chimeralang-mcp

Give Claude typed confidence, hallucination detection, and constraint enforcement as native MCP tools.

ChimeraLang is a programming language built for AI cognition. This MCP server exposes its runtime as 44 tools Claude can call during any conversation. No Anthropic permission needed, works today with Claude Desktop and Claude Code.

Install

pip install chimeralang-mcp
# or
uvx chimeralang-mcp

Development quick checks

pytest -q
python -m pip check

pytest is configured via pyproject.toml to include project root on PYTHONPATH.

For token counting internals, these optional environment variables are supported:

• CHIMERA_TOKEN_CACHE_MAX_ENTRIES (default 2048) — bounds in-memory token count cache size.
• CHIMERA_TOKEN_FALLBACK_LOG_INTERVAL_S (default 60) — throttles repeated fallback warning logs.

Claude Desktop Setup

Add to your config file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "chimeralang": {
      "command": "uvx",
      "args": ["chimeralang-mcp"]
    }
  }
}

Or with a pip-installed version:

{
  "mcpServers": {
    "chimeralang": {
      "command": "python",
      "args": ["-m", "chimeralang_mcp"]
    }
  }
}

Restart Claude Desktop. 44 ChimeraLang tools are now available.

Tools

The core language/runtime tools are executable and deterministic. Several higher-level reasoning, safety, and cognition helpers are lightweight local heuristics intended for planning, triage, and guardrails rather than authoritative verification.

Most stateful tools accept an optional namespace and persist data to ~/.chimeralang_mcp (or CHIMERA_MCP_DATA_DIR) so agents can carry memory, world state, traces, and cost history across sessions. The package also ships an embedded material pack plus offline CLI commands for sync, build, status, and licenses.

Core Language

Confidence and Safety

Reasoning and Cognition

Knowledge and Memory

Provenance and Verification

Token Budget, Cost, and Workflow

The token-efficiency stack now defaults to a deterministic, quantum-inspired compressor: it scores text spans by salience amplitude, boosts shared entities through entanglement, suppresses redundant spans through interference, and then "measures" the best units under a token budget. chimera_optimize, chimera_compress, chimera_fracture, and chimera_csm all accept algorithm="classic|quantum" and optional focus text so compression can stay task-aware instead of blindly deleting tokens.

Meta and Audit

What problem does this solve?

Claude's tool-use loop has no built-in mechanism for:

• Confidence gating - only proceed if confidence is above a threshold
• Typed output contracts - a result must satisfy a constraint before going downstream
• Genuine consensus detection - determine whether multi-path agreement is substantive
• Hallucination signals - structured detection rather than pure intuition
• Trust propagation - confidence and provenance should survive chained tool calls
• Evidence-backed verification - explicit claims checked against supplied evidence
• Persistent reasoning state - memory, world state, and traces carried across sessions
• Cost intelligence - token tracking and compression throughout long sessions

ChimeraLang provides a practical constraint layer between Claude and its tools. The language runtime is the strongest guarantee surface. The higher-level cognition and safety helpers are best treated as lightweight first-pass checks unless you pair them with stronger external evidence.

Example prompts

Gate a value before a critical action: "Before you submit that form, use `chimera_confident` to verify you're at least 0.95 confident the data is correct."

Consensus across reasoning paths: "Generate 3 different answers, then use `chimera_quantum_vote` to collapse to the most reliable one."

Hallucination scan on output: "After you get that search result, run `chimera_detect` with semantic strategy to check for absolute-certainty markers."

Full constraint pipeline: "Use `chimera_constrain` on that tool result with `min_confidence=0.85` and `detect_strategy=semantic`."

Evidence-backed fact checking: "Extract claims with `chimera_claims`, verify them against these sources with `chimera_verify`, then apply `chimera_policy` using `strict_factual`."

Trace and provenance inspection: "Merge the envelopes from those two tool calls with `chimera_provenance_merge`, then inspect the latest trace with `chimera_trace`."

Material-pack inspection and build flow: "List the bundled packs with `chimera_materials`, then run `chimeralang-mcp build` locally if you want generated JSON runtime and eval pack artifacts on disk."

ChimeraLang Quick Reference

Variable Declaration

Both val and let are supported:

val answer = Confident("Paris", 0.97)
let hypothesis = Explore("maybe dark matter is...", 0.4)

Probabilistic Types

emit Confident("verified fact", 0.97)
emit Explore("hypothesis", 0.60)

Assertions

assert Confident(0.78) > Confident(0.45)

Gate Declaration

gate verify(claim: Text) -> Converge<Text>
  branches: 3
  collapse: weighted_vote
  threshold: 0.80
  return claim
end

Logical Operators

Both keyword and symbolic forms are supported:

if a > 0.5 and b > 0.5
  emit Confident("both pass", 0.9)
end

if a > 0.5 && b > 0.5
  emit Confident("both pass", 0.9)
end

Hallucination Detection

detect temperature_check
  strategy: "range"
  on: temperature
  valid_range: [-50.0, 60.0]
  action: "flag"
end

If / Else

if confidence > 0.80
  emit Confident("high confidence result", 0.9)
else
  emit Explore("low confidence - needs review", 0.5)
end

For Loop

for item in items
  emit Explore(item, 0.6)
end

Match

match verdict
| "pass" => emit Confident("approved", 0.95)
| "fail" => emit Explore("rejected", 0.70)
| _      => emit Explore("unknown", 0.50)
end

Changelog

0.6.0

• Add a deterministic quantum-inspired token compression engine with salience amplitude, entanglement boosts, redundancy interference, and budgeted measurement
• Make chimera_optimize, chimera_compress, chimera_fracture, and chimera_csm query-aware through optional focus input plus classic|quantum algorithm selection
• Upgrade chimera_fracture to return the actual optimized documents and compressed message artifacts instead of only aggregate stats
• Fix TokenBudgetManager singleton behavior so token-count caches are reused across tool calls

0.5.0

• Add a material-pack subsystem with pinned source manifests, pack builders, a loader, and offline sync / build / status / licenses CLI commands
• Add chimera_materials and wire pack metadata into claims, verification, policy, detect, safety, trace, and audit flows
• Upgrade claim extraction with pack-driven claim typing plus hedge and abstention tagging
• Upgrade verification to FEVER-style verdicts with evidence matches, taint-aware attack flags, and pack-version reporting
• Add MCP security policies, OWASP MCP Top 10 mappings, CI material builds, and an HTTP transport path for conformance workflows

0.4.0

• Add a unified result envelope model with confidence, provenance, transform history, claims, constraints, warnings, and metadata
• Persist namespace-scoped knowledge, memory, world model, self model, meta-learning history, traces, and cost tracking to disk
• Add chimera_claims, chimera_verify, chimera_provenance_merge, chimera_policy, and chimera_trace
• Expose evidence-backed verification and reusable policy application directly through MCP tools
• Add regression coverage for namespace persistence, claim extraction, verification, provenance merge, policy enforcement, and trace inspection

0.3.2

• Preserve structured JSON values through chimera_confident, chimera_explore, and chimera_constrain instead of stringifying them
• Fix multimodal message token estimation fallback so chimera_cost_estimate, chimera_budget, and chimera_csm do not undercount text-array content
• Make chimera_fracture stop dropping history once the budget is satisfied instead of over-pruning to the minimum message floor
• Fix chimera_mode full to report the real live tool inventory and align README documentation with the current 38-tool surface

0.2.7

• Fix UnboundLocalError in chimera_cost_track caused by log variable shadowing the module-level logger in the chimera_audit handler
• Add let as a keyword alias for val in variable declarations
• Add && and || as lexer tokens (aliases for and and or)
• Expand tool count to 33 in the README

0.2.5

• Initial AGI component suite: causal reasoning, deliberation engine, quantum vote, safety layer, ethical reasoner
• Knowledge base, world model, session memory
• Cost tracking, budget management, dashboard
• Self-model and metacognition tools

Links

• ChimeraLang core: github.com/fernandogarzaaa/ChimeraLang
• OpenChimera: github.com/fernandogarzaaa/OpenChimera_v1

License

MIT (c) Fernando Garza

Materials CLI

The wheel includes a compact curated core material pack. Larger corpora stay external and are only touched through manual CLI commands:

chimeralang-mcp status
chimeralang-mcp licenses
chimeralang-mcp sync
chimeralang-mcp build

• status reports bundled pack counts and whether generated artifacts already exist in CHIMERA_MCP_DATA_DIR/materials
• licenses prints the machine-readable source and bundle-policy report
• sync fetches current upstream metadata snapshots for the pinned GitHub and Hugging Face sources
• build writes normalized JSON manifests plus runtime and eval pack files to disk

The runtime MCP tools never fetch from the network. They only consume the bundled core pack and any already-generated local artifacts.