_ _ ___ _ _ _
/ \ __ _ ___ _ __ | |_|_ _|_ __ ___(_) __ _| |__ | |_
/ _ \ / _` |/ _ \ '_ \| __|| || '_ \/ __| |/ _` | '_ \| __|
/ ___ \ (_| | __/ | | | |_ | || | | \__ \ | (_| | | | | |_
/_/ \_\__, |\___|_| |_|\__|___|_| |_|___/_|\__, |_| |_|\__|
|___/ |___/
Local-first observability for OpenAI Codex usage
Sessions · rollouts · token usage · costs · search · repositories · models · skills
Get started · Features · Architecture · Configuration · Contributing · Runbook
AgentInsight turns local Codex session data into a private analytics workspace. It discovers sessions, parses rollout events, estimates cost from token usage, indexes prompts and responses for search, and surfaces dashboard, repository, model, and skills analytics without uploading your provider files.
AgentInsight is in active alpha development. It is intended for local-only usage, private experimentation, and iterative validation against real Codex usage data. APIs, database tables, and analytics views may still change before a stable release.
- Session discovery - reads local Codex state and session files from configured provider paths.
- Rollout parsing - streams
sessions/**/*.jsonlevents without loading large files fully into memory. - Cost estimation - calculates usage from rollout token events with configurable pricing.
- Search - indexes prompts, responses, repositories, and models with results linked back to sessions.
- Dashboard analytics - summarizes sessions, costs, tokens, cache behavior, and activity over time.
- Repository analytics - groups usage and behavior by local project.
- Model analytics - compares model usage, cost, and token patterns.
- Skills analytics - surfaces usage patterns for coding-agent skills and workflows.
- Task and prompt insights - includes token waste analysis, task context recommendations, and prompt optimization.
- Local-first operation - reads provider data as source material and stores the AgentInsight index locally.
Local Codex home
state_*.sqlite
sessions/**/*.jsonl
|
v
AgentInsight API
provider adapters -> rollout parser -> local SQLite index -> analytics services
|
+--> AgentInsight UI
| dashboard -> sessions -> search -> repositories -> costs -> models -> skills -> insights
|
+--> AgentInsight MCP
read-only local context tools for MCP-compatible coding clients
AgentInsight treats Codex data as read-only input. Provider-owned files such as history.jsonl, sessions/, and
state_*.sqlite are never modified. Application-owned data is written only to the configured AgentInsight data path.
The fastest way to run the full stack is Docker Compose:
docker compose up --build -dThen open:
| Service | URL |
|---|---|
| UI | http://localhost:5173 |
| API | http://localhost:8081 |
| Health check | http://localhost:8081/api/health |
| MCP health check | http://localhost:8081/api/mcp/health |
Stop the stack with:
docker compose downFor native development, troubleshooting, and AI-assisted implementation workflows, see the AgentInsight runbook.
Runtime configuration lives in a root .env file. It is intentionally ignored by Git.
AGENTINSIGHT_CODEX_HOME=/path/to/.codex
AGENTINSIGHT_DATA_PATH=/path/to/.agentinsight
AGENTINSIGHT_DB_PATH=/path/to/.agentinsight/agentinsight.db
AGENTINSIGHT_API_PORT=8081
AGENTINSIGHT_UI_PORT=5173
SERVER_PORT=8081
For Docker Compose, AGENTINSIGHT_CODEX_HOME is mounted read-only at /codex and AgentInsight data is stored in the
agentinsight-data volume.
Backend:
cd agentinsight-api
set -a
. ../.env
set +a
./gradlew bootRunFrontend:
cd agentinsight-ui
npm install
npm run devMCP CLI:
cd agentinsight-mcp
go run ./cmd/agentinsight-mcp version
go run ./cmd/agentinsight-mcp tools
go run ./cmd/agentinsight-mcp doctor --api-url http://localhost:8081MCP client setup, including Claude Desktop stdio configuration, is documented in agentinsight-mcp/docs/client-config.md.
Validation commands:
npm run build
npm run test
cd agentinsight-mcp && go test ./...agentinsight-api/ Spring Boot API, parsing, persistence, analytics services
agentinsight-ui/ React/Vite analytics UI
agentinsight-mcp/ Go MCP adapter for local AgentInsight context
docs/ Local runbooks and operational notes
assets/images/ README screenshots
.github/ Issue, pull request, and release templates
AgentInsight uses requirement-driven development. New behavior should start from a documented requirement or issue, and pull requests should explain the requirements implemented, acceptance criteria satisfied, and tests added. See CONTRIBUTING.md for the workflow.
| Area | Description |
|---|---|
| Dashboard | Session, cost, token, cache, and activity summaries |
| Sessions | Session explorer and rollout transcript replay |
| Search | Prompt, response, repository, and model search linked to sessions |
| Costs | Configurable cost estimation from token events |
| Repositories | Usage and activity by local project |
| Models | Token, cache, and cost patterns by model |
| Skills | Analytics for coding-agent skill usage |
| Token Waste Advisor | Evidence-backed waste reports and prompt-only analysis |
| Task Context Analyzer | Repository-scoped context recommendations for new tasks |
| Prompt Optimization | Focused prompt generation, review, comparison, and local save workflows |
| MCP Adapter | Read-only local AgentInsight context exposed to MCP-compatible clients |
Cost calculations originate from rollout token events. The preferred source is token_count, including input tokens,
cached input tokens, and output tokens. Pricing is configurable in:
agentinsight-api/src/main/resources/pricing.yml
Pricing must not be hardcoded into business logic.
| Topic | Link |
|---|---|
| Runbook | docs/RUNBOOK.md |
| Contributing | CONTRIBUTING.md |
| Changelog | CHANGELOG.md |
| Security | SECURITY.md |
| License | LICENSE |
| Root agent instructions | AGENTS.md |
| API agent instructions | agentinsight-api/AGENTS.md |
| UI agent instructions | agentinsight-ui/AGENTS.md |
| MCP README | agentinsight-mcp/README.md |
| MCP agent instructions | agentinsight-mcp/AGENTS.md |
| MCP architecture | agentinsight-mcp/docs/architecture.md |
| MCP client config | agentinsight-mcp/docs/client-config.md |
| MCP tools | agentinsight-mcp/docs/tools.md |
| MCP privacy | agentinsight-mcp/docs/privacy.md |
| Docker Compose | docker-compose.yml |
Releases are created from version tags by the GitHub Actions release workflow. The first public release is v0.1.0;
release notes are tracked in CHANGELOG.md.
AgentInsight is designed for local analysis of local coding-agent data:
- Provider files are read-only.
- Imports continue past corrupt files and unknown events.
- Metrics are reproducible from imported source data.
- No random sampling or hidden calculations are used for analytics.
- Search and analytics results remain traceable to sessions.







