Zero-Trust Physics MCP server
com.zerotrustphysics/verifierzerotrustphysics.comstreamable-httpv0.15.3Scanned
Deterministic physics runtime with a proof on every answer: 61 models, open and keyless over MCP.
18 tools: 12 change things and 6 only read.
- Reads private data · 0
- Reads text others write · 0
- Can send data out · 0
- 5 findings
Tools listed without sign-in. Remote posture
With Anlyon
12 tools here change things. Govern them.
- Anlyon lists the tools and adopts none.
- Each tool you adopt gets approvals and receipts.
- Your agent never holds the key.
A tool your agent calls directly is not routed through Anlyon and is not governed by it. Anlyon Radar only lists and labels, and these labels are the same either way.
What the Zero-Trust Physics MCP server can do
The Zero-Trust Physics MCP server is published by zerotrustphysics.com, a domain the MCP Registry verified. It has 18 tools. It can create or change data (12 tools: compute_step, run_trajectory, verify_tape, solve and 8 more), but cannot delete data. 6 tools only read.
None of its tools reads private data, brings in text other people write, or sends data out.
The fixed checks matched 5 findings (annotation-conflict and tool-shadowing). Each one comes with the text it matched.
Its remote endpoint lists its tools without sign-in, and its tools changed 2 times since Anlyon Radar first saw them.
Last scanned Nov 16, 2026, version 0.15.3.
Tools
Grouped by what each tool does, most consequential first. Open one for its reasons and full description; descriptions are the publisher's own words.
18 tools
Changes things 12Claude Code: ask
citeGives how to cite a card or a solver: the service itself first (Zero-Trust Physics, its maker and its address) and the paper (https://doi.org/10.5281/zenodo.23517223), then the peo…No findings
Why it is in this group
- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Description, as the server wrote it Gives how to cite a card or a solver: the service itself first (Zero-Trust Physics, its maker and its address) and the paper (https://doi.org/10.5281/zenodo.23517223), then the people whose measurements the model stands on (authors, title, year, address, and what each asks of those who use its data), each as text and BibTeX. With domain and card, or solver, the answer is for that model; with neither, it is how to cite the service itself: its name, who made it, its address and the paper.
Annotations the server declares
- title: Cite
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
compute_stepComputes one anchored step of one card and returns the state it lands on, in decimals and exact hex, with its cryptographic proof (sha384_seal) and a receipt.No findings
Why it is in this group
- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Description, as the server wrote it Computes one anchored step of one card and returns the state it lands on, in decimals and exact hex, with its cryptographic proof (sha384_seal) and a receipt. parameters holds the card's phase fields, its load fields and dt_s, as describe_card gives them. A step whose request or whose landing state is outside the card is refused in one sentence, with the refusal's name in round brackets when it has one, and no proof is issued for that state. With from, the step starts from a sealed run, which is computed again first. On this open connection one address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Compute step
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
design_runDesigns a run: from a target end state to the loads of each segment that reach it, computed by the engine with no randomness, so the same request gives the same bits.1 finding
Why it is in this group
- The name "design_run" says "run": a word for a change.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Findings
- Warn
The server marks "design_run" read-only (readOnlyHint: true), and its own name, description or input give the class "write".
What this check looks forMatched text
annotations: {"readOnlyHint":true,"destructiveHint":false,"idempotentHint":true,"openWorldHint":false}; The name "design_run" says "run": a word for a change.
Description, as the server wrote it Designs a run: from a target end state to the loads of each segment that reach it, computed by the engine with no randomness, so the same request gives the same bits. request is a run_trajectory request; its numbers at the places vary names are the start, and vary gives each load's span inside the card's bounds. target is a number of the final state, a value and a tolerance. Every candidate is an ordinary run with its run_seal; one that crosses a limit is recorded refused by its name, and no value is moved back. The answer gives the timeline, its run_trajectory request and run_seal, the value reached and the miss, every candidate in order as a run_dataset plan with its dataset_seal, and a design_seal. Outcome reached; not_reached, meaning this search found none inside its caps, which is not a limit of the physics; or refused: TargetOutsideCard for a target outside the card's range, TargetBeyondReach where the card's map is monotone and a corner of the loads allowed falls short. The search is coordinate bisection, capped at 64 candidates (DESIGN_MAX_CANDIDATES), 1000000 steps (DESIGN_MAX_STEPS) and 10000 recorded frames (DESIGN_MAX_FRAMES) in all, and 4 sweeps (DESIGN_MAX_SWEEPS). verify_tape, given this whole answer as it came, designs it again and compares the two, byte for byte (the long form). On this open connection a design is long work whenever its candidates may take more than 10000 steps together. One address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Design run
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
run_datasetRuns a declared plan of runs on one card, or of solves of one solver, and seals the rows under one root.1 finding
Why it is in this group
- The name "run_dataset" says "run": a word for a change.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Findings
- Warn
The server marks "run_dataset" read-only (readOnlyHint: true), and its own name, description or input give the class "write".
What this check looks forMatched text
annotations: {"readOnlyHint":true,"destructiveHint":false,"idempotentHint":true,"openWorldHint":false}; The name "run_dataset" says "run": a word for a change.
Description, as the server wrote it Runs a declared plan of runs on one card, or of solves of one solver, and seals the rows under one root. A plan is a list of runs, a grid over a base run, or a seeded sample over a base run; a run is a run_trajectory request without domain and card. A plan of solves names solver in place of domain and card and is a list or a grid of solve requests. The answer is small by default: the head, which holds the request, the counts and the dataset_seal, and one summary row per run; answer "states" adds each final state in decimals and "full" gives every row whole, under the same dataset_seal. A row is completed, ended, refused or unadmitted: an ended or a refused row is a run sealed up to its last landed step, and an unadmitted row is a request turned away before computing, with no seal. The seal shows the dataset was not altered after it was sealed. It does not show the engine made it: a row's request sent to run_trajectory gives the run seal to compare. The input schema says how each input is written, with a worked plan of each kind. On this open connection a dataset is returned inline and holds at most 300 runs (OPEN_MAX_DATASET_ROWS), 1000000 steps in all (OPEN_MAX_DATASET_STEPS), 10000 frames and an answer counted from the plan at 4194304 bytes (OPEN_MAX_DATASET_ANSWER_BYTES) and 42500 characters, up to about 150 summary rows; a plan of solves holds the engine's own 64 and at most 15 s of them, or one, at each solver's measured time, which describe_solver states with the count (OPEN_PLAN_SECONDS). One address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Run dataset
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
run_trajectoryRuns one card from a start state through segments of steps, each under its own load, and seals the run.1 finding
Why it is in this group
- The name "run_trajectory" says "run": a word for a change.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Findings
- Warn
The server marks "run_trajectory" read-only (readOnlyHint: true), and its own name, description or input give the class "write".
What this check looks forMatched text
annotations: {"readOnlyHint":true,"destructiveHint":false,"idempotentHint":true,"openWorldHint":false}; The name "run_trajectory" says "run": a word for a change.
Description, as the server wrote it Runs one card from a start state through segments of steps, each under its own load, and seals the run. The answer holds the outcome, the final state in decimals and exact hex, a frame for each recorded step, the run_seal and a receipt. record_every sets which steps are recorded, about 262 characters each; the final state comes back in full whatever is recorded. readout gives named quantities in decimals without the frames, with each one's minimum and maximum over every landed step and the first step past a level; it is not part of the sealed request. from starts the run from a sealed run, which is computed again first. A run comes back in one of four ways. outcome "completed": every step landed. outcome "ended": the run reached an end its card declares, which describe_card lists under declared_ends; reject holds the end's name. outcome "refused": a step part way was refused; reject holds the name (null when the cause has none) and reject_detail the sentence. An ended or a refused run is sealed up to its last landed step, steps_landed. An error, not a result: the request was turned away before computing, in one sentence with the refusal's name in round brackets when it has one, and nothing is sealed. A sentence that begins "Physical Failure" says the state asked for or reached is outside the card; one that begins "Physical Violation" says the request, or a number in it or computed from it, is not one the engine takes. A load may be signed; load_source says by whom. The input schema says each of these in full. On this open connection one address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Run trajectory
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
solveSolves one field problem with one solver and seals the solution: the outcome, the result in decimals and exact hex, its invariants, the solution_seal and a receipt.No findings
Why it is in this group
- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Description, as the server wrote it Solves one field problem with one solver and seals the solution: the outcome, the result in decimals and exact hex, its invariants, the solution_seal and a receipt. parameters holds the solver's inputs as describe_solver gives them. example, in their place, is the name of a worked example describe_solver lists (example_solve, or a key beginning example_solve_), so a long worked input need not be typed again; the request is filled with that example's parameters, and the filled request is the one solved, sealed and returned as inputs. The receipt covers the request as it was written, number spellings included, while a seal covers its values: a request sent with 310 and the same request sent with 310.0 give different receipts and the same seal, so a receipt that differs between two such answers is not a sign of tampering. A chained solve, from a sealed card run through a coupling (from), takes each handed value in parameters bit for bit: decode it from the run's frame_hex or a _hex value, not a decimal readout, or it is refused as CouplingLedgerBreach. verify_solution verifies a solution later from this request, or its inputs, and its solution_seal. On this open connection one address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Solve
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
validate_seriesRuns a run_trajectory request (without a readout) and compares one published quantity with the agent's measurements, each at a landed step.1 finding
Why it is in this group
- The name "validate_series" says "validate": a word for reading.
- The description's first verb is "run": a word for a change.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Findings
- Warn
The server marks "validate_series" read-only (readOnlyHint: true), and its own name, description or input give the class "write".
What this check looks forMatched text
annotations: {"readOnlyHint":true,"destructiveHint":false,"idempotentHint":true,"openWorldHint":false}; The name "validate_series" says "validate": a word for reading.
Description, as the server wrote it Runs a run_trajectory request (without a readout) and compares one published quantity with the agent's measurements, each at a landed step. The answer is the error in numbers (rms, mean signed, greatest absolute, model minus measurement, each with its exact hex), the model value at every compared step in decimals (model_values) and exact (model_values_hex), and a hash of the measurements, under one seal. It sets no threshold: the numbers are the answer. The seal has no key, so it does not show who measured or who computed. commitment, optional, opens a salted SHA-384 the measurements were locked with before the comparison: data that are not the committed ones are refused before computing (CommitmentDoesNotMatch), and the seal covers it. genesis_seal verify-validation rebuilds the seal, the statistics and the commitment from the answer alone. In place of run, solves takes a plan of solves as run_dataset takes it (solver and plan, at most 64 solves), each measurement at a row of the plan: the solves are sealed under one dataset_seal, and the comparison by the same rules with the dataset_seal in the place of the run seal and the row in the place of the step; a measured row whose solve did not converge is counted in not_reached. The same request sent again gives the same seal. On this open connection a plan of solves holds the engine's own 64 and at most 15 s of them, or one, at each solver's measured time, which describe_solver states with the count (OPEN_PLAN_SECONDS). One address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Validate series
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
verify_datasetVerifies the seal of a dataset.No findings
Why it is in this group
- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Description, as the server wrote it Verifies the seal of a dataset. Short form: request is the run_dataset request as it was sent, and dataset_seal is the seal its answer gave. The plan is run again under run_dataset's permits and limits; the answer is holds, cause ("dataset_seal" when the seal given is not the seal the plan runs to), recomputed_dataset_seal when it holds (else no seal: run_dataset gives the true one), compared, and text_not_compared: no rows the agent holds are compared. A plan of solves is verified by this short form only. Long form: head, rows and frames_hex, the three parts a full answer carries (answer "full"), together, as they came. The rows, the root and the seal are recomputed, and row and fault name the first row changed, swapped, dropped or repeated; then the plan is run again under run_dataset's permits and limits and the engine's answer compared with it, value by value and byte for byte (cause answer_text, with differs_at and row): holds is true only when every byte is the engine's, values_hold when every value is. On this open connection a dataset is returned inline and holds at most 300 runs (OPEN_MAX_DATASET_ROWS), 1000000 steps in all (OPEN_MAX_DATASET_STEPS), 10000 frames and an answer counted from the plan at 4194304 bytes (OPEN_MAX_DATASET_ANSWER_BYTES) and 42500 characters, up to about 150 summary rows; a plan of solves holds the engine's own 64 and at most 15 s of them, or one, at each solver's measured time, which describe_solver states with the count (OPEN_PLAN_SECONDS). One address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Verify dataset
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
verify_receiptVerifies the receipt at the end of one response against the text before it.No findings
Why it is in this group
- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Description, as the server wrote it Verifies the receipt at the end of one response against the text before it. It takes no physics permit. The receipt covers every byte of the response before it, the decimal values included, and shows who produced the response or when in no way: it has no key. matches true proves only that those bytes are unchanged since the receipt was written; anyone can write a matching receipt for any text, so it is not a verification of the answer. A response that names an engine_version has it said beside matches, with whether it is this engine's; origin needs engine_version_is_this_engines true and a replay. match_proves says this in each answer. The receipt covers the request as it was written, number spellings included, while a seal covers its values: a request sent with 310 and the same request sent with 310.0 give different receipts and the same seal, so a receipt that differs between two such answers is not a sign of tampering. Replaying the request through run_trajectory, solve or compute_step shows the engine produced it; verify_tape, verify_solution and verify_dataset with the request and its seal do that replay and compare the seals.
Annotations the server declares
- title: Verify receipt
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
verify_solutionVerifies the seal of a solution.No findings
Why it is in this group
- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Description, as the server wrote it Verifies the seal of a solution. Short form, which needs only what is short: request is the solve request as it was sent (the answer's inputs), and solution_seal is the seal its answer gave. The request is solved again, under a physics permit and the limits and refusals of solve, and the answer is holds, cause ("solution_seal" when the seal given is not the seal the request solves to), recomputed_solution_seal and recomputed_receipt_sha384 when it holds (else no seal: solve gives the true one), compared, and text_not_compared: no answer text the agent holds is compared. Long form: the whole solve answer sent back as it came, without request. Its solution seal is recomputed from its exact hex values and the part that fails is named, then its request is solved again under a physics permit and the engine's answer compared with it, value by value and byte for byte, receipt included: holds is true only when every byte is the engine's, values_hold when every value is, else cause names the part, solution_seal (the request solves to another seal), answer_text (with differs_at, the first field that differs) or engine_version (an answer of another version). A chained solve: its whole answer with from beside it; the upstream run is computed again and the chain held too (cause chain), with recomputed_chain_seal. Alone it is refused (AnswerCarriesNoRequest). Nothing is stored, so each form answers the same on every connection. On this open connection one address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Verify solution
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
verify_stepComputes one anchored step of one card, as compute_step does: verify_step is the earlier name of compute_step, served for this one version beside it, and takes the same arguments a…No findings
Why it is in this group
- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Description, as the server wrote it Computes one anchored step of one card, as compute_step does: verify_step is the earlier name of compute_step, served for this one version beside it, and takes the same arguments and gives the same answer, byte for byte. The name compute_step is the one that stays. On this open connection one address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Verify step
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
verify_tapeVerifies the seal of a run.No findings
Why it is in this group
- Nothing in its name or description says what it does, so it is counted as a write.
- The server marks it read-only (readOnlyHint: true), but an annotation never lowers a class.
Description, as the server wrote it Verifies the seal of a run. Short form, which needs only what is short: request is the run_trajectory request as it was sent, with its readout and its from, whose request has none, when it had them (the answer's inputs holds neither), and run_seal is the seal its answer gave. The run is computed again, under a physics permit and the limits and refusals of run_trajectory, and the answer is holds (the seal given is the seal the request runs to), cause ("run_seal" when it is not), recomputed_run_seal and recomputed_receipt_sha384 when it holds (else no seal: run_trajectory gives the true one), compared, and text_not_compared: no answer text the agent holds is compared. Long form: the whole run_trajectory answer sent back as it came, without request. Its state digest and run seal are recomputed from its exact hex values, then its request is computed again under a physics permit and the engine's answer compared with it, value by value and byte for byte, receipt included: holds is true only when every byte is the engine's, values_hold when every value is, else cause names the seal, load_signature (a signed load that does not verify), answer_text (with differs_at, the first field that differs) or engine_version (an answer of another version). A chain is verified by its request with from, run_seal and chain_seal, or by links, its answers in order, each computed again from the links before it; a chained answer alone is refused (AnswerCarriesNoRequest). Nothing is stored, so each form answers the same on every connection. On this open connection one address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Verify tape
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
Only reads 6Show 6 read-only tools
describe_cardDescribes one domain and card pair: inputs, units, bounds, step limits, reject names and meanings, advisory flags, frame slots, and worked examples.No findings
Why it is in this group
- The name "describe_card" says "describe": a word for reading.
Description, as the server wrote it Describes one domain and card pair: inputs, units, bounds, step limits, reject names and meanings, advisory flags, frame slots, and worked examples. With form "short" the answer is a page, enough to write a first request: what the model is of, the evidence behind it in numbers, each input with its unit and range, the names of its stops, and a worked request. The full form, the answer when form is left out, adds each bound's reason, the validation record and each stop's meaning.
Annotations the server declares
- title: Describe card
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
describe_solverDescribes one solver: inputs and bounds, outputs, invariants and tolerances, refusal names and meanings, the published reference, and a worked example.No findings
Why it is in this group
- The name "describe_solver" says "describe": a word for reading.
Description, as the server wrote it Describes one solver: inputs and bounds, outputs, invariants and tolerances, refusal names and meanings, the published reference, and a worked example. With form "short" the answer is a page, enough to write a first request: what the model is of, the evidence behind it in numbers, each input with its unit and range, the names of its stops, and a worked request. The full form, the answer when form is left out, adds each bound's reason, the validation record and each stop's meaning.
Annotations the server declares
- title: Describe solver
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
find_flipFinds where an outcome changes, and gives the run on each side.No findings
Why it is in this group
- The name "find_flip" says "find": a word for reading.
Description, as the server wrote it Finds where an outcome changes, and gives the run on each side. request is a run_trajectory or solve request written out, its domain and card or its solver by the name find_model or the lists give, with no from, readout or example; vary is the place of the one number to change, such as "segments[0].load.current_a", "initial.soc", "dt_s" or "parameters.load_n"; low and high are the two ends of the span, in either order: the smaller is taken as low. The span is halved to the place where the outcome or the stop name changes or, with level, where a published quantity of the final state (of a solve: of result) crosses a value or a flag turns. When both ends give the same, 7 values between them (3 for solves) are run to find a band, and the answer says so. The answer gives the run on each side as a request ready to send again, with its outcome, its stop name and sentence and the run_seal or solution_seal that request has when sent alone, the two values in decimals and exact hex, and the width between them. A side that stops before computing has no proof. The same request gives the same pair. A search is bounded, and names the bound that stops it: 64 runs (FLIP_MAX_RUNS), 1000000 steps in all (FLIP_MAX_STEPS), or 64 solves (FLIP_MAX_SOLVES), the ends included, and no new solve once those made are counted at 20 s (FLIP_MAX_SECONDS); the pair it gives, sent back as the span, narrows it further. On this open connection a search is long work whenever it varies a solve or its runs may take more than 10000 steps together. One address holds at most 1 physics permit at a time, and long work (a solve, a plan of runs or of solves, a run of more than 10000 steps, or a verification that computes one again) runs 1 at a time. A call that finds no permit or no place is answered "Too many concurrent physics tasks." and nothing waits: the same call sent later is answered once one is free. An answer is waited for 60 seconds, and long work is stopped there.
Annotations the server declares
- title: Find flip
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
find_modelFinds the served model for a real thing described in plain words, such as "a lithium cell discharging in the cold".1 finding
Why it is in this group
- The name "find_model" says "find": a word for reading.
Findings
- Warn
The description contains "every other tool", wording about other tools or servers.
What this check looks forMatched text
description: "...escribe_card, describe_solver, cite and every other tool take; describe_card or describe_solver..."
Description, as the server wrote it Finds the served model for a real thing described in plain words, such as "a lithium cell discharging in the cold". It returns the best few cards and solvers, each with fit first (the words that matched, and model_of: the one real thing it is a model of, and its range) and then accuracy: against_measurement, the evidence behind it in one of three words, as list_cards and list_solvers give it ("compared with measurement", "held to a reference calculation" or "company specimen"); compared_with, what it was fitted on; and accuracy_against_measurement, what it was compared with outside the fit, within what, and the gap, in numbers. Fit and accuracy are separate: a model with a gap is still the served model of its thing. The names it gives (domain and card, or solver) are the names list_cards and list_solvers give and the names describe_card, describe_solver, cite and every other tool take; describe_card or describe_solver gives the range and every refusal. When nothing served fits it says so, having compared every card and solver, names the nearest by words and why they are not it, or says that there is no near one, and says where requests for a model go. A model is set aside when the description names another thing of its kind (another gas, another maker's cell, another planet's probe). It is matched against a fixed list of terms with no model call: the same words give the same answer.
Annotations the server declares
- title: Find model
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
list_cardsLists every domain and card this connection serves, each with one sentence, and count, how many there are.No findings
Why it is in this group
- The name "list_cards" says "list": a word for reading.
Description, as the server wrote it Lists every domain and card this connection serves, each with one sentence, and count, how many there are. against_measurement names the evidence behind the model, in one of three words: "compared with measurement" (a real object; its description says what it was fitted on, what it was compared with outside the fit and within what, the gap, and where it refuses), "held to a reference calculation" (compared with a published calculation), or "company specimen" (an object the company specified, every constant stated). The numbers say how good it is.
Annotations the server declares
- title: List cards
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
list_solversLists every field solver this connection serves, each with one sentence, and count, how many there are.No findings
Why it is in this group
- The name "list_solvers" says "list": a word for reading.
Description, as the server wrote it Lists every field solver this connection serves, each with one sentence, and count, how many there are. against_measurement names the evidence behind the model, in one of three words: "compared with measurement" (a real object; its description says what it was fitted on, what it was compared with outside the fit and within what, the gap, and where it refuses), "held to a reference calculation" (compared with a published calculation), or "company specimen" (an object the company specified, every constant stated). The numbers say how good it is.
Annotations the server declares
- title: List solvers
- readOnlyHint: true
- destructiveHint: false
- idempotentHint: true
- openWorldHint: false
This is not a safety rating, and no finding is not a safety guarantee. The groups and findings come from fixed checks of the tool definitions the server returned on the date shown, not from its code or what it does when called. How the checks work.
Exfil surface
Which tools read private data, which read text someone else can write, and which can send data out. One of each is a path a prompt injection can use.
No private-data, untrusted-input, outbound path in this server alone: no tool is tagged private-data or untrusted-input or outbound.
- Reads private data0
- Reads files, mail, messages, databases, secrets, calendars or repositories.
- Reads text others write0
- Brings in text someone else can write: web pages, inbound email, messages, issues and comments.
- Can send data out0
- Can carry data to people or systems: it sends, posts or publishes, or takes a recipient, channel or URL.
Other servers in the same agent add their own tools to these paths. Audit your whole setup to see them together.
Remote posture
What anyone on the internet sees at the remote endpoint: an anonymous initialize and tools/list, and the public OAuth metadata.
Looked at Nov 20, 2026, 18:50 UTC, unchanged since Oct 22, 2026. No credentials were sent and no tool was called. What is looked at.
- Reachable
- yes
- HTTP status of an anonymous initialize
- 200
- tools/list served without credentials
- yes
- Tools that change things, listed without credentials
- 12
- Protocol version
- 2025-06-18
- WWW-Authenticate on 401
- not observed
- resource_metadata in WWW-Authenticate
- not observed
- RFC 9728 protected resource metadata
- not observed
- Authorization servers
- not observed
- Legacy HTTP+SSE transport
- no
- Dynamic client registration
- not observed
- Client ID metadata documents
- not observed
- WarnThe whole server
The endpoint served tools/list without credentials, and the list has 12 tools that change things (12 write). Whether a call also works without credentials was not tested: tools are never called.
What this check looks forMatched text
https://mcp.zerotrustphysics.com/mcp: tools/list returned 18 tools to an anonymous client, including compute_step (write), run_trajectory (write), verify_tape (write), solve (write), verify_solution (write) and 7 more
Findings
Matches of fixed checks, with the text they matched.
- Warnrun_trajectory
The server marks "run_trajectory" read-only (readOnlyHint: true), and its own name, description or input give the class "write".
What this check looks forMatched text
annotations: {"readOnlyHint":true,"destructiveHint":false,"idempotentHint":true,"openWorldHint":false}; The name "run_trajectory" says "run": a word for a change.
- Warnrun_dataset
The server marks "run_dataset" read-only (readOnlyHint: true), and its own name, description or input give the class "write".
What this check looks forMatched text
annotations: {"readOnlyHint":true,"destructiveHint":false,"idempotentHint":true,"openWorldHint":false}; The name "run_dataset" says "run": a word for a change.
- Warnvalidate_series
The server marks "validate_series" read-only (readOnlyHint: true), and its own name, description or input give the class "write".
What this check looks forMatched text
annotations: {"readOnlyHint":true,"destructiveHint":false,"idempotentHint":true,"openWorldHint":false}; The name "validate_series" says "validate": a word for reading.
- Warnfind_model
The description contains "every other tool", wording about other tools or servers.
What this check looks forMatched text
description: "...escribe_card, describe_solver, cite and every other tool take; describe_card or describe_solver..."
- Warndesign_run
The server marks "design_run" read-only (readOnlyHint: true), and its own name, description or input give the class "write".
What this check looks forMatched text
annotations: {"readOnlyHint":true,"destructiveHint":false,"idempotentHint":true,"openWorldHint":false}; The name "design_run" says "run": a word for a change.
History
Changes to its tool definitions and to its remote endpoint, newest first.
Tools1 changed.~1 changedView diff
- Changed:solveChanges thingsdescriptioninputSchema
Tools5 changed.~5 changedView diff
- Changed:citeChanges thingsdescription
- Changed:verify_datasetChanges thingsinputSchema
- Changed:verify_receiptChanges thingsinputSchema
- Changed:verify_solutionChanges thingsinputSchema
- Changed:verify_tapeChanges thingsinputSchema
Install
Pinned to version 0.15.3. Values in angle brackets are yours to fill in.
claude mcp add --transport http verifier https://mcp.zerotrustphysics.com/mcpclaude mcp add --transport http verifier https://mcp.zerotrustphysics.com/mcpRun in a terminal.
Claude Code permissions
Allow the reads, ask before a change or a way out, deny the deletes. Paste into .claude/settings.json.
{
"permissions": {
"allow": [
"mcp__verifier__list_cards",
"mcp__verifier__describe_card",
"mcp__verifier__list_solvers",
"mcp__verifier__describe_solver",
"mcp__verifier__find_model",
"mcp__verifier__find_flip"
],
"ask": [
"mcp__verifier__compute_step",
"mcp__verifier__run_trajectory",
"mcp__verifier__verify_tape",
"mcp__verifier__solve",
"mcp__verifier__verify_solution",
"mcp__verifier__run_dataset",
"mcp__verifier__verify_dataset",
"mcp__verifier__verify_receipt",
"mcp__verifier__validate_series",
"mcp__verifier__cite",
"mcp__verifier__design_run",
"mcp__verifier__verify_step"
],
"deny": []
}
}Tools that only read are allowed, tools that change things ask first, tools that delete or cancel are denied. A read that can send data out (it takes a URL, say) asks first too. It assumes the server is installed as verifier. Review it before you use it.
Pin
Keep the decision: fail CI when the tools change, follow every change, or show a badge.
Pin in CI
Fail a build when the tools you reviewed are not the tools the server serves today.
curl -s "https://radar.anlyon.com/api/v1/servers/com.zerotrustphysics/verifier/verify?hash=c97089ea72eedc1f41cfbc1881be3c72cd3e26d4b037be82cf0f2d49bda95534" | jq -e '.changed == false'Pinned hash c97089ea72ee, from Nov 16, 2026, 03:17 UTC.
Watch
Get told when its tools change: a new tool, a new description, a new class.
https://radar.anlyon.com/servers/com.zerotrustphysics/verifier/changes.xmlAn Atom feed. Works in any feed reader or Slack's RSS app.
Badge
[](https://radar.anlyon.com/servers/com.zerotrustphysics/verifier)Source: MCP Registry. Synced Nov 2, 2026, 19:17 UTC.
Zero-Trust Physics MCP server: questions and answers
Answered from the scan above, not from the publisher's description.
Is the Zero-Trust Physics MCP server safe?
- No scan can promise that. What its definitions show: 18 tools, 12 of which can change, send or delete something, and 5 findings from fixed checks. Allow the reads, ask before each change, and deny the deletes; the permissions block on this page does that.
Can the Zero-Trust Physics MCP server delete data?
- No tool is classed as destructive: going by their definitions, none of them deletes, cancels or revokes anything.
Can the Zero-Trust Physics MCP server do things that cannot be undone?
- No tool is classed as irreversible: none of them sends, pays, publishes or merges, going by their definitions.
What tools does the Zero-Trust Physics MCP server have?
- 18 tools: 12 write, 6 read. compute_step, run_trajectory, verify_tape, solve, verify_solution, run_dataset, verify_dataset, verify_receipt and 10 more.
Does the Zero-Trust Physics MCP server need sign-in or OAuth?
- Its remote endpoint listed its tools to an anonymous client, without sign-in.
Does the Zero-Trust Physics MCP server send data to outside services?
- None of its tools sends, posts or publishes, or takes a recipient, channel or URL.
Has the Zero-Trust Physics MCP server changed its tools?
- Yes, 2 times since Anlyon Radar started tracking it, most recently on Oct 22, 2026. Every change is on this page and in its Atom feed.