Starfetch gives AI agents safe, reproducible access to public astronomy
catalogs through the
[Model Context Protocol (MCP)](https://modelcontextprotocol.io/docs/getting-started/intro).
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.
```text
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:
```text
Service: ESA Gaia Archive
Table: gaiadr3.gaia_source
Rows returned: 10
Query 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, ruwe
FROM gaiadr3.gaia_source
WHERE CONTAINS(POINT('ICRS', ra, dec), CIRCLE('ICRS', 56.75, 24.12, 0.5)) = 1
AND pm IS NOT NULL
ORDER 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](#connect-an-agent)
- [What the agent does](#what-the-agent-does)
- [Why Starfetch?](#why-starfetch)
- [Reliability without hidden assumptions](#reliability-without-hidden-assumptions)
- [Optional Starfetch skill](#optional-starfetch-skill)
- [Supported scope](#supported-scope)
- [CLI and TypeScript](#cli-and-typescript)
- [Run the Gaia proper-motion demo](#run-the-gaia-proper-motion-demo)
- [Development](#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](https://registry.modelcontextprotocol.io/?q=io.github.starfetch-js%2Fstarfetch).
#### Codex
Register Starfetch for the Codex CLI, IDE extension, and ChatGPT desktop app:
```sh
codex mcp add starfetch -- npx -y @starfetch-js/mcp
codex mcp list
```
These Codex surfaces share MCP configuration. See the
[official Codex MCP documentation](https://developers.openai.com/codex/mcp).
#### Claude Code
Register Starfetch in user scope:
```sh
claude mcp add --scope user --transport stdio starfetch -- npx -y @starfetch-js/mcp
claude mcp get starfetch
```
See the
[official Claude Code MCP documentation](https://docs.anthropic.com/en/docs/claude-code/mcp).
#### Cursor
Add this server entry to `~/.cursor/mcp.json` for global use or
`.cursor/mcp.json` for one project:
```json
{
"mcpServers": {
"starfetch": {
"command": "npx",
"args": ["-y", "@starfetch-js/mcp"]
}
}
}
```
See the
[official Cursor MCP documentation](https://docs.cursor.com/context/model-context-protocol).
Other MCP clients can use the same stdio command and arguments through their
own server-registration interface:
```text
command: npx
args: -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:
```text
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](https://github.com/starfetch-js/starfetch/blob/HEAD/apps/mcp-app/README.md),
[privacy notice](https://github.com/starfetch-js/starfetch/blob/HEAD/apps/mcp-app/PRIVACY.md),
[support](https://github.com/starfetch-js/starfetch/blob/HEAD/apps/mcp-app/SUPPORT.md), and
[terms of use](https://github.com/starfetch-js/starfetch/blob/HEAD/apps/mcp-app/TERMS.md).
## 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:
```text
starfetch_list_presets
starfetch_registry_search
starfetch_tap_availability
starfetch_tap_capabilities
starfetch_tap_tables
starfetch_tap_columns
starfetch_tap_query
starfetch_tap_submit_job
starfetch_tap_job_status
starfetch_tap_job_wait
starfetch_tap_job_fetch
starfetch_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:
| Layer | Role |
| ------------------------- | ------------------------------------------------------------------------------- |
| MCP tool descriptions | Minimum metadata-first and bounded-query contract available to every MCP client |
| MCP prompts and resources | Discoverable workflows, ADQL guidance, service notes, and examples |
| Optional Starfetch skill | Rich multi-step behavior across longer agent interactions |
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:
```sh
npx skills add https://github.com/starfetch-js/starfetch/tree/main/packages/skill/skill/starfetch
```
Inspect or install the packaged skill:
```sh
npx -y @starfetch-js/cli skill print
npx -y @starfetch-js/cli skill install --target codex
npx -y @starfetch-js/cli skill install --target claude-code --scope project
npx -y @starfetch-js/cli skill install --target cursor
```
Install into a custom final skill directory with:
```sh
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:
```sh
npm install -g @starfetch-js/cli
npx -y @starfetch-js/cli tap tables --service gaia
```
Inspect metadata before writing service-specific ADQL:
```sh
starfetch tap availability --service gaia
starfetch tap tables --service gaia
starfetch tap columns --service gaia --table gaiadr3.gaia_source
```
Run a bounded query:
```sh
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`:
```sh
starfetch tap query --service gaia --file query.sql --format csv --out result.csv
cat 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:
```sh
starfetch tap registry search gaia --maxrec 5 --format json
```
### Async jobs
Use explicit async jobs for larger justified queries:
```sh
starfetch tap jobs submit \
--service gaia \
--query "SELECT TOP 10 source_id, ra, dec FROM gaiadr3.gaia_source" \
--maxrec 10
starfetch tap jobs status
starfetch tap jobs wait --interval 2000 --timeout 120000
starfetch tap jobs fetch --format votable --out result.xml
starfetch tap jobs delete
```
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:
```sh
npm install @starfetch-js/core
```
```ts
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:
```sh
git clone https://github.com/starfetch-js/starfetch.git
cd starfetch
npm ci
npm run build
node 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](https://github.com/starfetch-js/examples). Each example
includes its exact ADQL, expected columns, and a cross-platform Node.js runner.
```sh
git clone https://github.com/starfetch-js/examples.git
cd examples
npm ci
node 01-gaia-nearby-stars/run.mjs
```
Launch MCP Inspector from that repository with:
```sh
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:
```sh
npm ci
```
Run the private Streamable HTTP development app with:
```sh
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:
```sh
npx -y @modelcontextprotocol/inspector@latest --cli http://127.0.0.1:3000/mcp --transport http --method tools/list
npx -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:
```sh
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:
```sh
npm --workspace @starfetch-js/mcp-app run check
npm --workspace @starfetch-js/mcp-app run test:browser
```
Build and smoke-test the production Linux container with Docker:
```sh
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:
```sh
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:
```sh
npm run format:check
npm run lint
npm run typecheck
npm test
npm run coverage
npm run build
npm run smoke:cli
```
Release-sensitive agent surface checks are:
```sh
npm --workspace packages/mcp run typecheck
npm --workspace packages/mcp run test
npm --workspace packages/mcp run build
npm --workspace packages/mcp run smoke
npm --workspace @starfetch-js/mcp-app run check
npm --workspace packages/skill run typecheck
npm --workspace packages/skill run test
npm --workspace packages/skill run build
```
Default tests use local fixtures and mocks only. Optional live TAP checks are
explicit:
```sh
npm run test:live:tap
```
## License
MIT
