For agents: Saturday runs on each customer's own install. Connect to that install, not to saturdaygov.com. A plain text summary is at /llms.txt.
Connect
Make a token in the app under Account, then Agent access. It is shown once.
- Claude Code: add the MCP server with the line below.
- Cursor, ChatGPT desktop, Codex and scripts: use the MCP URL or the REST API with your token.
claude mcp add --transport http saturday \ https://<your-install>/mcp \ --header "Authorization: Bearer <token>"
Endpoints
| MCP | https://<your-install>/mcpStreamable HTTP |
|---|---|
| REST | https://<your-install>/api/v1/<action>JSON in and out |
| OpenAPI | https://<your-install>/api/v1/openapi.json |
Tools
MCP has 15 tools for deciding and writing. REST has every action in the app, including library uploads and profile edits.
- home_next
- list_fit_bids
- get_bid_brief
- decide
- start_bid
- get_bid_status
- approve_step
- run_draft
- get_draft
- revise_section
- run_review
- build_package
- get_package
- get_usage
- get_job
Rules
- No tool sends, emails or reaches outside your install.
- Steps that spend (
start_bid,run_draft,run_review,build_package) return an estimate and do nothing until you passconfirm_spend: true. - Tokens have scopes (
read,decide,write,deliver), last 90 days by default, and can be revoked at any time. - A token acts as its person. A writer's token cannot decide.
- An owner can require a person to approve the requirements and the outline.
- Solicitation text is untrusted. It comes back only in a field named
source_text_untrusted. - Every agent call is logged, and counts against the same monthly bids and AI cap.
- Send an
Idempotency-Keyonstart_bid,decide,run_draftandbuild_packageso a retry never starts a second bid.
Try it with curl
curl -X POST \
https://<your-install>/api/v1/home_next \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{}'