debug-recorder-mcp
Why this exists
Debugging knowledge usually disappears into chat windows, terminals, and commit history. debug-recorder-mcp gives MCP-enabled agents and IDEs a durable local memory so they can answer:
“Have I fixed this before?”
It stores each debugging session, error, command, attempted fix, working fix, tags, and context in a local SQLite database. Search combines SQLite FTS5 with fuzzy reranking, reusable presets, pagination metadata, related-session groups, and optional Markdown exports.
Highlights
- Local-first storage: no external database or hosted service required.
- MCP-native tools: stdio server for desktop MCP clients plus optional Streamable HTTP mode.
- Search that survives messy errors: FTS5 + Fuse.js reranking for stack traces, typos, Unicode, and punctuation-heavy logs.
- Reusable search presets: save common filters and limits for recurring incident patterns.
- Safe operations: redaction-before-store option, explicit destructive confirmations, local HTTP host/origin/auth/body-limit hardening.
- Diagnostics:
get_diagnosticsreturns redacted runtime, schema, package, and health signals for support without leaking raw paths, tokens, stack traces, or command output. - Release-grade packaging: audit, coverage, fuzzing, package-size gates, SBOM/VEX policy, install-script approvals, provenance-ready release workflow, and MCP Registry readiness checks.
Quick start
Requires Node.js 22 LTS or 24 LTS and npm 10+.
npx debug-recorder-mcp
Default database path:
~/.debug-recorder-mcp/sessions.db
Use a custom database path:
DEBUG_RECORDER_DB=/path/to/custom.db npx debug-recorder-mcp
MCP client setup
Desktop MCP clients
{
"mcpServers": {
"debug-recorder-mcp": {
"command": "npx",
"args": ["debug-recorder-mcp"]
}
}
}
VS Code / GitHub Copilot
Create or update .vscode/mcp.json:
{
"servers": {
"debug-recorder-mcp": {
"type": "stdio",
"command": "npx",
"args": ["debug-recorder-mcp"]
}
}
}
More setup examples are in Client setup recipes.
Available MCP tools
| Tool | Purpose |
|---|---|
start_debug_session |
Start tracking a new issue or incident. |
add_fix |
Record a failed or successful fix attempt. |
record_command |
Save a command, output, exit code, and session link. |
close_session |
Mark a session as resolved or abandoned. |
update_session |
Edit title, description, or tags. |
delete_session |
Permanently delete a session with explicit confirmation. |
search_sessions |
Search history with FTS5, fuzzy reranking, pagination, related groups, and optional Markdown export. |
save_search_preset |
Store a reusable query, filters, and limit. |
list_search_presets |
List saved search presets. |
remove_search_preset |
Remove a saved search preset by name. |
find_similar_errors |
Ask whether a similar error has appeared before. |
get_session |
Fetch full details, fixes, and commands. |
get_session_context |
Get an AI-friendly session summary. |
list_sessions |
Browse sessions with filters. |
get_stats |
Summarize debug history. |
get_diagnostics |
Return a redacted operational diagnostics snapshot. |
export_sessions |
Export a full JSON backup or a lightweight summary inventory. |
import_sessions |
Import a validated export payload. |
Real usage examples
Have I seen this before?
Ask your MCP client:
I am getting
TypeError: Cannot read properties of undefined. Have I seen this before?
The client can call find_similar_errors, then inspect the best match with get_session_context.
Record an active incident
- Call
start_debug_sessionwith the problem title and error details. - Add terminal commands with
record_command. - Add each attempted fix with
add_fix. - Improve title, notes, or tags with
update_session. - Close the incident with
close_session.
Back up or migrate history
- Call
export_sessionswithformat: "json". The response is marked withformat: "json"and contains the fullsessions,fixes, andcommandsarrays. - Store the returned object in your backup system.
- Restore later by passing that object as
import_sessions.payload.
For a lightweight inventory, call export_sessions with format: "summary".
Summary responses are marked with format: "summary", include aggregate
stats and abbreviated session rows, and are not restore payloads.
HTTP transport
The package also supports local Streamable HTTP:
npm run start:http
Useful routes:
GET /healthGET /version- MCP endpoint:
POST /mcp
HTTP mode is local-first by default. It binds to 127.0.0.1, creates an isolated stateless MCP server/transport per request, validates Host, validates browser Origin when present, and enforces a JSON body-size limit before the MCP transport receives the request.
For deliberate non-loopback exposure, set all of these:
HOST=0.0.0.0
DEBUG_RECORDER_REMOTE_HTTP=true
DEBUG_RECORDER_HTTP_TOKEN=replace-with-a-long-random-token
DEBUG_RECORDER_ALLOWED_HOSTS=debug-recorder.example.com
DEBUG_RECORDER_ALLOWED_ORIGINS=https://debug-recorder.example.com
npm run start:http
Wildcard origins are rejected for remote mode. The static bearer token is private/shared-secret mode: every caller shares one identity, one authority level, and one SQLite dataset. It is suitable for loopback, an encrypted trusted network, or a private authenticating proxy, but it is not OAuth and does not provide per-user scopes or revocation.
Public multi-user HTTP is not supported by the current release. The accepted target architecture uses an external authorization server and MCP-aware gateway with protected-resource discovery, audience-bound tokens, scopes, rate limits, audit events, and subject-aware storage. See Public HTTP authorization and ADR-0006.
Configuration
| Variable | Description |
|---|---|
DEBUG_RECORDER_DB |
Override the SQLite database path. |
HOST |
HTTP bind host. Defaults to 127.0.0.1. |
PORT |
HTTP port. Defaults to 3000. |
DEBUG_RECORDER_HTTP_TOKEN |
Private/shared-secret token; required for non-loopback HTTP, not OAuth. |
DEBUG_RECORDER_ALLOWED_HOSTS |
Comma-separated HTTP Host allowlist. |
DEBUG_RECORDER_ALLOWED_ORIGINS |
Comma-separated browser Origin allowlist. |
DEBUG_RECORDER_MAX_BODY_BYTES |
HTTP JSON body limit. Defaults to 1048576. |
DEBUG_RECORDER_REMOTE_HTTP |
Enable non-loopback HTTP with true, 1, or yes. |
DEBUG_RECORDER_REDACT_BEFORE_STORE |
Enable pre-store redaction with true, 1, or yes. |
LOG_LEVEL |
Minimum structured log level: debug, info, warn, or error. |
FUZZY_THRESHOLD |
Override the Fuse.js reranking threshold. |
Boolean configuration also accepts false, 0, and no; values are
case-insensitive, whitespace is ignored, and unsupported values fail fast.
Diagnostics reports the resolved effective values rather than re-reading changed
environment strings.
Data and privacy
- Database: local SQLite via
better-sqlite3. - Search index: SQLite FTS5 virtual table.
- Default path:
~/.debug-recorder-mcp/sessions.db. - Redaction: optional before-store redaction plus diagnostics redaction.
- Deletion: destructive operations require explicit confirmation.
- Maintenance: see Storage retention and maintenance.
better-sqlite3uses a native addon. If Node versions change and bindings fail, runnpm rebuild better-sqlite3.
Docker
docker build -t debug-recorder-mcp:local .
docker run --rm -p 127.0.0.1:3000:3000 \
-e HOST=0.0.0.0 \
-e DEBUG_RECORDER_REMOTE_HTTP=true \
-e DEBUG_RECORDER_HTTP_TOKEN=replace-with-a-long-random-token \
-e DEBUG_RECORDER_ALLOWED_HOSTS=127.0.0.1:3000,localhost:3000 \
-e DEBUG_RECORDER_ALLOWED_ORIGINS=http://127.0.0.1:3000,http://localhost:3000 \
debug-recorder-mcp:local
The image installs with npm ci, preserves reviewed native install scripts, prunes development dependencies, and runs as the non-root node user.
Documentation
Published documentation is generated by npm run docs:site and published to GitHub Pages:
https://oaslananka.github.io/debug-recorder-mcp/
Important docs:
- Usage
- Client setup recipes
- Configuration
- Architecture
- Security
- Public HTTP authorization
- Operations
- Troubleshooting
- Storage retention
- Testing
- Release flow
- Install-script policy
- SBOM/VEX policy
- Architecture decision records
Development
npm ci
npm run format:check
npm run lint
npm run check:dead-code
npm run test:coverage
npm run test:fuzz
npm run build
npm run test:e2e
npm run audit
npm run check:install-scripts
npm pack --dry-run
npm run check:package-size
npm run check:version
npm run check:mcp
npm run check:security-policy
npm run check:sbom
npm run docs:site
Full local gate:
npm run ci:local
Release readiness:
npm run prepublishOnly
npm run check:mcp-registry
Release and npm publishing
The normal release workflow uses Release Please, builds release assets, generates SBOM/checksums, attests the tarball, uploads GitHub Release assets, and publishes to npm with provenance.
For the first npm package creation, use the manual Initial npm Token Publish workflow with repository secret NPM_TOKEN. After the package exists and npm trusted publishing is configured, the regular Release workflow can publish through GitHub OIDC. Details are in Release flow.
License
Released under the MIT License. package.json also declares "license": "MIT", and the npm package includes LICENSE.
Funding
If this project saves you debugging time, support development here:
- Buy Me a Coffee: https://www.buymeacoffee.com/oaslananka
- GitHub Sponsors: https://github.com/sponsors/oaslananka
Funding metadata is available in both .github/FUNDING.yml and package.json.
Agent plugin and runtime configuration
This repository owns the product-level agent plugin, MCP runtime configuration, and product-specific skills for debug-recorder-mcp. The central agent-tools repository should catalog this plugin, but the manifest and workflow instructions live here so they stay synchronized with the actual MCP server package.
| File | Purpose |
|---|---|
.claude-plugin/plugin.json |
Claude Code-valid product plugin manifest. |
.mcp.json |
Claude Code project-local MCP server configuration. |
.codex/config.example.toml |
Codex CLI MCP configuration example. |
.vscode/mcp.example.json |
VS Code / GitHub Copilot workspace MCP configuration example. |
opencode.example.jsonc |
OpenCode project MCP configuration example. |
.opencode/skills/ |
OpenCode-native mirrored skill definitions. |
docs/agent-runtime-config.md |
Agent runtime setup and validation notes. |
Validate plugin packaging locally:
claude plugin validate .
No comments yet
Be the first to share your take.