Earnie

The project’s rules, loaded before the first line of code.

For coding agents, and the people who connect them. Earnie’s MCP server puts the project’s policies in the agent’s instructions before it writes, and reviews its changes at the end of every turn.

  • Claude Code
  • Cursor
  • Codex
  • VS Code
Reading this as an agent? Here’s the plain-text version.

One session, start to finish

Here is what Earnie does during a coding agent’s session, from the moment it opens to the moment someone asks for the record.

  1. The session opens with the rules already loaded

    Set up with the Earnie CLI, the project’s policies sit in the agent’s instructions file as plain rules, one sentence each: CLAUDE.md, AGENTS.md, .cursor/rules/earnie.mdc or .github/copilot-instructions.md. A session-start hook refreshes them when the policies change.

    Without the CLI, the agent calls earnie_policy_context and gets the same rules.

    # CLAUDE.md
    
    <!-- earnie:policies:start -->
    Don’t add dependencies licensed under AGPL-3.0.
    Don’t introduce RSA keys shorter than 3072 bits.
    Don’t add AI models without a declared licence.
    <!-- earnie:policies:end -->
  2. It writes to those rules

    The agent knows what’s forbidden before it picks a package or a cipher, so most problems never get written in the first place.

    add PDF export to the invoices page
    
    Adding pdf-forge 2.3.0 to package.json
    Writing src/invoices/export.ts
  3. The turn ends with a review

    In Claude Code, Cursor and Codex, a review hook sends the changed files to Earnie at the end of each turn, without the agent having to ask. The agent gets a gate, each violation worst first with a suggested fix, and a list of the scanners that ran.

    earnie_review_code  2 files
    
    gate      block
    violation pdf-forge 2.3.0 is AGPL-3.0
    fix       use an MIT- or Apache-2.0-licensed
              alternative
    scanners  oss, crypto
  4. It fixes the code and checks again

    Each review is a Self-check of only the files it was sent. It creates no findings and leaves the project’s posture alone, and its result stays linked in the web app for 30 days.

    A pass means nothing sent broke a policy. The pull request check still decides what merges.

    earnie_review_code  2 files
    
    gate      pass
    scanners  oss, crypto
  5. Later, someone asks where things stand

    Questions come in plain language and the agent picks the tool. earnie_project_posture reports the gate, what needs attention and the key counts. earnie_search_findings and earnie_get_finding go into the detail.

    is invoices-service blocked?
    
    earnie_project_posture
    gate     warn
    needs    notice file missing
  6. And then for the record

    earnie_export_bom returns a CycloneDX 1.7 SBOM, CBOM or AIBOM from a snapshot Earnie has already made. earnie_export makes notices, reports and compliance packages, each one saved to the project’s Reports page under the key’s name.

    export the CBOM for invoices-service
    
    earnie_export_bom  include: crypto
    format   CycloneDX 1.7
    stale    false

An example session. The project, package and policies are made up; the tool names, gates and files are Earnie’s.

The key decides what a coding agent can do

The agent authenticates with an API key, created by an Admin or Operator with the Coding agent (MCP) preset. It sees only the tools the key’s permissions allow, and only the projects the key reaches.

Triage stays off unless someone turns on Allow agents to triage findings when they create the key. Every decision the agent records is logged under the key’s name.

PermissionTools it unlocks
Any valid keyplatform_health_ping
projects:readearnie_list_projects
policies:readearnie_policy_context
findings:readearnie_search_findingsearnie_get_findingearnie_list_vulnerabilities
scans:readearnie_project_postureearnie_export_bomearnie_list_exports
scans:writeearnie_review_codeearnie_export
findings:writeearnie_triage_finding, only with triage turned on
vulnerabilities:decideearnie_decide_vulnerability

Which coding agents can connect

The MCP server lives at your Earnie address followed by /mcp/v1. Every Earnie runs one, with nothing to switch on.

ClientWhat the Earnie CLI sets up
Claude CodePolicy block, review hooks and session-start hook
CursorPolicy block, review hooks and session-start hook
CodexPolicy block, review hooks and session-start hook, once you approve the hooks with /hooks
VS CodePolicy block. Hooks follow when VS Code’s hook support leaves preview.
Claude.ai, ChatGPTLimited. They connect from the vendor’s cloud and expect OAuth, which Earnie doesn’t offer yet.
Back to the full page
# Earnie for coding agents

Earnie's MCP server gives you the project's policies as plain rules
before you write, and reviews your changes before anyone sees them.

## Connection
- Endpoint: your Earnie address + /mcp/v1
- Auth: Authorization: Bearer <key> on every request
- Key: created with the Coding agent (MCP) preset
- Clients: Claude Code, Cursor, Codex, VS Code

## At the start of a session
If the Earnie CLI set up your repository, the rules are already in
your instructions file, between <!-- earnie:policies:start --> and
<!-- earnie:policies:end -->. Otherwise call earnie_policy_context.

## After you write or edit code
If you added or changed a dependency, copied code, used cryptography,
or called an AI SDK, model or agent framework, send the changed files
or a diff to earnie_review_code. (With review hooks, this happens at
the end of each turn.)

Gates: pass, warn, block, require, error, none.
- require: held until a Policy Approval is granted. Not mergeable.
- error: a policy applies but lacks the evidence it needs.
- none: no policy applied. Not a pass.

A review is a Self-check of only what you sent. It creates no findings
and doesn't change posture. The pull request check decides what merges.

## When asked where things stand
- earnie_project_posture   gate, what needs attention, key counts
- earnie_search_findings   OSS, dependencies, crypto, AI provenance
- earnie_get_finding       one finding in full, with policy verdicts

## When asked for the record
- earnie_list_exports      check first
- earnie_export_bom        CycloneDX 1.7 SBOM, CBOM or AIBOM
- earnie_export            notice, report or compliance package;
                           each call is a permanent record

## Limits
You see only the tools your key allows. Triage needs
"Allow agents to triage findings" on the key. Decisions are logged
under the key's name.

## Docs
- Overview: https://docs.scanoss.com/en/latest/earnie/mcp/overview
- Connecting a client: https://docs.scanoss.com/en/latest/earnie/mcp/connecting-a-client
- Tools: https://docs.scanoss.com/en/latest/earnie/mcp/tools
- Troubleshooting: https://docs.scanoss.com/en/latest/earnie/mcp/troubleshooting

Every turn ends with a review.

Request a demo
Earnie for coding agents: policies in, reviews out
Earnie

The project’s rules, loaded before the first line of code.

For coding agents, and the people who connect them. Earnie’s MCP server puts the project’s policies in the agent’s instructions before it writes, and reviews its changes at the end of every turn.

  • Claude Code
  • Cursor
  • Codex
  • VS Code
Reading this as an agent? Here’s the plain-text version.

One session, start to finish

Here is what Earnie does during a coding agent’s session, from the moment it opens to the moment someone asks for the record.

  1. The session opens with the rules already loaded

    Set up with the Earnie CLI, the project’s policies sit in the agent’s instructions file as plain rules, one sentence each: CLAUDE.md, AGENTS.md, .cursor/rules/earnie.mdc or .github/copilot-instructions.md. A session-start hook refreshes them when the policies change.

    Without the CLI, the agent calls earnie_policy_context and gets the same rules.

    # CLAUDE.md
    
    <!-- earnie:policies:start -->
    Don’t add dependencies licensed under AGPL-3.0.
    Don’t introduce RSA keys shorter than 3072 bits.
    Don’t add AI models without a declared licence.
    <!-- earnie:policies:end -->
  2. It writes to those rules

    The agent knows what’s forbidden before it picks a package or a cipher, so most problems never get written in the first place.

    add PDF export to the invoices page
    
    Adding pdf-forge 2.3.0 to package.json
    Writing src/invoices/export.ts
  3. The turn ends with a review

    In Claude Code, Cursor and Codex, a review hook sends the changed files to Earnie at the end of each turn, without the agent having to ask. The agent gets a gate, each violation worst first with a suggested fix, and a list of the scanners that ran.

    earnie_review_code  2 files
    
    gate      block
    violation pdf-forge 2.3.0 is AGPL-3.0
    fix       use an MIT- or Apache-2.0-licensed
              alternative
    scanners  oss, crypto
  4. It fixes the code and checks again

    Each review is a Self-check of only the files it was sent. It creates no findings and leaves the project’s posture alone, and its result stays linked in the web app for 30 days.

    A pass means nothing sent broke a policy. The pull request check still decides what merges.

    earnie_review_code  2 files
    
    gate      pass
    scanners  oss, crypto
  5. Later, someone asks where things stand

    Questions come in plain language and the agent picks the tool. earnie_project_posture reports the gate, what needs attention and the key counts. earnie_search_findings and earnie_get_finding go into the detail.

    is invoices-service blocked?
    
    earnie_project_posture
    gate     warn
    needs    notice file missing
  6. And then for the record

    earnie_export_bom returns a CycloneDX 1.7 SBOM, CBOM or AIBOM from a snapshot Earnie has already made. earnie_export makes notices, reports and compliance packages, each one saved to the project’s Reports page under the key’s name.

    export the CBOM for invoices-service
    
    earnie_export_bom  include: crypto
    format   CycloneDX 1.7
    stale    false

An example session. The project, package and policies are made up; the tool names, gates and files are Earnie’s.

The key decides what a coding agent can do

The agent authenticates with an API key, created by an Admin or Operator with the Coding agent (MCP) preset. It sees only the tools the key’s permissions allow, and only the projects the key reaches.

Triage stays off unless someone turns on Allow agents to triage findings when they create the key. Every decision the agent records is logged under the key’s name.

PermissionTools it unlocks
Any valid keyplatform_health_ping
projects:readearnie_list_projects
policies:readearnie_policy_context
findings:readearnie_search_findingsearnie_get_findingearnie_list_vulnerabilities
scans:readearnie_project_postureearnie_export_bomearnie_list_exports
scans:writeearnie_review_codeearnie_export
findings:writeearnie_triage_finding, only with triage turned on
vulnerabilities:decideearnie_decide_vulnerability

Which coding agents can connect

The MCP server lives at your Earnie address followed by /mcp/v1. Every Earnie runs one, with nothing to switch on.

ClientWhat the Earnie CLI sets up
Claude CodePolicy block, review hooks and session-start hook
CursorPolicy block, review hooks and session-start hook
CodexPolicy block, review hooks and session-start hook, once you approve the hooks with /hooks
VS CodePolicy block. Hooks follow when VS Code’s hook support leaves preview.
Claude.ai, ChatGPTLimited. They connect from the vendor’s cloud and expect OAuth, which Earnie doesn’t offer yet.
Back to the full page
# Earnie for coding agents

Earnie's MCP server gives you the project's policies as plain rules
before you write, and reviews your changes before anyone sees them.

## Connection
- Endpoint: your Earnie address + /mcp/v1
- Auth: Authorization: Bearer <key> on every request
- Key: created with the Coding agent (MCP) preset
- Clients: Claude Code, Cursor, Codex, VS Code

## At the start of a session
If the Earnie CLI set up your repository, the rules are already in
your instructions file, between <!-- earnie:policies:start --> and
<!-- earnie:policies:end -->. Otherwise call earnie_policy_context.

## After you write or edit code
If you added or changed a dependency, copied code, used cryptography,
or called an AI SDK, model or agent framework, send the changed files
or a diff to earnie_review_code. (With review hooks, this happens at
the end of each turn.)

Gates: pass, warn, block, require, error, none.
- require: held until a Policy Approval is granted. Not mergeable.
- error: a policy applies but lacks the evidence it needs.
- none: no policy applied. Not a pass.

A review is a Self-check of only what you sent. It creates no findings
and doesn't change posture. The pull request check decides what merges.

## When asked where things stand
- earnie_project_posture   gate, what needs attention, key counts
- earnie_search_findings   OSS, dependencies, crypto, AI provenance
- earnie_get_finding       one finding in full, with policy verdicts

## When asked for the record
- earnie_list_exports      check first
- earnie_export_bom        CycloneDX 1.7 SBOM, CBOM or AIBOM
- earnie_export            notice, report or compliance package;
                           each call is a permanent record

## Limits
You see only the tools your key allows. Triage needs
"Allow agents to triage findings" on the key. Decisions are logged
under the key's name.

## Docs
- Overview: https://docs.scanoss.com/en/latest/earnie/mcp/overview
- Connecting a client: https://docs.scanoss.com/en/latest/earnie/mcp/connecting-a-client
- Tools: https://docs.scanoss.com/en/latest/earnie/mcp/tools
- Troubleshooting: https://docs.scanoss.com/en/latest/earnie/mcp/troubleshooting

Every turn ends with a review.

Request a demo