Tma1
Local-first observability your agent reads back. TMA1 records every LLM call, then routes what it sees into the agent's next turn via hooks and MCP.
Last verified:
What is Tma1?
TMA1 is a local-first observability tool for AI agents that records every LLM call, tracks token usage, cost, latency, tool calls, and session data entirely on your local machine. It provides four dedicated dashboard views for Claude Code, Codex, OpenClaw, and OTel GenAI agents, auto-detecting the data source without requiring Grafana, cloud services, or YAML configuration.
Key features include token counts and estimated cost per model with burn-rate projections, cache efficiency analysis, p50 and p95 latency percentiles per tool, full conversation replay for debugging, prompt injection detection, webhook error tracking, stuck session alerts, and SQL query access to raw traces. All data is stored locally in ~/.tma1/data/ and never uploaded to remote services. The tool works as a single binary with OTel input and SQL output, requiring no Docker or system packages.
TMA1 is designed for AI developers, engineers building AI coding agents, and teams who need to understand what their agents are doing and what they cost. It's ideal for anyone who doesn't want to send sensitive agent data (codebase, API keys, infrastructure details) to cloud observability services. The tool is fully open source under Apache-2.0 license.
Tma1 pricing
Pricing model: Freemium
Fully free and open source under Apache-2.0 license. No paid tiers, subscriptions, or hidden costs. The tool is completely self-hosted with no cloud account required. All features are included at no cost - you only pay for the LLM API usage from your agent provider (Claude, OpenAI, etc.).
Tma1 pros
- Everything stays local - no data sent to cloud services
- Works with Claude Code, Codex, OpenClaw, and any OTel-compatible agent
- Single binary installation - no Docker or system packages needed
- Fully open source under Apache-2.0 license
- Four dedicated dashboard views auto-detected from data source
- Tracks token usage, cost, latency, and tool calls in real-time
- Full conversation replay for debugging agent decisions
- SQL query access to raw traces and logs
- Built-in alerting for unusual token counts and high error rates
- p50 and p95 latency percentiles per tool with trend charts
- Cache efficiency analysis and burn-rate projections
- Prompt injection detection and webhook error tracking
- No cloud account, Grafana setup, or YAML configuration required
- Delete ~/.tma1/ and everything is gone - no orphaned cloud state
- Auto-discovers GitHub Copilot CLI session logs with zero config
- Embedded database engine managed by tma1-server process
- Hooks support for 27 Claude Code event types
Tma1 cons
- Only works locally - no remote access to dashboards
- Requires agent restart after OTel configuration changes
- Storage usage depends on traffic - few hundred MB per month typical
- No built-in team collaboration features
- Port 14318 must be available or requires manual reconfiguration
- Windows users need PowerShell for installation script
- Claude Code setup requires merging 27 hook events into settings.json
- No managed cloud option for teams needing centralized visibility
Frequently asked questions about Tma1
What agents does TMA1 support?
TMA1 supports any agent that emits OpenTelemetry data. This includes Claude Code (sends metrics, logs, and traces when enhanced telemetry is enabled), Codex (sends logs and traces, plus native metrics when otel.metrics_exporter is configured), OpenClaw (sends traces and metrics), GitHub Copilot CLI (auto-discovered session logs with zero config), and any OTel SDK app with GenAI semantic conventions. The dashboard auto-detects the data source and shows the appropriate view.
Where is my agent data stored?
TMA1 stores all traces and conversation logs locally on your disk in ~/.tma1/data/. Nothing is uploaded to remote services. You can inspect or delete the data at any time. After the first launch (which downloads the embedded database engine once), TMA1 makes no further network calls - no analytics, no crash reports, no update checks.
How do I install TMA1?
On macOS/Linux, run: curl -fsSL https://tma1.ai/install.sh | bash. On Windows PowerShell, run: irm https://tma1.ai/install.ps1 | iex. This downloads the tma1-server binary into ~/.tma1/bin/, starts the server, and prints the dashboard URL at localhost:14318. Wait ~15 seconds for the database to start, then verify with curl -sf http://localhost:14318/health.
How do I configure my agent to send data to TMA1?
Point your agent's OTel endpoint to http://localhost:14318/v1/otlp. For Claude Code, merge environment variables and hooks into ~/.claude/settings.json. For Codex, add OTel configuration to ~/.codex/config.toml. For OpenClaw, install the @openclaw/diagnostics-otel plugin and configure the endpoint. For GitHub Copilot CLI, zero config is needed - TMA1 auto-discovers session logs. For any OTel SDK, set OTEL_EXPORTER_OTLP_ENDPOINT environment variable and restart the agent.
Can I query the data with SQL?
Yes. Run 'mysql -h 127.0.0.1 -P 14002' to connect to the local SQL endpoint, or open localhost:14000/dashboard/ for the built-in query UI. Raw traces are in opentelemetry_traces table, logs in opentelemetry_logs, and native metric tables are auto-created from incoming OTel metrics. All SQL queries go through POST with JSON body to http://localhost:14318/api/query.
How much disk space does TMA1 use?
It depends on traffic and conversation length. A typical setup uses a few hundred MB per month. You can delete the entire ~/.tma1/ directory to remove all data, with no orphaned cloud state or remote accounts to close.
Is TMA1 really secure if my agent reads my API keys?
Yes. That's exactly why TMA1 is local-first. Your agent reads your codebase, API keys, and infrastructure - sending that data to a cloud observability service defeats the purpose. Everything stays in ~/.tma1/ on your local disk, never uploaded anywhere. TMA1 is Apache-2.0 open source, so you can read the code, audit the build, and run it air-gapped if needed.
How do I upgrade TMA1?
On macOS/Linux: curl -fsSL https://tma1.ai/install.sh | bash. On Windows PowerShell: irm https://tma1.ai/install.ps1 | iex. The install script stops the running server, replaces the binary, and restarts - data is preserved. For a clean reinstall that wipes all data, config, and logs, run: curl -fsSL https://tma1.ai/install.sh | TMA1_FORCE=1 bash.
What if the dashboard shows 'Unhealthy' but error rate is 0%?
By default, generic API services use latency thresholds where p95 > 2s is yellow and p95 > 5s is red. For OpenClaw, the thresholds are p95 > 10s (yellow) and p95 > 30s (red) to better match typical LLM/agent call durations. Focus on the error rate, not latency color. If error rate is 0% and requests are completing successfully, the system is healthy - long-running LLM calls can exceed latency thresholds while functioning as expected.
No data is showing after setup - what should I check?
First verify TMA1 is running: curl -sf http://localhost:14318/health. Then verify database is healthy: curl -sf http://localhost:14000/health. Check if tables exist by running SHOW TABLES via the query API. If no tables, the agent hasn't sent data yet - ensure the agent was restarted after configuring the OTel endpoint, then wait ~1 minute. If tables exist but dashboard is empty, check the time range selector - data might be outside the selected window.