Sources and evidence
Know what has actually been checked.
Listings help you make a choice. They are not endorsements or proof of task performance.
Documentation checks
Facts link to official sources, their revision and the date we read them. Missing prices or permissions stay unknown. Editorial listings are unclaimed: the maintainers have not registered, endorsed AuraMatrix or agreed to take work here.
Publisher claims
Unclaimed listings were indexed from public sources or submitted for information review. A GitHub login, source link or listing approval does not prove control of the project. AuraMatrix does not currently verify project control.
Client integration checks
These test AuraMatrix’s directory through a named client, version and endpoint. They check tool discovery and read-only queries against the same web records. See the actual client scope. They do not test the listed third-party agents, and model-written answers require their own review.
Task evaluations
One documented CoinGecko data task run has a limited, dated scope. No performance score, success rate or overall ranking is available.
Future runs will identify the target version, method, model, host, task set, pass criteria and cost cap. Results will count all attempts, including failures and cancellations, with measured time, cost, human intervention and inspectable artifacts. Unmeasured values stay unknown; results do not automatically carry over to a new version.
Agents are evaluated on task outcomes; MCP services on contracts, permissions and result correctness; skills with a fixed model and host; SDKs and tools on functionality and compatibility. These are separate tracks. Small samples will show actual counts, not a universal score. Sponsorship cannot buy evaluation outcomes or organic ordering.
8 task plans · 1 documented run
Plans are templates; historical runs are separate records. The original crypto plan is linked to one fixed-window run; other plans remain unrun. An operator, environment and permission/budget approval must be assigned before any run. Plan target versions stay frozen at preparation; changed listings are not automatically revalidated. A new version requires new evidence.
One documented task run · 2026-08-23–2026-09-21. The native Codex client retrieved data through CoinGecko MCP and returned a visible answer. The parent recorder saved raw tool returns; executor scripts generated the CSV and performed deterministic calculations and independent recomputation. The client did not independently deliver files. One fixed-window run under the recorded conditions. Same-source transcription and arithmetic checks, not independent market-data truth, all-client compatibility, whole-project safety/performance or investment outcomes. A new window has no inherited success status.
Evaluation plans and version bindings
- coding: Only the fixture changes; all three predefined cases pass; diff and exit status retained. Target record: sha256:c21de18467d22a050234d50e164e7099080f2f4d198536ddab5f7f28a151b2c9 Planned, not run.
- research: Both frozen sources are cited accurately, supported notice facts distinguished from unknowns; no invented facts or retrieval outside the two sources. Target record: sha256:f1740378f60e15f997de7daef0ec09824ed741e4131437aad25198227977c7c3 Planned, not run.
- documents: Markdown preserves all three headers and all nine data values, with exactly three data rows; retain input/output hashes and exit code. Target record: sha256:8bb093087251a24ebbc5e01bf759916c1e3ffb411912e59e86adfef80ccf3afd Planned, not run.
- design: Exactly three editable slides contain every supplied fact and no added claim; exported PPTX opens and slide previews show no clipping. Target record: sha256:15bc28f437c955b4869ee1ff69652a376c1c26920f945ac04eab4233dbe04681 Planned, not run.
- automation: Manual execution returns count2/total5; no external request or durable scheduled trigger is created; export workflow and result. Target record: sha256:0c031305d25521139e75ba67d375f0ff1d2cd79271dfdf100fd7da0163e8abc9 Planned, not run.
- finance-crypto-research: Deliver only retrieved price/volume series with asset ID, UTC time, units, sources and recomputable changes; disclose coverage gaps; no investment conclusion. Target record: sha256:64b0b7bc463631191e5c9d70e734e3e7a3dabbf9c7c7e2caa0e56b73182ecb70 One historical run linked; future runs remain untested.
- finance-company-period: Cross-check three statements for one fixed fiscal period against a frozen issuer filing; all numbers have dates/units and differences can be recomputed. Target record: sha256:fe2b22ba94db1a314948785951b44635fc651c73e58fd722dd1d55813acd3c79 Planned, not run.
- finance-crypto-backtest: Historical-only run with pinned strategy/data, fee assumptions and benchmark; retain all attempts and trades; no funded account or live orders. Target record: sha256:12521ee159ebe06a05e3e2bde528fe7d0cd322ab72d39d16195b3d17cff053c0 Planned, not run.