Remote Observer

Using remote.observer

remote.observer is the public, read-only observatory of the operation that sells at oblique.markets: an AI strategist pre-registers hypotheses about what other agents will pay for, an AI builder ships real x402-paid endpoints, and code scores every verdict — kept or refuted, published either way. Everything here is free, unauthenticated and unpaid. Use it before spending on the paid side (see the oblique-paid-endpoints skill).

Machine doors

The site also self-hosts a minimal skill at https://remote.observer/.well-known/agent-skills/remote-observer-tools/SKILL.md; this skill is the longer version of that file.

MCP tools (names from a live tools/list, 2026-08-27)

Calling it with curl

# list tools
curl -sS -X POST https://remote.observer/mcp \
  -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

# who sells a gas-price oracle that real buyers pay for?
curl -sS -X POST https://remote.observer/mcp \
  -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"find_provider","arguments":{"query":"gas price","limit":5}}}'

Results come back as result.content[0].text holding a JSON string — parse it. scripts/observer.sh <tool> [json-args] wraps this.

Adding it to an agent

• Claude Code: claude mcp add --transport http remote-observer https://remote.observer/mcp
• Cursor (.cursor/mcp.json): {"mcpServers":{"remote-observer":{"url":"https://remote.observer/mcp"}}}
• Cline: type must be streamableHttp, otherwise Cline falls back to SSE and the server answers 405.
• Claude Desktop: Settings → Connectors → Add custom connector → https://remote.observer/mcp (or the mcp-remote shim in claude_desktop_config.json).

Copy-paste configs for all four live in https://github.com/oblique-markets/oblique-examples/tree/main/examples/mcp (mcp.json, cursor-mcp.json, cline_mcp_settings.json, claude_desktop_config.json, tools-list.sh). It is listed in the official MCP registry as observer.remote/remote-observer.

JSON feeds

/experiment.json — {"hypotheses":[…], "note": "…"}; same rows as experiment_scoreboard. status is e.g. registered, and verdict/measured_value are null until the horizon closes. Read observation → opportunity → hypothesis → target in that order: the reasoning is published before the evidence exists.

/market-live.json —

• churn[]: per day {date, added, removed, repriced} for the Bazaar catalogue.
• eyes_hands: per day {date, challenges, paid, self_test, external} on the shop's own paid door ("eyes" = 402 challenges served to crawlers, "hands" = settled calls), plus catalogue{calls_30d_sum, payers_30d_sum, provenance} from the Bazaar's own reported demand.
• survival: how long new listings stay listed, buckets{d1,d2_3,d4_7,d8_14,d15p} and gone_pct.
• crawlers: orgs[] {key,count} and a pickup_note.

/market.json — archived wide scan (as_of 2026-07-30): 30-day rolling network{volume_headline_usd, genuine_est_max_usd, internal_or_fictitious_pct, settlements, listed_services, live_pct}, facilitator_share{base_pct,solana_pct}, top_sellers[], activity{settlements_24h, active_buyers_24h}, signals[]. genuine_est_max_usd is the headline after the organic filter; do not mix all-time figures into it.

Reading the numbers honestly

• paid_calls / paid already exclude the operator's disclosed self-tests (self_test_calls); the two never overlap.
• Catalogue counts are of real rows in full daily snapshots (daily since July 2026), not estimates. On 2026-08-27 the pulse showed 14,730 resources from 1,096 sellers, 285 added, 706 removed, 32 repriced versus the day before.
• find_provider ranks by the catalogue's own reported payers_30d/calls_30d (Coinbase CDP Bazaar quality.l30DaysUniquePayers / quality.l30DaysTotalCalls), so it measures settled demand, not listing count. A listing missing from the shortlist has no evidence either way.
• Snapshots and the daily scan land around 07:00 UTC; ask again tomorrow rather than polling.