Starfetch

Starfetch

[Starfetch]

[CI] [Coverage] [MCP Registry] [Starfetch on Glama] [skills.sh] [npm version]

Starfetch gives AI agents safe, reproducible access to public astronomy catalogs through the Model Context Protocol (MCP).

Ask an astronomy question in natural language. A Starfetch-enabled agent can select an appropriate service, inspect its live tables and columns, construct a bounded ADQL query, execute it, and return the result with the exact service, table, query, limit, units, and assumptions it used.

[A Starfetch-enabled agent inspecting Gaia metadata, ranking high-proper-motion sources near the Pleiades, and summarizing the results]

You: Find the 10 Gaia DR3 sources with the highest proper motion within 0.5     degrees of the Pleiades center at RA 56.75°, Dec +24.12°. What stands out?
Agent: selects Gaia → inspects metadata → runs bounded ADQL →       returns catalog rows, exact ADQL, units, and assumptions

Representative captured result:

Service: ESA Gaia ArchiveTable: gaiadr3.gaia_sourceRows returned: 10Query limit: TOP 10 / MAXREC 10
Highest returned proper motions:- Gaia DR3 66780900298410496: 244.48 mas/yr- Gaia DR3 66524409149512064: 184.50 mas/yr
Exact ADQL:SELECT TOP 10 source_id, ra, dec, pm, pmra, pmdec, parallax, parallax_error,  phot_g_mean_mag, bp_rp, ruweFROM gaiadr3.gaia_sourceWHERE CONTAINS(POINT('ICRS', ra, dec), CIRCLE('ICRS', 56.75, 24.12, 0.5)) = 1  AND pm IS NOT NULLORDER BY pm DESC

Gaia, SIMBAD, VizieR, the NASA Exoplanet Archive, and IRSA are available as built-in service presets. Agents can also discover and query other public TAP services by URL. Starfetch remains TAP-native and keeps ADQL visible, so the agent workflow is convenient without becoming a scientific black box.

Start here

• Connect an agent
• What the agent does
• Why Starfetch?
• Reliability without hidden assumptions
• Optional Starfetch skill
• Supported scope
• CLI and TypeScript
• Run the Gaia proper-motion demo
• Development

Connect an agent

The local npm server is Starfetch's stable public connection path. It runs on your computer and can be registered with any compatible MCP client.

Local npm server

Register Starfetch with the agent client that will launch it. Running the MCP package by itself only starts a stdio server; it does not connect that server to an agent. Starfetch is also discoverable through the official MCP Registry.

Codex

Register Starfetch for the Codex CLI, IDE extension, and ChatGPT desktop app:

codex mcp add starfetch -- npx -y @starfetch-js/mcpcodex mcp list

These Codex surfaces share MCP configuration. See the official Codex MCP documentation.

Claude Code

Register Starfetch in user scope:

claude mcp add --scope user --transport stdio starfetch -- npx -y @starfetch-js/mcpclaude mcp get starfetch

See the official Claude Code MCP documentation.

Cursor

Add this server entry to ~/.cursor/mcp.json for global use or .cursor/mcp.json for one project:

{  "mcpServers": {    "starfetch": {      "command": "npx",      "args": ["-y", "@starfetch-js/mcp"]    }  }}

See the official Cursor MCP documentation.

Other MCP clients can use the same stdio command and arguments through their own server-registration interface:

command: npxargs: -y @starfetch-js/mcp

Restart or reload the client after registration, then ask a normal astronomy question. You should not need to write ADQL or name Starfetch tools in the prompt. Starfetch requires Node.js 22 or newer.

Remote MCP

A remote MCP service runs on the internet and accepts MCP connections over HTTPS. No Starfetch package has to be installed or launched on the user's computer. Compatible clients can connect to the production endpoint:

https://starfetch-mcp-app-production-wumkqeiqaa-ew.a.run.app/mcp

The remote service is public, unauthenticated, bounded, and best effort. Do not send credentials, private archive URLs, sensitive personal data, or unbounded workloads. The local npm server remains the default when you want local control or when a client does not support remote Streamable HTTP MCP.

See the remote service's overview, privacy notice, support, and terms of use.

What the agent does

For a service-specific catalog question, Starfetch guidance teaches the agent to:

1. choose an explicit service preset or public TAP URL;
2. check service availability when appropriate;
3. inspect relevant tables and the selected table's columns;
4. construct ADQL only from discovered schema information;
5. bound exploratory work with ADQL TOP, TAP MAXREC, or both;
6. execute the smallest useful query;
7. return the service, table, exact ADQL, effective limit, format, units, and relevant assumptions;
8. return to metadata after a schema or syntax failure instead of guessing.

The agent should never present a timeout, availability failure, parse error, or query error as an empty scientific result. A successful zero-row result and a failed request are different outcomes.

Starfetch exposes tools for the complete workflow:

starfetch_list_presetsstarfetch_registry_searchstarfetch_tap_availabilitystarfetch_tap_capabilitiesstarfetch_tap_tablesstarfetch_tap_columnsstarfetch_tap_querystarfetch_tap_submit_jobstarfetch_tap_job_statusstarfetch_tap_job_waitstarfetch_tap_job_fetchstarfetch_tap_job_delete

Query tools return result data separately from diagnostics. They preserve the exact submitted ADQL and effective row limit for reproduction and review. Synchronous queries and async submissions send TAP MAXREC=100 when maxrec is omitted.

The remote HTTP service returns an opaque jobCapability with each async submission. Supply that value unchanged to every remote status, wait, fetch, or delete call. The capability remains usable while the remote job exists and the service signing secret is unchanged. Remote deletion is marked destructive so MCP clients can obtain user approval. The stdio MCP server and CLI keep their existing job-reference behavior and do not advertise remote-only fields.

Why Starfetch?

Starfetch is a useful middle layer when an agent needs live public catalog data without turning the workflow into a black box:

• it inspects live schemas instead of guessing table and column names;
• it bounds public-service queries by default and preserves service failures;
• it returns exact ADQL, limits, units, and assumptions for reproduction;
• it provides one metadata-first interface across several TAP services; and
• its CLI and TypeScript library can reproduce an agent's query outside the agent client.

Use an archive's own interface, PyVO/Astropy, or local analysis tools instead when you need authenticated/private archives, extensive local analysis, image data processing, or authoritative astrophysical interpretation.

Reliability without hidden assumptions

The MCP server works without installing a filesystem skill. Starfetch carries the same canonical guidance through three overlapping layers:

The server exposes the retrievable prompts query_astronomy_catalog, explore_service, run_cone_search, and troubleshoot_adql. Canonical Markdown resources are available under starfetch://guides/, starfetch://services/, and starfetch://examples/.

Prompt and resource support depends on the MCP client. Tool descriptions remain self-sufficient for basic safe operation when a client exposes tools only. The optional skill contains the full workflow, service references, and examples.

Optional Starfetch skill

Install the skill when the agent client supports filesystem skills and you want the strongest multi-interaction behavior. The skill is recommended, not required by the MCP server.

Install the canonical skill directly from GitHub through skills.sh:

npx skills add https://github.com/starfetch-js/starfetch/tree/main/packages/skill/skill/starfetch

Inspect or install the packaged skill:

npx -y @starfetch-js/cli skill printnpx -y @starfetch-js/cli skill install --target codexnpx -y @starfetch-js/cli skill install --target claude-code --scope projectnpx -y @starfetch-js/cli skill install --target cursor

Install into a custom final skill directory with:

npx -y @starfetch-js/cli skill install --path ./starfetch-skill

Use --dry-run to preview file actions. Default destinations are:

• Codex user scope: ~/.codex/skills/starfetch
• Claude Code user scope: ~/.claude/skills/starfetch
• Codex project scope: .codex/skills/starfetch
• Claude Code project scope: .claude/skills/starfetch
• Cursor project scope: .cursor/rules/starfetch.mdc

Codex and Claude Code default to user scope. Cursor defaults to project scope because its rules are project files.

Supported scope

Starfetch is designed for public astronomical Table Access Protocol services. It currently provides:

• built-in presets for gaia, simbad, vizier, exoplanetarchive, and irsa;
• VO registry search for additional TAP endpoints;
• VOSI availability, capabilities, table, and column metadata;
• bounded synchronous ADQL queries;
• explicit TAP/UWS async job submission, status, wait, fetch, and deletion;
• VOTable, CSV, and TSV requests, plus safe JSON and JSONL conversion;
• exact query and limit diagnostics.

Starfetch does not accept credentials or implement authenticated TAP workflows. The MCP server does not execute shell commands or write local result files. Starfetch retrieves catalog data; it does not validate astrophysical interpretations or reconcile scientific differences between catalogs.

VOTable TABLEDATA and inline base64 BINARY/BINARY2 rows can be converted. VOTable FITS rows, remote streams, and compressed streams remain pass-through or unsupported for local row conversion.

CLI and TypeScript

MCP is the primary agent interface. The CLI is useful for scripting, inspecting a query outside an agent, and reproducing the exact request an agent reported. The TypeScript library supports applications and custom adapters.

CLI quickstart

Install or run the CLI once:

npm install -g @starfetch-js/clinpx -y @starfetch-js/cli tap tables --service gaia

Inspect metadata before writing service-specific ADQL:

starfetch tap availability --service gaiastarfetch tap tables --service gaiastarfetch tap columns --service gaia --table gaiadr3.gaia_source

Run a bounded query:

starfetch tap query \  --service gaia \  --query "SELECT TOP 5 source_id, ra, dec FROM gaiadr3.gaia_source" \  --format json

ADQL can come from --query, a file, or stdin. Result data can be written with --out:

starfetch tap query --service gaia --file query.sql --format csv --out result.csvcat query.sql | starfetch tap query --service gaia --format jsonl

Use --service for a preset or --url for an explicit TAP base URL. If both are supplied, --url selects the endpoint and the service name remains as diagnostic context.

Discover additional services through the VO registry:

starfetch tap registry search gaia --maxrec 5 --format json

Async jobs

Use explicit async jobs for larger justified queries:

starfetch tap jobs submit \  --service gaia \  --query "SELECT TOP 10 source_id, ra, dec FROM gaiadr3.gaia_source" \  --maxrec 10
starfetch tap jobs status <job-url>starfetch tap jobs wait --interval 2000 --timeout 120000 <job-url>starfetch tap jobs fetch <job-url> --format votable --out result.xmlstarfetch tap jobs delete <job-url>

Absolute job URLs are sufficient for follow-up commands. Bare job IDs require --service or --url so Starfetch can resolve the TAP /async endpoint.

TypeScript API

Install @starfetch-js/core when a script, app, or custom agent adapter needs direct TAP access:

npm install @starfetch-js/core

import { registry, tap } from "@starfetch-js/core";
const client = tap("gaia");const columns = await client.columns("gaiadr3.gaia_source");const result = await client.query(  "SELECT TOP 5 source_id, ra, dec FROM gaiadr3.gaia_source",  { format: "votable", maxrec: 5 },);
console.log(columns.length);console.log(await result.fields());console.log(await result.overflow());console.log(await result.json());
const services = await registry().searchTapServices({  query: "gaia",  maxrec: 5,});console.log(services[0]?.accessUrl);

tap(target) accepts a known preset, a TAP base URL, or an object containing a service and/or URL. Metadata methods read TAP /availability, /capabilities, and /tables; sync queries use /sync, and explicit jobs use /async.

Run the Gaia proper-motion demo

Reproduce the demo's metadata-first Gaia query and print the exact ADQL, effective limit, and returned rows:

git clone https://github.com/starfetch-js/starfetch.gitcd starfetchnpm cinpm run buildnode examples/quickstart/run.mjs

This command queries the public Gaia TAP service. For more CLI, TypeScript API, live TAP, and MCP Inspector workflows, see starfetch-js/examples. Each example includes its exact ADQL, expected columns, and a cross-platform Node.js runner.

git clone https://github.com/starfetch-js/examples.gitcd examplesnpm cinode 01-gaia-nearby-stars/run.mjs

Launch MCP Inspector from that repository with:

npm run inspect:mcp

Packages

• @starfetch-js/mcp: primary stdio MCP server and packaged agent guidance.
• @starfetch-js/skill: optional distributable Starfetch agent skill.
• @starfetch-js/cli: scripting, TAP inspection, query, async job, and skill installation commands.
• @starfetch-js/core: reusable TAP, VOSI, UWS, VOTable, registry, and output conversion primitives.

Development

Install dependencies with the committed lockfile:

npm ci

Run the private Streamable HTTP development app with:

npm run dev:http

It serves MCP at http://127.0.0.1:3000/mcp and process health at http://127.0.0.1:3000/health. Each MCP request gets a fresh stateless @starfetch-js/mcp server. The HTTP surface keeps the 12 canonical Starfetch tools unchanged and adds two UI-linked tools through the immutable ui://starfetch/table/v1 MCP Apps resource. starfetch_render_table presents an existing table view without submitting a TAP request. starfetch_query_table runs a synchronous JSON query for the interactive scientific table; its full loaded table is widget-only while the model receives a 20-row preview. Ordinary MCP clients can continue using the canonical text and structured results without loading the widget. Configuration is environment-only:

• HOST defaults to 127.0.0.1; set 0.0.0.0 explicitly for all interfaces.
• PORT defaults to 3000.
• ALLOWED_ORIGINS is a comma-separated list of exact browser origins and defaults to none.
• SHUTDOWN_GRACE_MS defaults to 10000 and accepts 1 through 60000.
• STARFETCH_JOB_CAPABILITY_SECRET is a base64url secret of at least 32 bytes. It is required when HOST is not loopback; loopback development uses an ephemeral per-process secret when omitted.

The anonymous HTTP policy is a fixed, tested product profile rather than a set of independently tunable environment variables. It caps MCP requests at 2 MiB, TAP responses at 8 MiB, inline uploads at 1 MiB, MAXREC at 10,000, redirects at 3, outbound requests at 4 concurrent operations, and tools at 60 seconds. Canonical synchronous queries still default to MAXREC=100; the interactive table query defaults to MAXREC=1000 and caps its widget payload at 6 MiB. Job waits default to 30 seconds, cap at 45 seconds, and poll between 1 and 10 seconds. The process admits 100 MCP requests per minute globally; deployments that need per-client limits should enforce them at a trusted HTTPS ingress.

The HTTP host accepts only credential-free HTTPS TAP targets whose complete DNS result is public, pins validated addresses for each request hop, keeps redirects same-origin, never automatically follows write redirects, and rejects remote-URI TAP uploads. These restrictions apply to the anonymous HTTP app, not the local stdio MCP, CLI, or TypeScript API.

To verify the protocol surface without opening the widget, start the app and run MCP Inspector's CLI in another shell:

npx -y @modelcontextprotocol/inspector@latest --cli http://127.0.0.1:3000/mcp --transport http --method tools/listnpx -y @modelcontextprotocol/inspector@latest --cli http://127.0.0.1:3000/mcp --transport http --method tools/call --tool-name starfetch_list_presets

To render the widget in MCP Inspector's Apps tab:

1. Keep npm run dev:http running, then start the Inspector UI without a stdio server command:

   sh

   npx -y @modelcontextprotocol/inspector@latest

2. In Inspector, select Via Proxy, choose Streamable HTTP, enter http://127.0.0.1:3000/mcp, leave authentication empty, and connect.

3. Open Apps, select Refresh Apps, and choose starfetch_render_table. The Apps tab lists UI-linked tools rather than the server name.

4. Paste a valid bounded table view into App Input and select Open App. This minimal preset view exercises the widget without making a live TAP request:

   json

   {  "contractVersion": 1,  "resultKind": "presets",  "title": "TAP service presets",  "columns": [    { "key": "name", "label": "Name" },    { "key": "url", "label": "TAP URL" }  ],  "rows": [    {      "name": "gaia",      "url": "https://gea.esac.esa.int/tap-server/tap"    }  ],  "source": { "tool": "starfetch_list_presets" },  "state": "populated",  "clipping": {    "reasons": [],    "sourceRows": 1,    "sourceColumns": 2  }}

The Apps tab requires the Streamable HTTP endpoint; the canonical stdio server exposes the core Starfetch tools without UI resources. Inspector's Via Proxy mode works with the default origin policy. To use Direct mode, allow Inspector's browser origins explicitly when starting the app:

ALLOWED_ORIGINS=http://localhost:6274,http://127.0.0.1:6274 npm run dev:http

The widget uses the standard MCP Apps host bridge for tool results, theme variables, display-mode requests, model-context updates, and file downloads. It also reads ChatGPT's initial tool globals when present and uses ChatGPT's file APIs only as a download fallback. Its React UI uses semantic table markup, global TanStack Table sorting, page-scoped row selection, 100-row desktop pages, 10-row mobile pages, and fine-grained Shiki SQL highlighting for exact ADQL. Desktop tables use a capped two-axis scroll viewport; mobile tables expand to the page height and retain horizontal scrolling for wide results. An icon-only control in the action toolbar requests fullscreen or inline display mode. Another action can expose selected rows or the current page to the model for the next turn; it does not expose the whole loaded table. The action remains available, and each press replaces the previously pending model-context snapshot. The resource declares no network or static-resource domains and requests only clipboard-write permission.

Run its unit, single-file build, and browser-host acceptance checks with:

npm --workspace @starfetch-js/mcp-app run checknpm --workspace @starfetch-js/mcp-app run test:browser

Build and smoke-test the production Linux container with Docker:

npm run smoke:container

The image workflow publishes immutable containers to ghcr.io/starfetch-js/starfetch-mcp-app.

For ChatGPT Developer Mode or another remote MCP Apps host, expose the local MCP endpoint through HTTPS, add the resulting /mcp URL to the host, call a metadata tool to inspect the target, then call starfetch_query_table. Verify light and dark themes, fullscreen and inline modes, horizontal scrolling, sorting, paging, highlighted ADQL and copying, current-page model context, and TSV, CSV, and JSON copy and download actions. Host-specific visual differences should be handled through the bridge theme variables rather than a second component implementation.

For a temporary remote URL, the development machine can run:

cloudflared tunnel --url http://127.0.0.1:3000 --http-host-header 127.0.0.1:3000

The HTTP app has no authentication. It applies a coarse, per-process fixed window limit of 100 MCP requests per minute, not per-client abuse protection. A quick tunnel is public and temporary: use only non-sensitive test traffic and stop it immediately after testing. This development app is not production deployment infrastructure.

The workspace requires Node.js >=22.13.0. Run:

npm run format:checknpm run lintnpm run typechecknpm testnpm run coveragenpm run buildnpm run smoke:cli

Release-sensitive agent surface checks are:

npm --workspace packages/mcp run typechecknpm --workspace packages/mcp run testnpm --workspace packages/mcp run buildnpm --workspace packages/mcp run smoke
npm --workspace @starfetch-js/mcp-app run check
npm --workspace packages/skill run typechecknpm --workspace packages/skill run testnpm --workspace packages/skill run build

Default tests use local fixtures and mocks only. Optional live TAP checks are explicit:

npm run test:live:tap

License

MIT