Black-and-white view of a controlled pressure boundary holding a precise line
Free and open source

Code analysis that checks before it claims.

Agents can report confident similarities without comparing the files. Ground makes verification a prerequisite, so every duplicate, dead-code, or orphan claim starts with evidence.

  • Rule Verify first
  • Package Ground MCP
  • Cost Free
Ground operating path

Verification happens before the answer, not after.

One indexed surface keeps installation, enforcement, and production evidence together.

01 / 03 Connect

Put the check beside the agent.

Choose one client. Every command points to the same open-source Ground MCP package.

Claude DesktopAdd to claude_desktop_config.json
npx --yes -p @createsomething/ground-mcp ground-mcp
WindsurfSettings -> MCP -> View raw config
{"mcpServers":{"ground":{"command":"npx","args":["@createsomething/ground-mcp"]}}}
Codex CLIRegister Ground as a local MCP server
codex mcp add ground --command "npx @createsomething/ground-mcp"
npmWorks with any MCP client
npm install -g @createsomething/ground-mcp
Evidence
  • Four client-ready configurations
  • Commands remain visible without JavaScript
Receipts
FreeOpen source@createsomething/ground-mcp
02 / 03 Verify

Check first. Then claim. Block otherwise.

Ground turns verification from a suggestion into a required sequence before an agent reports a code finding.

  1. 01Check first

    Compare the files or count the uses.

  2. 02Then claim

    Attach the finding to what was checked.

  3. 03Blocked otherwise

    Stop an unsupported conclusion before it spreads.

# First, compare the files
ground compare utils.ts helpers.ts

# Then make the evidence-bound claim
ground claim duplicate utils.ts helpers.ts "same validation logic"

Claim blocked: compare these files first.
Evidence
  • Check first
  • Then claim
  • Blocked otherwise
Receipts
Compared inputsEvidence-bound claim
03 / 03 Trust

A smaller tool surface produces stronger claims.

The tool inventory supports one operating rule, and the Kickstand audit shows the rule applied to production code.

Verify

  • ground_compareSee how similar two files actually are
  • ground_count_usesFind whether a function is actually used
  • ground_check_connectionsSee whether a module is connected
  • ground_check_environmentCatch runtime APIs in the wrong environment

Find problems

  • ground_find_duplicate_functionsFind copy-pasted code
  • ground_find_orphansFind files nothing imports
  • ground_find_dead_exportsFind exports nobody uses
  • ground_find_driftFind divergence from an owned pattern

Understand patterns

  • ground_adoption_ratioMeasure token or pattern adoption
  • ground_suggest_patternSuggest from the code already present
  • ground_mine_patternsDiscover repeated structures

Report findings

  • ground_claim_duplicateReport a verified duplicate
  • ground_claim_dead_codeReport verified dead code
  • ground_claim_orphanReport a verified orphan
Evidence
  • 20+ evidence tools
  • 155 scripts became 13
  • Zero reported false positives
Receipts
92% script reductionPublished case study
Ground handoff

Put evidence before the next code claim.

Install Ground in one agent client, run a real verification, and preserve the checked inputs with the finding.

Owner
Codebase operator
Authority
Verify-before-claiming policy
Proof
Checked inputs + finding
State
ready
Performance system

Make the handoff faster without losing the judgment.

Start with one workflow that still needs manual rescue. We make the owner, allowed actions, approval pauses, and proof record visible before a build expands what automation can do.

Outcome Routine work moves. Judgment stays with the operator.

Evidence can be prepared before review while approval, rejection, and consequential action remain with a named person.

Worked example Marketplace template review

The system prepared 49 of 50 selected evidence packets before a human reviewed the templates.

  1. Map the frictionSee the systems, handoffs, and decision owner.
  2. Test one pathLet routine work move while consequential work pauses.
  3. Keep the proofReview what ran, what stopped, and what needs recovery.