worklog

Personal time-tracker for the developer who hates timers. Pulls activity from Claude Code, GitHub, Google Calendar, and Jira into one local SQLite file, clusters it into blocks, uses claude -p to pick a ticket and write a description for each block, and syncs the result to Tempo Cloud after you review it in a local web UI.

Everything runs on your machine. Nothing leaves until you push a Tempo sync — and you review every block before that happens.

Install

curl -fsSL https://raw.githubusercontent.com/TomasPalsson/worklog/main/install.sh | bash

The script downloads a signed release binary for your platform (macOS arm64 or linux x86_64) from GitHub Releases, verifies it, and drops it at ~/.local/bin/worklog. After that:

worklog setup          # one-shot onboarding: db + secrets + Claude hook
worklog day            # end-of-day: collect → infer → estimate → review UI

Upgrading is a single command and always verifies the signature:

worklog upgrade        # signed self-update via the release pipeline

Migrating from the old uv tool install? See docs/MIGRATION.md — one step.

Architecture

┌──────────────────────────── your Mac ─────────────────────────────┐
│                                                                    │
│   ~/.local/share/worklog/worklog.db   (single source of truth,     │
│                                        WAL mode — safe RO readers  │
│                                        during concurrent writes)   │
│             ▲                     ▲                                │
│             │                     │                                │
│     writes (unix socket           │ reads direct                   │
│      OR TCP 127.0.0.1:9323)       │ via bun:sqlite                 │
│             │                     │                                │
│    ┌────────┴────────┐        ┌───┴───────────────────┐            │
│    │ worklog (Rust)  │        │ Docker: worklog-web   │            │
│    │  · CLI          │        │  · Bun + Next.js 15   │            │
│    │  · Collectors   │        │  · Server Components  │            │
│    │  · Estimator    │───────▶│  · Server Actions →   │            │
│    │  · axum daemon  │  TCP   │     host.docker       │            │
│    │  · web orch.    │        │     .internal:9323    │            │
│    └────────┬────────┘        └───────────────────────┘            │
│             │  spawns/manages                                       │
│             ▼                                                       │
└─────────────────────────────────────────────────────────────────────┘

Events are classified to a Jira ticket by the AI estimator; unmatched ones appear as "unassigned" in the web UI and are reassigned by hand. The web container reads SQLite directly (WAL mode = concurrent reads are safe), and writes flow through Next.js Server Actions → the Rust daemon over TCP (Docker Desktop can't proxy live unix sockets through its macOS VM, hence TCP between the container and host).

Only the Rust daemon writes to the DB — Server Actions are a thin shim that forwards the mutation and calls revalidatePath.

Quickstart

worklog setup          # preflight + interactive secrets + db migrate
worklog hook install   # register the Claude Code hook
worklog day            # full daily pipeline (collect → infer → estimate → UI)

To pull everything without running the UI:

worklog day --no-serve

Commands

| Command | Purpose | |---|---| | worklog version | Print the embedded version | | worklog setup | Preflight + secrets capture + db migrate | | worklog doctor | Environment, DB, secrets sanity report | | worklog day [--day] [--no-serve] [--model] | Full daily pipeline: collect → infer → estimate → web UI | | worklog collect [all\|jira\|github\|gcal] [--days] | Pull remote activity | | worklog infer [--day] | Gap-timeout clustering of events into blocks | | worklog estimate [--day] [--model] | claude -p fills jira + description + minutes | | worklog sync [--day] [--dry-run] | POST reviewed blocks to Tempo Cloud | | worklog web up [--port] | Bring up the dockerised Next.js UI | | worklog web down / status / logs / build | Container lifecycle | | worklog daemon [--socket] [--tcp] | Axum API server (unix + TCP) | | worklog hook [install\|uninstall\|status] | Claude Code hook wiring | | worklog schedule [install\|uninstall\|status] | Scheduled collection (launchd / systemd --user) | | worklog secret [set\|get\|rm\|list] | Credentials in the OS keychain | | worklog db [migrate\|info\|path] | DB operations | | worklog upgrade | Signed self-update | | worklog self-update [--check\|--dry-run\|--force] | Lower-level alias for upgrade | | worklog dev [keygen\|sign\|make-patch\|apply-patch] | Maintainer tooling |

Every subcommand accepts --json for structured output.

Storage

• DB: ~/.local/share/worklog/worklog.db (SQLite, WAL mode).
• Config: ~/.config/worklog/ (.env, google_credentials.json, google_token.json).
• Secrets: OS keychain under service name worklog (macOS Keychain, Linux secret-service, Windows Credential Manager). The .env file still works — secrets migrate lazily as you set them via worklog secret set … or re-run worklog setup.
• Binaries + releases: ~/.local/share/worklog/{bin,releases}/.

Override everything with $WORKLOG_HOME=<dir> — collapses db, socket, config, bin, logs, releases into one root. Primarily for tests and power users.

Switching estimator provider

The estimator has two interchangeable backends:

• claude_subprocess (default) — shells out to claude -p. Zero configuration. Requires the Claude Code CLI on PATH.
• litellm — POSTs to any OpenAI-compatible proxy (LiteLLM is the reference implementation). Lets you route to Anthropic, OpenAI, local Ollama, Bedrock, etc. without changing worklog.

Select via worklog setup (interactive) or set the env var:

export WORKLOG_ESTIMATOR_PROVIDER=litellm
worklog secret set litellm_base_url http://localhost:4000
worklog secret set litellm_api_key  <your-proxy-key>   # empty OK for unauthed local
worklog secret set litellm_model    anthropic/claude-haiku-4-5

worklog doctor reports the active provider and a reachability probe. Unset WORKLOG_ESTIMATOR_PROVIDER (and run worklog secret rm worklog_estimator_provider) to fall back to the subprocess path.

Dev

cargo test  --manifest-path rust/Cargo.toml
cargo clippy --manifest-path rust/Cargo.toml --all-targets --all-features -- -D warnings
cargo fmt   --manifest-path rust/Cargo.toml --all

cd web && bun test && bun run typecheck && bun run build

Local release smoke (no network, no tag push):

bash scripts/release-smoke.sh
bash tests/install/smoke.sh

CI: .github/workflows/rust.yml runs fmt + clippy + tests on Linux and macOS on every push / PR; .github/workflows/release.yml builds, signs, and publishes on every v* tag.

Files of interest

• rust/crates/worklog-core/ — data layer: paths, db, repo, secrets, collectors, infer, estimator, signed updater
• rust/crates/worklog-cli/ — the worklog binary: CLI + setup wizard
• rust/crates/worklog-core/sql/schema.sql — canonical SQL schema
• web/ — the Next.js + Bun review UI (dockerised)
• install.sh — curl-piped installer
• scripts/release-smoke.sh — host-side dry-run of the release pipeline