or how cost evolved over a week.

Reproducible experiments (optional)

The R&D Lab creates a variant matrix with a shared task and keeps the rendered prompt, the context, the repository HEAD and a SHA-256 digest to make repetition easier. Enable it when starting the dashboard:

NOSTRAXIS_EXPERIMENTS_ENABLED=1 npm run dev

The available strategies are raw-repo, knowledge-base and llm-wiki. Unset the variable or use a value other than 1 to hide this feature again.

Usage, credits and costs

At the top right there are cards for Codex / ChatGPT, Claude, GitHub Copilot and Hermes Agent. Click a card to open the source detail, the known model, the connection status and the last update. That way you can check from a single place what each provider allows you to observe.

The dashboard distinguishes three concepts that should not be confused:

• Subscription limit or quota: counter and reset date provided by the provider.
• Provider credits: provider-specific units, such as AI credits or Copilot premium requests. They are not dollars and are not comparable across providers.
• Estimated cost: a USD amount computed only when you configure per-model prices and there are enough tokens. It does not replace the provider's bill.

To enable estimated cost, define USD prices per million tokens before starting the server:

export NOSTRAXIS_PRICING_JSON='{"model-id":{"inputPerMillion":1.25,"cachedInputPerMillion":0.25,"outputPerMillion":10}}'
npm run dev

Privacy and data boundaries

Nostraxis is local-first. Its SQLite database lives by default at .nostraxis/dashboard.sqlite inside the dashboard project. You can change that location with NOSTRAXIS_DATA_DIR.

The watcher imports visible prompts and answers, tool metadata and usage the source reported. It does not import system prompts or hidden reasoning. External sessions are observation only; it does not take control of them.

For Hermes, Nostraxis opens state.db in read-only and query-only mode. It never mutates Hermes state, and it does not launch or manage Hermes sessions in this POC.

For VS Code Copilot Chat, Nostraxis reads the editor's local journal but excludes hidden transcript entries, agent instructions, tool payloads and reasoning blocks. The journal format is internal to VS Code, so unknown future records are ignored rather than inferred.

Copilot launched from Nostraxis enables the official OpenTelemetry exporter to an isolated JSONL at .nostraxis/copilot-otel, with message-content capture disabled. If you want to enrich external Copilot sessions with telemetry you already have, point to the file or directory:

export NOSTRAXIS_COPILOT_OTEL_PATH='/absolute/path/copilot-otel.jsonl'

For mounted or shared histories, replace the source list and adjust the discovery limits:

export NOSTRAXIS_SESSION_ROOTS_JSON='[{"provider":"codex","root":"/absolute/path/codex-sessions"}]'
export NOSTRAXIS_SESSION_MAX_FILES=200
export NOSTRAXIS_SESSION_MAX_AGE_DAYS=30

Advanced configuration

Troubleshooting

Local API

See architecture.md for the architecture, the adapters and the internal data flow.

Contributing

Contributions are welcome: bug reports, provider adapters, translations and documentation improvements are all useful.

Read CONTRIBUTING.md for the development workflow, architecture boundaries and pull-request checklist. Good first contributions include parser fixtures, provider adapters, accessibility improvements and translations.

The English README is canonical. When a user-facing behavior changes, update the relevant translated section when possible.

Security

Nostraxis reads local agent histories and can launch commands in registered repositories, so security reports should not be filed as public issues. See SECURITY.md for the private reporting process and the local trust model.

License

Released under the MIT License. See LICENSE for the full text.