Loading ToolBench
Exposes ASTRA epistemological engine as MCP tools for scientific hypothesis validation and deliberative research cycles. Enables any agent (Claude, Codex, Gemini) to verify physics/math conjectures through symbolic solvers, numerical computation, and formal proof systems.
ASTRA exposes 7 tools with detailed descriptions (50-300 chars), but critical gaps undermine production readiness. All tools have input schemas with types and descriptions, meeting baseline requirements. However, output schemas are entirely undocumented, LLMs cannot plan downstream calls or extract structured data. Tool naming is domain-specific but lacks action verbs (astra_execute, astra_cycle, astra_job are acceptable but not verb_noun patterns). Error handling is absent: no guidance on retryability, user-fixable vs fatal errors, or recovery paths. No parameter validation rules, enums, or constraints visible. Security: code execution tools (astra_execute, astra_submit) accept arbitrary Python scripts with no sandboxing hints, timeout enforcement, or injection guards documented. Composition: tools are well-separated (execute vs cycle vs job polling), but no idempotency guarantees or chaining documentation.
Query the available computational capacity: visible CPU cores, threads, and the safe parallelism policy for the current oracle.
Run ASTRA's minimum application-facing assurance package. The validator router sends formal invariants to pinned Lean 4, constraints to Z3, symbolic formulas to SymPy, numerical models to SciPy, dimensional checks to Pint, and project cases to their scientific package. Every result includes an explicit claim verdict, artifact hash, assumptions, limitations, structured evidence, provenance and a reproduction command.
Run ASTRA's FULL deliberative multi-model pipeline and return a verdict. Codex and agy propose and cross-critique hypotheses against a shared final objective; Codex synthesizes the consensus. Claude writes the falsifiable validation program, Codex independently reviews it, the selected oracle executes it, Codex audits the evidence, and agy proposes the next direction. Use astra_execute when YOU wrote the code and only need the oracle. Slower than astra_execute. Compact cycles may finish in several minutes; adversarial scientific audits commonly require 15-30 minutes.
Queue a FULL ASTRA deliberative cycle as a persistent background job. This is the production route for complex scientific audits. It is not bound by the synchronous MCP wall, checkpoints every completed phase, waits for the single shared model-account slot, and survives the calling task. Poll the returned job_id with astra_job.
Output schemas completely undocumented. LLMs cannot infer what fields astra_cycle, astra_job, or astra_client_validate return, blocking downstream tool chaining and data extraction.
Code execution tools (astra_execute, astra_submit) accept arbitrary Python with no documented sandboxing, injection guards, or resource limits. Timeout parameters exist but no guidance on what happens if code hangs or consumes unbounded memory.
No error handling guidance. Tools lack recovery hints, retryability classification, or actionable error messages. An LLM receiving 'timeout' or 'execution failed' has no next step.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Run a verification script through ASTRA's oracle and return real results. Use this to VERIFY a physics/math hypothesis with actual computation instead of trusting an LLM's judgment. Write a self-contained script (sympy, einsteinpy, z3, scipy, numpy, qutip, pint, in-house packages, or a Sage/Maxima/Cadabra/Lean script with an '# ASTRA_ENGINE: ...' marker) that prints its evidence and ends with a line 'VERDICT: PASS' or 'VERDICT: FAIL'.
Poll the status and result of a background job submitted via astra_cycle_submit or astra_submit. Returns job metadata, current status, and (when done) the full scientific result.
Queue a long-running Python script as a persistent background job. The job is decoupled from the MCP server and survives the calling process. Use this for computation that exceeds the synchronous MCP wall (~15 min). Poll with astra_job.
Parameter constraints missing. 'oracle' accepts 'local|astrum|auto' but no enum declared. 'timeout' is unbounded integer, no min/max guidance. LLMs may pass invalid values.
Tool naming lacks action verbs. 'astra_execute' and 'astra_cycle' are domain-specific but do not follow verb_noun convention (e.g., 'run_verification', 'start_deliberation'). Reduces LLM intent inference.