AI agent observability & security

Every action leaves a trace.

See what your AI agents actually do — every tool call, file change and command, as it happens.

Tracemason lets engineering and security teams observe agent execution, investigate code changes, catch risky operations, and reconstruct what happened afterwards.

One session, as recorded

  1. session.started14:02:07claude-code · payments-api
  2. tool.completed14:02:19Edit · src/ledger/refund.ts
  3. tool.completed14:02:31Bash · package.test
  4. tool.completed14:02:44Bash · vcs.commit a41c9e7
  5. tool.requested14:03:02Bash · database.destructive

Finding opened: destructive database command ▲ HIGH

Requested against production. Evidence: the event above.

Illustrative · the product's views are drawn from your data

Scroll to follow one session, step by step

The trace

Make the invisible observable.

From tool calls to code changes, Tracemason records what an agent actually did — as ordered events you can read, search and export.

  • 01

    Agents act with real permissions.

    They run commands, edit files and reach infrastructure using the developer's own access. What they did should be a record, not a recollection.

  • 02

    A chat transcript is not an audit trail.

    Transcripts live on one laptop and mix intent with outcome. Tracemason records the actions themselves: tool, target, timing, result.

  • 03

    Observability, not surveillance.

    Collection is metadata-first and configured in the open. Tracemason watches agent activity — not keystrokes, screens or people.

sess_7f3a91c2·3m 12s
Illustrative
  1. +00:00
    session.started claude-code · acme/payments-api · branch fix/refund-rounding
  2. +00:12
    tool.completed Read · src/ledger/refund.ts
    ok
  3. +00:31
    subagent.started Explore · search call sites
  4. +00:33
    tool.completed Grep · 14 files matched
    ok
  5. +00:52
    subagent.stopped Explore · 3 tool calls
  6. +01:04
    tool.completed Edit · src/ledger/refund.ts
    ok
  7. +01:16
    tool.failed Bash · package.test · exit 1
    failed
  8. +02:40
    tool.completed Bash · package.test · exit 0
    ok
  9. +02:58
    tool.requested Bash · database.destructive
    flagged
  10. +03:12
    session.ended 9 tool calls · 1 failure · 1 finding

Investigations

From uncertainty to evidence.

Understand what happened, where it happened, and what to investigate next. Start from an alert, a commit or an agent and walk the chain — every hop states why two records are related, and how sure we are.

investigation / refund-table-dropIllustrative
  1. Finding

    Destructive database command

    high · open

  2. Event

    tool.requested · Bash

    14:03:02.415

    Verified

    finding references this event id

  3. Session

    sess_7f3a91c2

    claude-code · 3m 12s

    Verified

    event carries the session id

  4. Commit

    a41c9e7 Fix refund rounding

    fix/refund-rounding

    Verified

    collector observed this SHA created in-session

  5. Pull request

    #482 Fix refund rounding

    checks passing

    Verified

    GitHub reports the commit in this PR

  6. Deployment

    production

    17 min later

    Correlated

    deployment SHA descends from the PR merge

Verified
Both sides report the same identifier.
Explicitly linked
A person or tool declared the link, e.g. a commit trailer.
Correlated
Multiple independent signals agree. No shared identifier.
Possible association
Same repository and time window only. Treat as a lead.
Unknown
No supporting evidence. Shown so gaps are visible.

How it works

From action to evidence.

One tool call, followed end to end. Each stage adds structure — none of them needs your source code.

  1. Action1 / 4
    hook PreToolUse
    tool_name Bash
    tool_input.command
      ▒▒▒▒ ▒▒▒▒▒▒▒▒▒▒▒▒▒ ▒▒ ▒▒▒▒▒
      stays on the machine

    An agent requests a tool

    Claude Code fires a lifecycle hook. The collector reads it locally and classifies the command without uploading it.

  2. Trace2 / 4
    type tool.requested
    tool.name Bash
    tool.category database.destructive
    source claude-code.hook
    schema_version 1.0

    It becomes a structured event

    A versioned, provider-neutral record with separate occurrence and ingestion timestamps and its source preserved.

  3. Risk3 / 4
    Destructive database command
    policy
    destructive-commands
    severity
    ▲ high
    status
    open

    Policies evaluate it

    Deterministic rules run on every event. A match opens a finding with the evidence and next steps attached.

  4. Evidence4 / 4
    Fix refund rounding#482
    Verifiedcommit observed in session

    Git activity is correlated

    Commits, pull requests and deployments attach to the session — each link labelled with how strong the evidence is.

Security by design

Built to be trusted with what it sees.

A platform that watches development environments has to earn its place there. These are properties of the system, not settings you have to find.

  • Metadata-first ingestion

    By default the collector sends tool names, categories, paths, timing and outcomes. Not prompts, not file contents, not terminal output.

  • Redaction before upload

    Fields are allowlisted, then scanned for keys, tokens and credentials on the developer's machine — and again on the server.

  • Encrypted transport

    The collector and SDKs refuse non-TLS endpoints outside localhost. Ingestion credentials are stored hashed and shown once.

  • Organization isolation

    Every query is scoped to the organization resolved from the session or credential on the server — never from a client-supplied ID.

  • Configurable retention

    Choose how long events live. A scheduled worker deletes what's past the window and records that it did.

  • Fine-grained access

    Owner, Admin, Security Analyst, Developer and Viewer roles, enforced server-side, with an audit log of sensitive changes.

Monitoring is not enforcement

Tracemason can ask Claude Code to deny a tool call through its documented hook mechanism, and reports a block only when the hook confirms it. Hooks run on a developer-controlled machine and are not a tamper-proof boundary — keep critical controls in repository permissions, sandboxes, CI and infrastructure authorization.

What we don't claim

Tracemason is pre-release software and holds no security certifications. Encryption at rest depends on how you deploy the database. The documentation states exactly what is collected at each collection level so teams can tell their developers.

Pricing

Simple, predictable pricing.

Start free with Claude Code. Upgrade when you need more events, longer retention and Git correlation.

  • Developer

    For individuals instrumenting their own agents.

    Free

    Start tracing
    • 50k events / month
    • 7-day retention
    • Claude Code integration
    • Basic traces and built-in policies
  • Most teams

    Team

    For engineering teams adopting agents together.

    $49starting price, per month

    Start with Team
    • 1M events / month
    • 30-day retention
    • Git activity for up to 15 repositories
    • Policy alerts by email and webhook
    • GitHub correlation
    • Team roles and invitations
  • Business

    For organizations with security and audit requirements.

    $199starting price, per month

    Start with Business
    • 10M events / month
    • 180-day retention
    • Custom policy rules
    • Audit and trace exports
    • Security Analyst and Viewer roles
  • Enterprise

    For regulated and large-scale deployments.

    Custom

    Contact us
    • Custom retention
    • Custom ingestion limits
    • Everything in Business
    • SSO / SAML · planned
    • Private deployment options · planned
    • Enterprise support SLAs · planned

Prices are starting prices in USD. Limits are enforced as listed; items marked planned are not yet available.

You see what your agents produce. Tracemason shows what they did.

See beyond the output.

Connect Claude Code in a few minutes and watch your first session arrive, event by event.