Skip to main content
Glama
Sourcery MCP Server

Sourcery MCP Server

Sourcery Agent Bundle

A hybrid plugin bundle for agent-driven Sourcery security workflows: an MCP server (12 tools over Sourcery's public security API, pinned OpenAPI snapshot), two Agent Skills, and adapters for Claude Code, Cursor, Codex/ChatGPT, VS Code, Devin, Grok (CLI + Bot), Amp, Hermes, OpenClaw, and Rovo Dev CLI. The server ships as the npm package @abyssbugg/sourcery — most hosts just register npx -y @abyssbugg/sourcery@latest. See docs/COMPATIBILITY.md for per-host install steps.

What is confirmed

Sourcery's current docs say the only public API is a Team-plan REST API for security findings. The documented API base is https://api.sourcery.ai/api, and the first documented request is:

curl https://api.sourcery.ai/api/v1/security-issues \
  -H "Authorization: Bearer $SOURCERY_API_KEY"

Verified and pinned. The live OpenAPI document (https://api.sourcery.ai/api/openapi.json) is committed as a snapshot at openapi/sourcery-openapi.json:

  • Info: Sourcery API 0.1.0 (OpenAPI 3.1.0)

  • Fetched: 2026-09-10

  • SHA-256: 9fdb45e657a6406c86ea0d0def2963d534e20bec1a9c88f7d62dbc2e6807f877

Every tool and the client allow-list are wired from that snapshot. The verified surface is exactly eight operations:

Method

Path

Purpose

GET

/v1/security-issues

List issues (filters + cursor pagination)

GET

/v1/security-issues/stats

Counts by status and severity

GET

/v1/security-issues/{id}

Fetch one issue

PATCH

/v1/security-issues

Bulk update (max 100 ids)

GET

/v1/security-issue-groups

List groups (same filters)

GET

/v1/security-issue-groups/stats

Counts by status and severity

GET

/v1/security-issue-groups/{id}

Fetch one group including its issues

PATCH

/v1/security-issue-groups

Bulk update groups (max 100 ids)

Related MCP server: vibecheck

Why the Sourcery API + Git connector split

Sourcery's PR review controls are currently documented as GitHub/GitLab comment or label commands (review, summary, guide, title, resolve, dismiss, create issue) rather than public REST endpoints. Those actions are therefore best implemented through your GitHub/GitLab connector, while Sourcery is used for security findings.

Bundle shape

Repo root = plugin root. The portable core (plugin.json + skills/ + mcp.json) follows the Agent Plugins 1.0.0 standard, which Cursor and OpenAI (Codex, ChatGPT desktop) load natively; the other files are thin per-host adapters.

sourcery-agent-bundle/
  plugin.json                      # Agent Plugins core manifest
  mcp.json                         # portable stdio wiring -> bin/sourcery
  skills/
    sourcery-triage/SKILL.md       # findings triage workflow
    sourcery-remediate/SKILL.md    # minimal-change fix workflow
  bin/
    sourcery                       # launcher: dist/cli.js when built, else npx @abyssbugg/sourcery
  agents/
    sourcery-triager.md            # read-only triage sub-agent (Claude format)
  .claude-plugin/plugin.json       # Claude Code manifest + userConfig API key
  .mcp.json                        # Claude Code MCP wiring (npx)
  .agents/plugins/marketplace.json # Codex/ChatGPT local marketplace entry
  scripts/
    install-rovodev.sh             # Rovo Dev CLI wiring (mcp.json + skills)
    install-local-hosts.sh         # grok, amp, hermes, openclaw, VS Code, Devin
  examples/
    mcp.http.json                  # remote deployment template
    hooks/                         # optional hooks example (SessionStart echo; not enabled)
  docs/
    COMPATIBILITY.md               # per-host setup matrix
    grok-bot-skill.md              # paste-in skill for Grok Bot
  openapi/
    sourcery-openapi.json          # pinned spec snapshot (SHA-256 above)
  src/                             # TypeScript: MCP server + findings CLI (@abyssbugg/sourcery)
    constants.ts                   # pinned facts: operations, enums, spec hash
  tests/                           # TypeScript tests (constants drift pin, tools, client)
  python/                          # Python reference implementation (server, client, prompts)
    src/sourcery_agent/
    tests/
  package.json                     # npm package manifest (@abyssbugg/sourcery)
  .env.example
  README.md

Tools

  • sourcery_security_snapshot: counts by status/severity plus the first page of active issues — start here for a triage overview.

  • sourcery_list_findings: list issues with repository_ids, issue_types, statuses, search, limit, cursor.

  • sourcery_get_finding: fetch one issue (full record incl. source snippet and dependency graph).

  • sourcery_get_security_counts: aggregate counts by status and severity.

  • sourcery_bulk_update_findings: bulk status/severity changes (≤100 ids; SOLVED is scanner-owned and rejected).

  • sourcery_list_groups, sourcery_get_group, sourcery_get_group_counts, sourcery_bulk_update_groups: the same capabilities for issue groups.

  • sourcery_build_fix_prompt: builds the Sourcery-style minimal-change agent prompt from a finding; the DEPENDENCY path uses fixed_versions + manifest_file_path and renders the dependency chain.

  • sourcery_capabilities: describes the verified surface, spec pin, and enum values.

  • sourcery_api_request: compatibility bridge restricted to the eight verified operations.

CLI

npx -y @abyssbugg/sourcery@latest snapshot                          # counts + first page of active findings
npx -y @abyssbugg/sourcery@latest list --status ACTIVE --limit 50   # filterable list
npx -y @abyssbugg/sourcery@latest get 1234                          # one finding in full
npx -y @abyssbugg/sourcery@latest fix-prompt 1234                   # minimal-change agent prompt

Inside an enabled plugin host, bin/sourcery runs the same CLI (it prefers a locally built dist/cli.js and otherwise falls back to npx). Add --json to any query command for scripting.

Setup

Requirements: Node 20+. Most hosts register the MCP server directly:

export SOURCERY_API_KEY='...'
npx -y @abyssbugg/sourcery@latest        # stdio MCP server (no args)
npx -y @abyssbugg/sourcery@latest http   # Streamable HTTP (127.0.0.1:8765/mcp)

Or install it as a plugin bundle — one step per host (details in docs/COMPATIBILITY.md):

Host

Quick start

Claude Code

claude --plugin-dir . — prompts for the API key on enable

Cursor

symlink the repo into ~/.cursor/plugins/local/, reload

Codex CLI / ChatGPT desktop

codex plugin marketplace add ./

ChatGPT web

run npx -y @abyssbugg/sourcery@latest http behind HTTPS, add via developer mode

VS Code / Grok / Amp / Hermes / OpenClaw

scripts/install-local-hosts.sh (idempotent; registers plugin, MCP, and skills)

Devin

skill links via the installer; devin plugins install --local . for the full plugin

Grok Bot

paste-in skill from docs/grok-bot-skill.md

Rovo Dev CLI

scripts/install-rovodev.sh

For development and tests, see the Python reference implementation (its suite doubles as the bundle-packaging checker) and run npm test for the TypeScript suite. The server is built on the official MCP TypeScript SDK (@modelcontextprotocol/sdk).

Regenerating the pinned snapshot

When Sourcery publishes a spec change:

curl -sS --fail -o openapi/sourcery-openapi.json https://api.sourcery.ai/api/openapi.json
shasum -a 256 openapi/sourcery-openapi.json   # update SPEC_SHA256 in src/constants.ts (+ python/src/sourcery_agent/constants.py, README/docs)
npm test                                      # tests/constants.spec.ts fails on drift

Sourcery's PR review commands (review, summary, guide, title, resolve, dismiss, create issue) remain GitHub/GitLab comment or label commands — implement those through your Git provider connector, not the Sourcery REST API.

Available Tools

12 tools
sourcery_api_requestB

Compatibility bridge: call one of the eight verified operations directly.

Prefer the typed tools. The path allow-list is enforced in the client; anything outside the pinned OpenAPI surface is rejected before a request is made.

Inputs: path (required), method (required), body_json, params_json.

sourcery_build_fix_promptB

Build a minimal-change agent prompt from a Sourcery finding object.

Handles DEPENDENCY findings via fixed_versions + manifest_file_path and renders the dependency chain when dependency_graph is present.

Inputs: finding_json (required).

sourcery_bulk_update_findingsB

Bulk-update issue status/severity (max 100 ids; SOLVED cannot be set manually).

Returns updated_ids plus failed entries with reason not_found or not_eligible.

Inputs: ids (required), reason, status, snoozed_until, +1 more.

sourcery_bulk_update_groupsC

Bulk-update groups (max 100 ids); a group updates when at least one issue changes.

Inputs: ids (required), reason, status, snoozed_until, +1 more.

sourcery_capabilitiesC

Return the verified Sourcery integration boundary for this bundle.

Inputs: no named parameters.

sourcery_get_findingB

Fetch a single security issue (full record incl. source snippet and dependency graph).

Inputs: finding_id (required).

sourcery_get_groupB

Fetch a single issue group including all its issues and any linked tracker task.

Inputs: group_id (required).

sourcery_get_group_countsC

Aggregate group counts by status and severity.

Inputs: issue_types, repository_ids.

sourcery_get_security_countsC

Aggregate issue counts by status and severity.

Inputs: issue_types, repository_ids.

sourcery_list_findingsB

List security issues with spec filters; pass a previous next_cursor to page.

Inputs: limit, cursor, search, statuses, +2 more.

sourcery_list_groupsC

List security issue groups (same filters as findings; groups aggregate one rule/package).

Inputs: limit, cursor, search, statuses, +2 more.

sourcery_security_snapshotC

Triage overview: counts by status/severity plus the first page of active issues.

Inputs: limit, issue_types, repository_ids.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3/5.0

    Scored across 12 tools

    Disambiguation3/5

    Tools are organized around three distinct resources (findings, groups, counts) with clear actions, but there is overlap between sourcery_security_snapshot, sourcery_get_security_counts, and sourcery_get_group_counts, which could cause confusion about which to use for simple count retrieval. The compatibility bridge sourcery_api_request overlaps with all typed operations.

    Naming Consistency4/5

    Nearly all tools follow a consistent 'sourcery_verb_noun' pattern (list_findings, get_group, bulk_update_findings, build_fix_prompt). Slight inconsistency in 'security_snapshot' and 'api_request' which are nouns rather than verb_noun, but overall very predictable.

    Tool Count4/5

    12 tools is a reasonable number for a security finding management server. The presence of both typed tools and a generic API request bridge is slightly redundant but justified as a compatibility layer.

    Completeness4/5

    The server provides full lifecycle coverage for findings and groups (list, get, bulk update) and helpful aggregation tools. However, it lacks single-item update operations (only bulk), and does not support creating or deleting findings/groups, which may be by design but creates minor gaps for common workflows.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      B
      quality
      C
      maintenance
      Allows developers to query security findings (SAST issues, secrets, patches) using natural language within AI-assisted tools like Claude Desktop, Cursor, and other MCP-compatible environments.
      17
      11
      MIT
    • A
      license
      A
      quality
      D
      maintenance
      Agent-native "safe to ship?" security gate for AI-generated code. Uses real parsers and inter-rocedural taint analysis (JS/TS, Python, Go) to flag the classes AI coding agents get wrong — secrets, SQL injection, SS, SSRF, path traversal, command injection, weak JWT/CORS — and ranks findings by confidence. Exposes a scan tool over MCP.
      1
      8 npm
      2
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Enables AI agents to scan code for security vulnerabilities using multiple static analysis tools, with support for filtering, deduplication, and CI/CD integration.
      27
      48 PyPI
      4
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Exposes six security tools (scan_repo, scan_diff, list_findings, get_finding, verify_finding, list_detectors) over stdio or HTTP so AI coding agents can run SAST/SCA/secrets/IaC scans and inspect prioritized findings without writing to the repository. It also lets agents verify a finding by re-running its generated proof-of-concept in an isolated sandbox, returning a confirmed or unavailable verdict.
      17
      Apache 2.0