io.github.mrhpython/lens-mcp

Soulfield Lens — MCP server

Long jobs can use validation_submit followed by validation_receipt on a fresh MCP connection. Save the returned job ID with the candidate SHA-256 locally. Retrieve that ID after a polling timeout; do not submit the content again. Pending, failed, expired and incomplete jobs confer no clearance. An uncertain submission without a returned ID must not be automatically retried. The existing validate_content tool is unchanged. See test/validation-jobs.test.cjs for offline recovery and Rights-preservation fixtures.

Outside-in validation for AI-generated text, as an MCP tool.

Every AI tool asks the same model that wrote the answer whether it's any good. It says yes. The Soulfield Lens is outside-in: a separate model runs a fixed gate over your output. It checks text — it doesn't write it. This package puts that gate inside Claude Code, Cursor, and any other MCP-compatible agent, so output can be validated on the path where it's generated.

The gate is fail-closed: a borderline case returns UNKNOWN, never a silent pass. There is no generation step, so it can't fabricate claims of its own — only check. It can still be wrong about a judgment; that is exactly why borderline cases return UNKNOWN instead of a confident yes.

This is a local stdio boundary around the hosted Lens API (api.soulfield.one). It keeps repository access on the caller's machine and sends only the finished artifact text to the API gate.

It exposes two tiers. The gate tier (3 tools) needs nothing but an API key. The validator tier (7 tools) adds one local orchestration entry point plus the optional lens-kit cross-file checks and defect memory that the single-document gate cannot see. Skip local evidence inputs and the gate tier works exactly as before.

Try it before you install anything

The keyless demo endpoint runs the same gate — a few runs a day per IP, no signup:

curl -s https://api.soulfield.one/v1/demo \
  -H 'content-type: application/json' \
  -d '{"text": "<paste the AI output you are about to ship>"}'

Install

npm install -g @soulfield/lens-mcp

Or run without installing: npx @soulfield/lens-mcp.

Claude Code

claude mcp add soulfield-lens \
  -e SOULFIELD_API_BASE=https://api.soulfield.one \
  -e SOULFIELD_API_KEY=<your-key> \
  -- npx @soulfield/lens-mcp

Any MCP client (JSON config)

{
  "mcpServers": {
    "soulfield-lens": {
      "command": "npx",
      "args": ["@soulfield/lens-mcp"],
      "env": {
        "SOULFIELD_API_BASE": "https://api.soulfield.one",
        "SOULFIELD_API_KEY": "<your-key>"
      }
    }
  }
}

Production calls need an API key — request one at hello@soulfield.one. lens_health works without one.

Tools

Gate tier — hosted API, works out of the box

Validator tier — local; consistency checks require lens-kit

Version note: the validator tier lands in 1.1.0. If npm view @soulfield/lens-mcp version still reports 1.0.0, the registry has not caught up with this repo yet and npx @soulfield/lens-mcp will give you the three gate-tier tools only. Install from source in the meantime.

Artifact orchestration: validate_artifact lands in 1.2.0. Its bounded local evidence search is built into this MCP package. Requested consistency checks still require lens-kit.

Side effect worth knowing: every validator-tier call appends a row to RUNS.md in its working directory — that is the kit's run ledger, by design. The directory is the cwd argument, or the server's own cwd if you omit it, so pass cwd explicitly if you care where the ledger lives. Sensitive flag values are redacted in the row (--deny <redacted>), so deny terms do not land on disk.

Prerequisite: pip install lens_kit (Apache-2.0, github.com/mrhpython/lens-kit), or set LENS_KIT_BIN to its path. Without it these six tools return UNKNOWN with an error — never a silent pass. No API key needed: they run locally and make no LLM call.

Why they run locally and not on the hosted API: they take file paths off your disk. A hosted endpoint that accepted arbitrary local paths would be a file-disclosure vector, not a feature. On stdio the paths are your own machine, so the capability is safe here and only here — and for that reason it will not be added to the hosted API.

Choosing deny terms and markers. These three are tripwires, not oracles — on a live run over this project's own copy they produced six flags and zero true defects, across three distinct false-positive classes (negation, superseded-figure, casing). That is the designed behaviour, and it is why the doctrine is adjudicate, never auto-apply. Deny terms work best as strings that are wrong in every context — a real client name, an internal codename — rather than claims you do not make, which legitimately appear inside disclaimers. Markers work best when their casing is stable across source and render. | lens_catches_relevant | Reads the defect bank before you validate: prior named defects for an artifact type, most-recurrent first. Patterns at threshold are marked [PROMOTE] — they recur often enough to deserve a fixed check. | — | | lens_catches_add | Records a named defect so it is caught next time: what was wrong, the general pattern, the forward rule. Routine passes are rejected by design — only real defects. | — | | lens_catches_stats | Per-pattern recurrence counts with promote suggestions. Tells you what to harden next. | — |

The two tiers are complementary, not alternatives. The gate has no tools and no file access — that is precisely what makes it an independent check, and it is also why it cannot see a contradiction spread across two files. The validator tier sees the disk; the gate owns the score. Pair them: gather substrate evidence with the local tools, hand the text to the gate, and never argue a gate FAIL into a PASS. Full protocol: docs/VALIDATOR-AGENT.md.

One local entry point

validate_artifact makes the boundary explicit in one call:

{
  "artifact_path": "output/candidate.md",
  "evidence_roots": ["reports", "output/proof"],
  "evidence_queries": [
    {
      "claim": "measured-result",
      "query": "113/113 rows scored",
      "case_sensitive": true
    }
  ],
  "domain": "agency",
  "context": "Internal technical handoff"
}

The response includes evidence_scope, files_read, searches_and_commands, claims_verified, and claims_unverified. claims_verified is deliberately empty: grep proves that text exists at a source location, not that the claim is true. The separate validator agent must read the located evidence, reason across the files, and attach the final full-validator receipt.

Evidence file contents and local absolute paths are never added to the API request. Directory searches skip .git, node_modules, symlinks, and binary files and fail closed when configured safety limits are exceeded.

What you get per run: receipts — what was checked, what passed, what was held, and why. Machine-readable, not a badge. We won't hand you a guaranteed accuracy number for your data: scores don't transfer across models, datasets, and runtimes, and a tool that promises a fixed figure on data it has never seen is making the exact claim this gate exists to catch.

Long inputs

Inputs of ~4,000 characters and up are submitted as an async job and polled to completion automatically, so a single long validation never dies on a request timeout. Short inputs use the fast synchronous path. No configuration needed.

Configuration (env vars)

The rest of the product

This wrapper is one of several surfaces on the same engine:

• Free one-output audit — api.soulfield.one/audit. The audit is the demo.
• Wire it in (SDK stop-hook and middleware) — api.soulfield.one/developers.
• Own it — the kit: lenses, compiler, self-improve loop, validator agent, Apache-2.0. Train it on your own data. Public repo: github.com/mrhpython/lens-kit — clone it, pip install -e ".[dev]", and the test suite runs offline with no key. Installing it is also what activates the validator tier above.

We hold our own marketing copy to the same gate this package exposes.

License

MIT — see LICENSE. (The lens-kit product is licensed separately under Apache-2.0.)