TMA1

What does this agent do, and when should you use it?

TMA1 is a local-first observability layer for LLM agents, built on GreptimeDB and shipped as a single binary with no Docker, Grafana, or cloud account required. It ingests traces, metrics, and logs over OTLP/HTTP from Claude Code, Codex, and any GenAI app, plus hook events and JSONL transcripts, and stores everything in a local GreptimeDB instance. It then closes the loop: hook responses become a <tma1-context> block in the agent's next turn, and ten MCP tools answer on-demand questions about session state, anomalies, build status, peer sessions, and past conversations. The embedded dashboard covers token usage, cost estimates, latency, tool activity, conversation replay, and raw SQL. Deployment is a background service on one machine — a launchd agent, a systemd user unit, or a Windows scheduled task — writing to ~/.tma1/, with no authentication because it is explicitly a local-only tool.

tma1-server binds port 14318 and does four jobs at once: it reverse-proxies OTLP, serves the /api/hooks endpoint, hosts the dashboard, and runs an MCP stdio server. OTLP traffic (/v1/otlp, or per-signal /v1/traces, /v1/metrics, /v1/logs) is forwarded to GreptimeDB, which TMA1 downloads and starts as a child process on first launch. Adapters register agent hooks; Claude Code gets all 27 hook events for telemetry but only five inject content (SessionStart, UserPromptSubmit, PostToolUse, Stop, PreCompact), while Codex registers five and injects on four. The tma1 build wrapper (with --watch, --tag, --filter-regex, --debounce) captures build and test output into tma1_build_events for build status and repeated-failure detection. MCP exposes get_context_bundle, get_session_state, get_anomalies, get_build_status, get_external_changes, get_project_state, get_peer_sessions, search_sessions, get_session_transcript, and exec_query, the last accepting a single read-only SELECT. Data is reachable through four surfaces: the TMA1 dashboard, MCP, MySQL protocol on port 14002, and GreptimeDB's own dashboard.

1. A developer running long Claude Code or Codex sessions who wants each new turn to automatically know which files changed, whether the build broke, and what anomalies are active.
2. Someone working with a persistent npm run dev or cargo watch process, using tma1 build --watch so failures reach the agent without manual log pasting.
3. A team mixing Claude Code and Codex on the same project, using /tma1-peer codex so one agent reads the other's work verbatim.
4. An engineer asking "which past session discussed retry backoff?" and answering it with /tma1-search retry backoff plus get_session_transcript.
5. A lead who wants token spend, cache-read share, tool failures, and compaction events compared side by side across two sessions.
6. A data engineer who prefers querying raw traces, metrics, and logs directly over MySQL or the GreptimeDB dashboard.

How do you install or deploy this agent?

One-line installer for macOS or Linux, with TMA1_ADAPTER selecting the agent to wire up:

curl -fsSL https://tma1.ai/install.sh | TMA1_ADAPTER=claude-code bash
curl -fsSL https://tma1.ai/install.sh | TMA1_ADAPTER=codex bash
curl -fsSL https://tma1.ai/install.sh | TMA1_ADAPTER=all bash

Server only, wiring an agent later:

curl -fsSL https://tma1.ai/install.sh | bash

Windows PowerShell, adapter passed as an environment variable:

$env:TMA1_ADAPTER = 'claude-code'; irm https://tma1.ai/install.ps1 | iex

The script registers a background service and starts it; the dashboard is at <http://localhost:14318> when it exits. First start downloads GreptimeDB (pinned to v1.3.0 by default). No Docker or cloud account. From source:

git clone https://github.com/tma1-ai/tma1.git
cd tma1
make build
./server/bin/tma1-server

How do you use this agent?

Once wired, each agent turn receives a <tma1-context> block automatically. Manage adapters per project with:

tma1-server install --adapter claude-code --project .
tma1-server uninstall --adapter claude-code --project .

Capture build and test output:

tma1 build --tag test -- make test
tma1 build --watch --tag dev -- npm run dev

Run the MCP server (normally spawned by the agent, but useful to verify):

tma1-server mcp-serve

Use the skills and tools from inside the agent:

/tma1-search retry backoff
/tma1-search --session 3d8f1d0a
/tma1-peer self --limit 3

Key environment variables:

TMA1_PORT=14318
export TMA1_DATA_DIR=~/.tma1
export TMA1_DATA_TTL=60d
export TMA1_LLM_API_KEY=...   # only for optional prompt evaluation
export TMA1_LLM_PROVIDER=anthropic
export TMA1_DISABLE_INJECTION=1  # record hooks but inject nothing

Raw SQL is available on MySQL port 14002 or through GreptimeDB's dashboard.

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

1. v0.3.0-beta4 Pre-release 2026-10-30
   Release notes excerpt

   What's Changed * feat(pricing): price cache reads, cache writes and OpenAI Fast mode by @killme2008 in https://github.com/tma1-ai/tma1/pull/113 Full Changelog: https://github.com/tma1-ai/tma1/compare/v0.3.0-beta3...v0.3.0-beta4

   View on GitHub ↗

FAQ