casino-mcp

An MCP control plane over the Fortran CASINO quantum Monte Carlo code: write the input for the next calculation — and the blank Jastrow factor, backflow function and geminal wave function for the first one — start it, know what is running, stop it, and read the result as structured data instead of shipping 4000 lines of text into a model's context, including from a DMC run that is still going, which has no energy in out at all until its last block.

Beta (0.6.0). Everything below is tested against a real CASINO: the recipes against runqmc's own input check, and every file the wave function writer produces against a CASINO test run. Interfaces may still move before 1.0.

What it is, and what it is not

CASINO already has the primitives — opt_plan, runqmc --auto-continue, multirun, envmc, make_E_v_dt. What it has no place for is the layer between them: machine-readable results, a memory of what was run, and the judgement calls that sit between the steps ("has the variance plateaued?", "is this timestep still in the linear regime?"). That layer is what this package is, and three rules keep it honest:

1. No number is produced by the model. Every value a tool returns is read from a file and carries the line it came from. What CASINO did not print comes back as null with a reason, never a guess.

2. Every result is reproducible from its record. A job record freezes the command, the process count, and the path, size and mtime of the casino binary that ran.

3. Nothing destructive is implicit. A run refuses to start in a directory that already holds results, and refuses harder when that out is committed reference data.

There is deliberately no execute_shell(command) tool. Every tool is a named CASINO operation with typed arguments.

Related MCP server: AutoDL MCP Server

Install

pip install casino-mcp          # needs Python 3.11+ and a working CASINO installation

From a checkout:

pip install -e '.[dev]'

The package does not bundle, build or replace CASINO. It drives runqmc, which stays the runtime: arch detection, MPI variants, batch-queue submission and the lock file are its job, not ours.

Register it with Claude Code

.mcp.json, project scope:

{
  "mcpServers": {
    "casino": {
      "command": "casino-mcp",
      "args": ["serve"],
      "env": {
        "CASINO_HOME": "/home/you/bin/CASINO",
        "CASINO_ARCH": "linuxpc-gcc-parallel.openblas"
      }
    }
  }
}

Tools

casino_results and casino_input answer different questions on purpose: what a run did against what it was told to do. The keywords in a result are CASINO's own echo, which is neither the file nor a superset of it — it holds every default CASINO applied, 70 entries against the 23 a file typically sets, and drops what it does not print, random_seed among them. casino_input also reads a directory nothing has run in, which is how a prepared calculation is checked before there is a job to name it by, and keeps the input a stopped job was started from, since haltqmc -u rewrites that file in place.

fields is what a scan wants from casino_results: a whole parsed run is 10–16 kB of JSON, and 38 directories' worth of it is 600 kB to say six numbers a point. A path is written in the run's own keys — vmc, opt, dmc_equil, dmc_stats mean the last phase of that kind, opt[3] the cycle CASINO itself numbered, phases[-1] a position, and anything else is a key: keywords.DTVMC, cpu_time, vmc.energy.error. A path that does not exist comes back in problems naming what is there instead; a path that exists but holds a number CASINO never printed comes back as null with its reason. keywords.DTVMC against vmc.dtvmc is the pair worth asking for together — the step the input asked for against the step the run used.

Every job_id above is either the id casino_run returned or the calculation directory, which is what a chain of runs actually holds: the registry knows which job ran where, so nothing has to be written into the calculation directory to record it.

The runtype (vmc, vmc_opt, vmc_dmc, …) comes from the input file in workdir; there is no tool per runtype, because that multiplies the surface without adding a capability. What casino_prepare adds is the other half of that: it writes the input, filling in the keywords a runtype requires and the source directory does not set, and refusing to write one that CASINO would reject.

Starting a chain: the blank wave function

The first calculation of a chain comes out of an orbital code with a wave function file and nothing else, and use_jastrow : T needs a correlation.data that does not exist yet. No CASINO utility writes one — the manual's own instruction is to copy an example and delete its numbers by hand — so casino_prepare(..., jastrow=['u', 'chi', 'f'], backflow=['eta', 'mu', 'phi']) writes it, both blocks in the one file:

• the atoms come from the orbital file's own header, because input says how many electrons there are and never how many nuclei; one set per element, every atom labelled;

• which atoms are pseudo-atoms comes from the *_pp.data files, each of which states its own atomic number. In the Jastrow that decides where the chi cusp is refused, because CASINO errstops on it; in the backflow it decides the cusp type of every mu and phi set, which CASINO believes without checking — 1 for a bare nucleus, 0 behind a pseudopotential;

• every coefficient starts at zero, which is what the first optimisation cycle is for;

• the cutoffs are written as zero, which CASINO reads as use your own default: 2 or 5 a.u. for u, 4 for chi, 3 for f, 4.5 for mu and phi, and 1 or 4 for eta depending on whether the channel carries the e-e cusp. warnings says which values that will be. No AE CUTOFFS block is written either — it is optional, and CASINO picks those lengths itself.

jastrow_settings overrides any of the shape, for both blocks: trunc_order, bf_trunc_order, n_u, n_chi, n_f_en, n_f_ee, n_eta, n_mu, n_phi_en, n_phi_ee, every spin_dep_*, cusp_chi, irrotational, every cutoff_*, and cusp_bf for the rare all-electron orbital set that does not satisfy the cusp condition. Finite systems so far: a periodic Jastrow wants a P term, whose stars of reciprocal lattice vectors come from CASINO's own make_p_stars.

A block is written only if the input turns its keyword on, and a keyword that is on with no block is refused rather than left for CASINO to errstop over — the two halves of the same mistake.

The geminal wave function

psi_s : geminal replaces the Slater determinant with a sum of geminal determinants — the electrons are paired by Φ(r,r′) = Σ g_mk φ_m(r) φ_k(r′) instead of put in orbitals — and every parameter of it lives in the GEMINAL block of a parameters.casl. CASINO ships no utility that writes one either, so casino_prepare(..., geminal=[...]) does:

casino_prepare('./hf', './gem', geminal=[])  # the Hartree-Fock geminal alone
casino_prepare(
    './hf',
    './gem',
    geminal=['p:2', 'd:1'],  # ... and a correlating one over
    geminal_settings={'anchors': [1]},
)  #     the first two p and first d levels

• geminal=[] is the Hartree-Fock determinant, exactly. g_m,m = 1 over the doubly occupied orbitals and one u_m,k per singly occupied one; the manual recommends it as the check to make before correlating anything, and tests/integration/test_geminal_casl.py makes it — a VMC run over it has to land on the energy the same system gives with psi_s : slater. Being channel-less it reads no orbital file and works for any basis.

• A channel is a degenerate level, not an orbital. p:2 means the first two p levels of the orbital file, and the whole of each is tied together in Constraints, component by component — a correlating geminal built out of one component of a level is not spherically symmetric, and optimizing it breaks the symmetry of the state it describes. The levels are read off the orbital coefficients of gwfn.data, so a channel needs a gaussian basis.

• A level whose orbitals are not one clean m-component each is demoted, not guessed at. It gets a diagonal-only tie and a line in warnings saying so, because component-wise off-diagonal ties between two levels that are mixed differently constrain orbitals that are not each other's counterparts.

• The unpaired columns are not optional. An open shell needs one u_m,k in every geminal with a non-zero c, since an empty unpaired column makes the geminal matrix singular at every configuration — CASINO's check_umat errstops on it — and they are written fixed, because parse_umat_el refuses an optimizable one.

geminal_settings holds the rest: seed and seed2 (−0.05 and −0.02, the two leading correlating diagonals, which start away from zero because a geminal holding only its anchors is singular and has no gradient to move it), anchors (derived — every occupied orbital no correlated level holds — unless given), mirror (a third geminal with c = -1, tied to the second parameter for parameter), and purity.

Reading a DMC run before it ends

A DMC calculation runs for hours and has no energy in out until the last block: CASINO writes the mixed estimators once, at the end. Until then the current estimate lives in dmc.status, which it rewrites after every statistics block and deletes when the run finishes — copying the same text into out at that moment, so nothing is lost, but nothing is available either while it matters most.

casino_results reads that file when it is there, and points result at it. So a running job answers with the estimate as of its last block, and never with the VMC energy of the configuration-generation phase — which is the trial wave function's, not the calculation's. A run stopped by casino_stop keeps its dmc.status, so the last estimate it reached survives the stop; a run still equilibrating has none, and result says so rather than reaching for an earlier phase.

Starting, stopping and continuing a calculation all go through CASINO's own scripts, and only through them: runqmc starts, haltqmc ends and tidies, runqmc --continue or a plain runqmc over the input that haltqmc -u rewrote carries on. Nothing here moves a config file, edits an input, or decides what a half-finished calculation should do next.

A directory that already holds an out

runqmc appends to out, vmc.hist and dmc.hist rather than replacing them, so running twice in one directory produces files that are two runs glued together. That is refused by default, and there are two ways past it — opposites, so pass one:

On the command line these are --restart and --resume, and --continue is accepted for the latter, which is what runqmc calls it. The tool parameter cannot be spelled that way: continue is a Python keyword.

One machine, one run at a time

A second run is refused while a job this server started is still going. Not because the machine is busy — because of what sharing it does to the numbers. Two CASINO runs land on the cores the scheduler gives them, and Total CASINO CPU time counts seconds of CPU across the MPI processes: two jobs that ended up on the same core measured 97.2 s of CPU against 194.41 s of real time, and efficiency, which is computed from the CPU time, was wrong by the same factor. Nothing in the output says this happened. The ratio of cpu_time to real_time does, which is one casino_results(fields=['cpu_time', 'real_time']).

Only jobs this server started are known here. A pgrep casino over the machine is deliberately not done: someone else's process is someone else's business, and "something is computing" with no owner and no job id is a refusal the caller has nothing to answer with. So the reading afterwards is not a backstop but the check that actually holds — a run started around the server spoils the timings just the same, and the registry never sees it.

Rerunning a directory without losing the last answer

restart deletes out, and with it what the previous run found. keep_previous moves it to the first free out.1, out.2, … instead; the rest of what restart deletes goes as before, and the reply says where the file went, under kept. The numbering ends up being the order the runs happened in.

The archives are deliberately not in the list of things restart deletes, and they are not called out_part.N — that name is runqmc's own, for the finished segments of a --continue, and it is in that list.

The half that matters is in the registry rather than in the directory. A job record holds a directory, not a file, so before this a casino_results for the earlier job read whatever out was in the directory now — and answered under the earlier job's id, with its status and its start time, and nothing in the reply saying the physics belonged to the run after it. The record follows the file, so each job goes on answering with its own numbers:

first = casino_run(workdir='calc')['job_id']  # writes calc/out
second = casino_run(workdir='calc', restart=True, keep_previous=True)  # out -> out.1
casino_results(first)['cpu_time']  # from calc/out.1
casino_results(second)['cpu_time']  # from calc/out
casino_results('calc')['cpu_time']  # a directory still means its newest run

The alternative is a fresh directory through casino_prepare, which copies the wave function too — for a blip calculation, a bwfn.data of hundreds of megabytes to keep a text file of ten kilobytes.

Stopping a run, and continuing it

casino_stop sends SIGTERM to that job's casino processes and to nothing else — the same signal haltqmc -k sends, except that haltqmc's is a pkill -x casino over the whole account, which would take down every other job on the machine. mpirun puts each rank in a process group of its own, so the ranks are found by session id: the session is the launcher's, and the whole tree shares it. runqmc itself is left alive to finish its epilogue, and only a job still running after timeout has its process group signalled and then killed.

Then the directory goes to haltqmc -f -u, which is the part that makes a stopped run continuable: config.out becomes config.in, the lock and marker files go, and input is rewritten for the work that is left — newrun : F, the finished blocks subtracted, the runtype moved on to the next stage. The reply carries what it did under halt. The input as it was before that is copied into the job directory, and halt.input_saved says where.

Which continuation route resume=true then takes is decided by the last run in out:

restart=true is refused on a directory whose input says newrun : F, because restarting deletes the config.in that CASINO then demands. Put back the saved input first.

Command line

The same runtime without a model in the loop — which is also how you debug the server:

casino-mcp config                  # the resolved configuration, and the files it came from
casino-mcp run ./calc -p 4         # start a calculation
casino-mcp run ./calc --restart    # ... after deleting what an earlier run left there
casino-mcp run ./calc --continue   # ... or carrying that run on instead
casino-mcp status 20260823-164511-qobn
casino-mcp status ./calc           # every job argument takes a directory: its newest job
casino-mcp wait   ./calc           # block until that calculation ends
casino-mcp stop   20260823-164511-qobn   # stop the run, then hand the directory to haltqmc
casino-mcp jobs                    # the registry, newest first
casino-mcp jobs -C ./calc          # ... or only what ran in one directory
casino-mcp results 20260823-164511-qobn   # the physics of that job, live runs included
casino-mcp prepare ./vmc ./dmc --runtype vmc_dmc -s dtdmc=0.005   # the next calculation
casino-mcp prepare ./hf ./opt --runtype vmc_opt --jastrow u,chi,f # ... and the first one
casino-mcp prepare ./hf ./bf --jastrow --backflow -s backflow=T   # ... with backflow in it
casino-mcp prepare ./hf ./gem --geminal p:2,d:1 -g anchors=1      # ... as a geminal wave function
casino-mcp input ./calc            # the keywords it was given, random_seed included
casino-mcp parse ./calc            # the `out` file as JSON
casino-mcp serve                   # the MCP server on stdio

Every subcommand prints JSON and exits non-zero when that JSON carries an error.

Configuration

There is no configuration file. An MCP server is configured where it is registered — the env block of the .mcp.json above — and CASINO's own variables keep their names, so setting them once configures both layers:

Everything else — one MPI process, the opt binary, twenty seconds for a stopped job to end on its own, a minute for haltqmc to tidy, two hundred job records kept — is a constant in settings.py. casino-mcp config prints what the server would use right now and which variable said so; run it first when a tool call refuses.

CASINO_MCP_FORBID is the one guard with no per-call override. restart=true/resume=true and unlock=true unlock the other two; a directory listed here cannot be run in at all, which is what makes it the right place for a tree of committed reference calculations.

How it works

Claude Code ──stdio──> server.py ──spawn──> launcher.py ──> runqmc ──> mpirun ──> casino
                          │                     │
                          │                     └─ writes status.json (exit code, end time)
                          └─ reads/writes jobs.json + one directory per job

State lives outside the calculation, under $XDG_STATE_HOME/casino-mcp/:

jobs.json                    index: job_id -> record
jobs/<job_id>/meta.json      what was launched, frozen at spawn
jobs/<job_id>/status.json    written by the launcher when the run ends
jobs/<job_id>/runqmc.log     runqmc's own output (not CASINO's `out`)
jobs/<job_id>/input.before_halt   the input as it was, kept when a stop rewrites it

The calculation directory only ever gets what CASINO puts there.

Why a launcher process. runqmc is a bash script that execs mpirun -np N casino; signalling its pid orphans the tree. The launcher runs in its own session, which is what makes the tree identifiable — killpg reaches runqmc and mpirun, and the session id finds the ranks that mpirun put in process groups of their own — its exit code survives the MCP server being restarted, and runqmc's output goes to a log instead of the JSON-RPC stream. A recycled pid cannot pass for a live job: /proc/<pid> start time is compared, and a zombie does not count as running.

The out parser

parse_out is a plain function with no MCP and no dependencies. An out file is a sequence of phases, not one result — vmc_opt writes a VMC and an OPTIMIZATION phase per cycle, vmc_dmc writes VMC, DMC equilibration and DMC statistics accumulation — so it returns phases, and result points at the last phase that carries an energy.

from casino_mcp.parse_out import parse_out

parsed = parse_out('./calc')
parsed['result']['energy']  # {'value': -2.861829862553, 'error': 0.000659077167, 'line': 237}

The one derived number in it is the sample-variance error of a single-block run, which CASINO does not print; it is taken from the one block exactly as envmc does, and labelled derived. Nothing shells out to envmc or endmc at runtime — endmc misparses numbers under a non-C locale.

parse_dmc_status reads the dmc.status of a run that has not finished, through the same parser: write_dmc_status in CASINO's dmc.f90 writes that file and the out section from one place, so reading them with two would be one more thing to keep in step. parse_out picks it up on its own when the file is next to the out it was given.

The input writer

input_file is the same shape in the other direction: text in, text out, no MCP.

from casino_mcp import input_file

current = input_file.read('./vmc')
filled, missing = input_file.recipe('vmc_dmc', {'dtdmc': '0.02083'}, present=current['keywords'])
text = input_file.apply(current['text'], filled)  # edits; it does not regenerate
input_file.check(*input_file.parse_text(text))  # [] when CASINO would take it

apply only touches the lines it is named for, so hand comments, %blocks and expert keywords no recipe has heard of all survive a rewrite — a calculation's input is a document, and the parts nobody can reconstruct are exactly the parts a template would drop. build writes a whole file from a recipe for callers that have no source to start from.

The recipes and the rules come from runqmc's own checks rather than from reading the manual, and tests/integration/test_recipes_check_only.py puts every one of them back to runqmc --check-only: a recipe is right when CASINO says the input is runnable, not when our own check does.

The correlation.data writer

correlation_data is the same shape again, and the layer under the jastrow and backflow arguments above:

from casino_mcp import correlation_data

geometry = correlation_data.read_geometry('./hf/gwfn.data')  # atoms, not orbitals
pseudo = correlation_data.pseudo_species('./hf')  # {8}, out of o_pp.data
problems = correlation_data.check(geometry, terms=('u', 'chi', 'f'), backflow=('eta', 'mu', 'phi'))
text = correlation_data.blank(geometry, backflow=('eta', 'mu', 'phi'), pseudo=pseudo)

Every label and every line of it is CASINO's own: the unit suite strips the numbers out of two committed files — an optimised Jastrow and a hand-written blank backflow — and asserts that what is left is exactly what this writes for the same atoms. runqmc --check-only is no oracle here, it never opens the file, so tests/integration/test_blank_correlation.py uses testrun : T, which makes CASINO read the input files, impose the cusp, no-duplication and no-cusp constraints, count what is left free, check that they hold, and stop. That is also how the one rule nobody could read off the source was found: an all-electron phi set with N_eN = 1 has no free parameters left, whatever N_ee is, while a pseudo-atom set at the same order is fine.

The parameters.casl writer

geminal is the third writer of the same shape, and the layer under the geminal argument above:

from casino_mcp import geminal

orbitals = geminal.read_orbitals('./hf/gwfn.data')  # orbitals, not atoms
levels = geminal.mo_levels(orbitals)  # {1: [([3, 5, 4], True), ...], ...}
shells, diagonal, problems, notes = geminal.select(levels, [(1, 2)])  # the first two p levels
text = geminal.geminal_section([1, 2], [], [1], shells, diagonal)

CASL is not YAML — a constraint line reads 2^g_5,5=2^g_4,4, which is a bare scalar no YAML parser accepts — so the block is generated as plain text. Each MO is classified by the (l, m-slot) its coefficients live on, after the solid-harmonic constants CASINO premultiplies into d coefficients (and, per molden2qmc.py, not into f and g ones) are divided back out; MOs of the same l are grouped into levels of 2l+1 in file order.

The oracle is again the committed examples plus a testrun : T CASINO: the unit suite asserts that what this writes for the geminal calculations under examples/ declares the same parameters and imposes the same constraint groups as their hand-written parameters.casl, and tests/integration/test_geminal_casl.py puts the files to CASINO itself, which parses the block, resolves the constraint groups, checks them for contradictions and calls check_umat before it stops.

Tests

pytest                      # 337 tests, ~6 s, no CASINO needed

The unit suite runs anywhere: the parser is checked field by field against five real out files under tests/data/ — each with the input that produced it — and over all eighteen calculations under examples/, while the launcher, the process group and the guardrails are exercised against a fake runqmc shell script.

pytest -m integration

The integration suite needs a real CASINO, but nothing outside this repository. It checks parse_out against CASINO's own envmc over every out in examples/, puts every input recipe to runqmc --check-only and every blank correlation.data and parameters.casl to a testrun : T CASINO, re-runs the whole tree against the installed binary, and drives the server over real stdio MCP, running and stopping actual VMC calculations.

examples/ holds eighteen calculations chosen as a cover of the settings CASINO can be run with — every runtype, basis type, optimiser and wavefunction option appears at least once, and so do the two files a parser gets wrong quietly: a run that never printed an energy, and one interrupted between optimisation cycles. examples/README.md says what each is there for.

They are short and seeded, so the tree doubles as a check on CASINO itself: re-run it on a new release and any line the parser reads that has been renamed or dropped is named, rather than silently becoming a None. Only that is asserted — moved numbers are reported for a person to judge.

python tools/refresh_examples.py --nproc 4     # run the tree, report, touch nothing
python tools/refresh_examples.py --nproc 4 --write   # adopt the new output

tools/protocol_dump.py speaks the JSON-RPC by hand with no SDK and prints every line in both directions. Read it before adding a tool.

Licence

Apache-2.0 — see LICENSE.