textlog

A macOS clipboard + screenshot capture daemon that exposes its archive to Claude Code as a Model Context Protocol (MCP) server, with on-device OCR via Apple Vision Framework.

textlog runs quietly in the background, watches your clipboard and screenshots, OCRs images with the same engine the macOS system uses, and makes everything searchable by Claude Code through seven textlog__* MCP tools. Nothing leaves your machine. There is no LLM inside textlog — Claude Code is the LLM, textlog is its eyes on your clipboard.

Why this exists

Pasting context into Claude Code one selection at a time is the default, and it works. But it has three failure modes:

You paste the wrong slice. A 400-line stack trace gets trimmed to the boring middle, the actual panic line is in the 20 lines you missed.
You can't paste at all. The bug is in a screenshot from your designer, a render of a PDF, or an image in a Notion doc. Claude can read images, but you spend tokens uploading the raw PNG when you only needed the text.
You forgot what you copied. You know you saw the right error message ten minutes ago. It was on your clipboard. Now it's not.

textlog fixes all three. Every clipboard transition lands in a SQLite ring buffer (1000 captures by default) and a permanent daily Markdown file. Images are OCR'd at capture time, so by the time you ask Claude "what was in that screenshot?" the answer is already a plain string in the index — no image upload, no token bloat.

It is built specifically for the Claude Code workflow: the daemon's job is to be a high-recall, low-latency clipboard backend for an LLM that lives next door, not a standalone clipboard manager with its own UI. The MCP server is the entire user-facing surface.

Features

MCP server with 7 tools — textlog__get_recent, __search, __list_today, __get_capture, __ocr_latest, __ocr_image, __clear_since. Registers in one command: claude mcp add --scope user textlog -- tl mcp.
Preview-by-default responses (v0.1.7+). List-style tools (get_recent, list_today, search) return a 200-char text_preview plus a truncated: bool flag instead of the full body. A 5-row get_recent against large copies dropped from ~10 k to ~250 tokens. Claude calls textlog__get_capture(id) only for rows it actually needs to expand.
On-device OCR via VNRecognizeTextRequest (Apple Vision). No cloud call, no API key, no rate limit. Honours recognition_level ("accurate" | "fast") and a configurable language list.
SQLite + FTS5 ring buffer. Bounded full-text index (default 1000 captures), automatically trimmed. The daily Markdown archive is never trimmed — the SQL index is for query speed; the MD files are the durable record.
SHA-256 dedup at query time. get_recent and list_today collapse duplicates by content hash; search returns every match but marks later occurrences with duplicate_of, so Claude sees each unique snippet exactly once per call.
md_path on every result. Every capture row in every MCP response carries the absolute path of the daily Markdown file it was mirrored into, so Claude can Read the whole day's context as a single file attachment without any clipboard round-trip.
Privacy filter. A RegexSet is compiled once at startup from monitoring.ignore_patterns. Defaults catch OpenAI-style API keys, *_KEY= env-var assignments, 16-digit credit-card numbers, and password = … lines. Hits are dropped before they reach storage and optionally surface a discreet macOS notification.
Self-write skip. When the daemon writes the daily-MD path back to your clipboard (copy_log_path_on_complete), it publishes the resulting NSPasteboard.changeCount into a shared atomic so the next poll skips its own write — no infinite recursion.
LaunchAgent integration. tl install writes a plist and runs launchctl bootstrap gui/$UID so the daemon survives reboots. tl uninstall does the inverse.
tl doctor runs eight health checks across config, storage, permissions, LaunchAgent state, MCP registration, and a live Vision smoke test, then exits non-zero if anything is broken.

Release history

See CHANGELOG.md. Latest version is always the one shown in the crates.io badge above.

Quickstart

# Build and install (requires Rust 1.80+ and macOS).
cargo install textlog                   # from crates.io
# or
cargo install --path .                  # from a local clone

# First-run health check — should print 4 PASS, 4 WARN, 0 FAIL.
tl doctor

# Optional: register as a LaunchAgent so it auto-starts on login.
tl install

# Register as an MCP server in Claude Code (user scope recommended —
# see "Registration scope" below for why).
claude mcp add --scope user textlog -- tl mcp

# Confirm.
claude mcp list

That's it. Open any Claude Code session and ask "what's in my recent clipboard?" — Claude calls textlog__get_recent and reads the answer.

If you'd rather run only when you ask, skip tl install and run tl start --foreground in a terminal tab — the daemon prints to stdout and stops on Ctrl-C.

Registration scope

Claude Code's claude mcp add supports three scopes, picked with --scope <local|user|project>. For textlog — a single-user clipboard daemon with a per-user MD archive — user scope is almost always the right choice:

# User scope (recommended).
# Stored in ~/.claude.json. Loads in every Claude Code session you
# start on this machine, from any directory. Best fit for textlog.
claude mcp add --scope user textlog -- tl mcp

# Local scope (the default when --scope is omitted).
# Stored per-project in ~/.claude.json under the current working dir.
# textlog only appears in sessions launched from this exact directory,
# which is almost never what you want for a personal clipboard tool.
claude mcp add textlog -- tl mcp

# Project scope.
# Writes a `.mcp.json` at the repo root, intended to be committed so
# every collaborator auto-registers the same servers. Not a great fit
# for textlog: every teammate would still need `tl` installed locally,
# the MD archive is strictly per-user, and the daemon has to be running
# on each machine.
claude mcp add --scope project textlog -- tl mcp

Switching scope means removing the existing registration first:

claude mcp list                         # shows which scope textlog is in
claude mcp remove textlog               # drops it from whichever scope
claude mcp add --scope user textlog -- tl mcp

If you previously followed an older version of this README (no --scope flag) and then moved to a different project directory, your MCP registration seems to "disappear" — that's local scope working as designed. Re-add with --scope user and it follows you everywhere.

Real scenarios

1. The "what just blew up" loop

Terminal spits out a 40-line stack trace. Instead of selecting carefully, Cmd+A; Cmd+C the whole pane and ask Claude:

"Look at my last clipboard entry — what's the root cause and what file should I open?"

Claude calls textlog__get_recent { n: 1 } and points you at the line.

2. Screenshot a UI bug, get a CSS fix back

A designer slacks you a screenshot of a misaligned button. Cmd+Shift+Ctrl+4 to copy the screenshot to clipboard, then:

"OCR the last image I copied and suggest CSS fixes for what it shows."

textlog__ocr_latest returns the OCR'd text from the screenshot — no image upload, no token waste.

3. Compiler error from a different terminal tab

A 300-line Rust error spans two screens. You don't want to scroll-select.

"Search my clipboard log for the most recent 'cannot find function' error and explain it."

textlog__search { query: "cannot find function", limit: 1 }.

4. Recovering yesterday's API response

Postman returned a JSON blob you reasoned about, then you closed the tab.

"Find that Stripe webhook payload from yesterday."

textlog__search { query: "evt_", since: "2026-04-16T00:00:00Z" }.

5. Dense docs you OCR'd from a PDF

PDFs with broken text selection are a known pain. Screenshot a code block from one and:

"OCR this and convert the Swift to Rust."

6. "Why did this work yesterday?"

You're debugging a regression. You copied the working version of a function six hours ago.

"Search clipboard for fn calculate_total, show the most recent three unique versions."

FTS5 search + sha256 dedup means Claude only sees distinct revisions — not 40 copies of the same paste.

7. Pre-commit context dump

You've been scribbling into your clipboard for an hour: error messages, command outputs, dbg! results.

"Summarise my clipboard from the last hour into a draft commit message."

textlog__list_today hands Claude the entire investigation trail.

8. Privacy-respecting paste

You accidentally Cmd+C'd OPENAI_API_KEY=sk-proj-…. The default filter dropped it before storage. A discreet notification confirms: "textlog dropped a sensitive clipboard entry". The key was never written to disk.

9. Cross-session memory

Three days later, after a reboot:

"Did I look at the Stripe webhook docs this week?"

textlog__search { query: "webhook", since: "2026-04-14T00:00:00Z" }.

10. Daily-archive paste trick (v0.1.1+ — no clipboard side-effect)

Every capture row in every MCP response now carries md_path. So:

"Read the daily clipboard log and summarise my morning."

Claude calls textlog__get_recent { n: 1 }, sees md_path: "/Users/you/textlog/logs/2026-04-17.md" in the response, then calls its built-in Read tool on that path. Claude receives the entire day's chronological transcript as a single attachment.

In v0.1.0 the same workflow needed notifications.copy_log_path_on_complete, which copied the path to the clipboard after each capture and could cascade with some clipboard managers. v0.1.1 made that side-effect optional; v0.1.6 turns it off by default — Claude still finds the file via md_path in the MCP response.

11. Code review prep

Reviewing a PR? Click through every changed file in your IDE, copy the unified diff for each, then:

"For each clipboard entry from the last 10 minutes that looks like a diff, list the function-level changes."

12. Slack/Linear ticket triage

Copy the full ticket body from Linear, copy the relevant log line from your terminal, then:

"Cross-reference my last two clipboard entries and tell me whether the log proves the bug from the ticket."

13. Comparing two API responses

Copy yesterday's response, copy today's response, then:

"What changed between my last two clipboard entries (both JSON)?"

textlog__get_recent { n: 2 } → Claude diffs two JSON blobs without you opening a diff tool.

14. "Save this for later" without manual notes

You're in the middle of fixing bug A and notice symptom B. Cmd+C the relevant piece of evidence for B and keep working on A. Tomorrow:

"Find clipboard entries from yesterday that look like error messages I haven't followed up on."

15. AI coding pair-prog feedback loop

Claude writes a function. You copy the test failure. Claude reads textlog__get_recent, sees its own previously-suggested code in context with the new failure, iterates without you re-pasting.

Architecture

                    +-------------------+
                    |   Claude Code     |
                    +---------+---------+
                              | stdio JSON-RPC 2.0
                              v
+--------------------------------------------------------+
|                     tl mcp (MCP server)                |
|   textlog__get_recent / __search / __list_today        |
|   textlog__ocr_latest / __ocr_image / __clear_since    |
+----------------------+---------------------------------+
                       |
                       v
        +--------------+----------------+
        |   src/storage/sqlite.rs       |   <-- SQLite + FTS5 ring buffer
        |   src/storage/markdown.rs     |   <-- daily MD archive (durable)
        +---------------+---------------+
                        ^
                        |  insert(CaptureRow)
                        |
+-----------------------+--------------------------+
|             tl start (pipeline)                  |
|                                                  |
|  clipboard::poll_once  -->  PrivacyFilter        |
|        ^                          |              |
|        |                          v              |
|  NSPasteboard      ocr::ocr_image (Vision)       |
|        ^                          |              |
|        |                          v              |
|        +----------- notifier::notify_complete    |
|                     (+ optional clipboard write- |
|                      back of the daily MD path)  |
+--------------------------------------------------+

Two long-running tasks joined via tokio::select!:

Monitor loop — polls NSPasteboard.changeCount directly on the async task (it's a microsecond i64 read, no spawn_blocking needed). Only falls through to clipboard::poll_once (FFI content read) via spawn_blocking when the counter actually advances. Exponential idle backoff: 500 ms active, doubling to a 2 s ceiling after 20 unchanged ticks; any real change snaps back to active. Pushes ClipboardEvent into a bounded mpsc(16) channel with drop-on-full backpressure.
Consumer — drains the channel, runs filter → OCR → SHA-256 → Storage::insert → notifier, all SQLite work spawn_blocking'd so concurrent MCP queries are never blocked by a slow disk.

Why these specific choices

Polling, not event-driven. macOS NSPasteboard has no public notification API for change events. The default 500 ms interval is perceptually instant; combined with exponential idle backoff (up to 2 s) and a direct changeCount fast-path (no spawn_blocking when unchanged), CPU at idle is effectively zero — only a content read happens when the counter actually advances.
SQLite + FTS5, not a fancier search engine. FTS5 is built into SQLite, supports MATCH 'foo' queries with prefix wildcards, and ships zero extra binaries. Latency for a single-keyword search over 10k rows is sub-millisecond on Apple Silicon.
Ring buffer + permanent MD archive. The bounded SQL index is what makes tl logs search fast forever; the MD archive is what makes "what did I copy three weeks ago" still answerable. Neither alone is enough.
Apple Vision, not Tesseract or a cloud OCR. Vision ships with macOS, requires no setup, supports 30+ languages, and runs on the Neural Engine when available. A typical screenshot OCRs in <100 ms with recognition_level = "accurate".
rmcp 1.4 stdio. Stdio JSON-RPC is the only MCP transport Claude Code's claude mcp add supports out of the box. No HTTP server, no port collisions.

CLI reference

tl mcp                          Run MCP stdio server (Claude Code spawns this)
tl version | -v | --version     Print version (subcommand or short flag)
tl update                       Self-update via `cargo install textlog --force`
tl config show                  Print effective config as TOML
tl config path                  Print config file path
tl config reset                 Overwrite config with v2.0 defaults

tl logs today                   List today's captures (one line each)
tl logs search <QUERY> [--limit N]   FTS5 search; canonical row only
tl logs path                    Print log directory

tl doctor                       Run 8 health checks; non-zero exit on FAIL
tl perf [--duration SECS]       Sample daemon CPU/RSS via `ps`
       [--interval-ms MS]        (default 10 s @ 1 s)

tl install                      Install LaunchAgent + bootstrap
tl uninstall                    Bootout + remove plist
tl start                        launchctl kickstart (background)
tl start --foreground           Run pipeline inline (Ctrl-C to stop)
tl stop                         launchctl kill SIGTERM
tl status                       Loaded? PID? Last exit?

Global flag: --config-dir <PATH> (env: TEXTLOG_CONFIG_DIR) overrides ~/textlog/.

MCP tools

Tool	Args	Returns	Purpose
`textlog__get_recent`	`n` (default 5), `kind?` (`text`\|`image`\|`any`)	`{ captures: [{ id, ts, kind, sha256, size_bytes, text_preview, truncated, md_path, source_app?, source_url?, ocr_confidence? }] }`	Latest N captures, deduped by sha256
`textlog__list_today`	`kind?`	same shape as `get_recent`	Everything from today (UTC)
`textlog__search`	`query`, `limit` (default 20), `since?` ISO 8601	`{ hits: [{ capture, duplicate_of? }] }`	FTS5 search; later occurrences marked
`textlog__get_capture`	`id` (from any summary above), `offset?` (default 0), `limit?` (default 8000 chars, max 32000)	`{ id, ts, kind, sha256, size_bytes, text, text_offset, text_total_chars, truncated, md_path, source_app?, source_url?, ocr_confidence? }`	Windowed slice of a single capture body; page until `truncated == false`
`textlog__ocr_latest`	none	`{ text, confidence, captured_at }`	Cached OCR text from the last image
`textlog__ocr_image`	`path` (absolute)	`{ text, confidence, block_count }`	Ad-hoc OCR of any file
`textlog__clear_since`	`ts` ISO 8601	`{ deleted_count }`	Privacy cut-off (SQLite only; MD untouched)

All input/output schemas are published via MCP tools/list.

Two-tier body retrieval (v0.1.7+). The list-style tools (get_recent, list_today, search) return a text_preview capped at 200 characters plus a truncated: bool flag. When truncated is true, Claude fetches the body via textlog__get_capture({ id }). This keeps a default 5-row get_recent under ~250 tokens regardless of how large each underlying copy was; previously a single big paste could blow the response past 10 k tokens. size_bytes on every summary tells Claude how much it's missing so it can decide whether to expand.

Paginated body fetch (v0.1.8+). textlog__get_capture now slices the body to a configurable window so even a 50 k-char build log can be retrieved without overflowing the MCP per-tool token budget. The default limit is 8000 chars (≈2000 tokens), capped server-side at 32000 chars (≈8000 tokens). The response carries text_offset, text_total_chars, and truncated — page forward with offset = text_offset + text.chars().count() until truncated is false. Bodies smaller than the default window are returned in one shot exactly as before.

md_path was added in v0.1.1 — it's the absolute path of the daily Markdown file the row was mirrored into, so Claude can Read the entire day's context without any clipboard round-trip. This is also the recovery path for captures older than storage.ring_buffer_size (default 1000): SQLite trims them, but the daily MD on disk doesn't.

Configuration

Default location: ~/textlog/config.toml. Override with TEXTLOG_CONFIG_DIR.

Reset to defaults at any time:

tl config reset
$EDITOR $(tl config path)

Keys you'll most likely want to touch:

key	default	why touch it
`monitoring.poll_interval_ms`	`500`	Active-rate ceiling; idle backoff slows to a 2 s cap automatically. Lower = snappier; higher = even less CPU
`monitoring.min_length`	`10`	Drop tiny copies (tab-switching noise). Lower to 1 if you want every clipboard transition recorded
`monitoring.ignore_patterns`	API keys, CC numbers, passwords	Add your own regexes — anything matched is silently dropped
`notifications.copy_log_path_on_complete`	`false`	Write the daily MD path back to the clipboard after each capture. Off by default since v0.1.6 — Claude finds the path via `md_path` in MCP responses (v0.1.1+), so most users don't need it
`notifications.on_capture`	`false`	A toast for every clipboard event — usually too noisy
`storage.ring_buffer_size`	`1000`	SQLite cap; MD archive is never trimmed
`storage.log_dir`	`~/textlog/logs`	Where daily MD files live
`ocr.recognition_level`	`"accurate"`	`"fast"` if Vision is too slow on your hardware
`ocr.languages`	`["en-US"]`	Add more — Vision's polyglot is excellent
`ocr.min_confidence`	`0.4`	Lower to keep blurrier OCR text; higher to drop low-confidence results

After editing: tl uninstall && tl install so the LaunchAgent re-reads the (cached) program args. (For tl start --foreground, just Ctrl-C and re-run.)

Tuning recipes

Minimal-noise mode

For users who only want substantive captures and no clipboard side-effects.

[monitoring]
min_length = 50                     # ignore tab-switch noise

[notifications]
enabled = false                     # no toasts at all
copy_log_path_on_complete = false   # no clipboard write-back

Maximum-recall mode

For users who want every clipboard transition recorded, even single chars.

[monitoring]
min_length = 1
poll_interval_ms = 100              # 4× faster polling

[storage]
ring_buffer_size = 10000            # ~10× the default; SQLite handles this fine

Multilingual capture

Vision supports 30+ languages. List them in priority order — earlier languages bias the recogniser when text is ambiguous.

[ocr]
languages = ["en-US", "ja-JP", "zh-Hans", "fr-FR", "de-DE"]
recognition_level = "accurate"

Strict privacy

Goes beyond the defaults — drops anything that looks remotely like a secret or a finance number.

[monitoring]
ignore_patterns = [
    "^sk-[A-Za-z0-9]{20,}",
    "^\\w+_KEY\\s*=",
    "^\\w+_TOKEN\\s*=",
    "^\\w+_SECRET\\s*=",
    "Bearer\\s+[A-Za-z0-9._-]{20,}",
    "\\b\\d{4}[- ]?\\d{4}[- ]?\\d{4}[- ]?\\d{4}\\b",   # credit cards
    "\\b\\d{3}[- ]?\\d{2}[- ]?\\d{4}\\b",                # US SSN
    "(?i)password\\s*[=:]\\s*\\S+",
    "-----BEGIN [A-Z ]*PRIVATE KEY-----",
]

[privacy]
filter_enabled = true
log_sensitive = false
show_filter_notification = true     # always notify on a hit

Privacy

Nothing leaves your machine. No telemetry, no auto-updates, no remote calls of any kind. The only outbound process is claude mcp spawning tl mcp as a local subprocess.
Privacy filter drops sensitive content (API keys, credit cards, passwords) before it reaches storage. See monitoring.ignore_patterns for the default regex set; add your own. privacy.show_filter_notification toggles the discreet macOS notification on filter hits.
textlog__clear_since lets you (or Claude) wipe SQLite rows at or after a timestamp. The Markdown archive is intentionally untouched — delete those files yourself if you want them gone.
Config file mode 0600 is checked by tl doctor; warn-level if loose. The default config contains no secrets but you may add API keys (e.g. for future provider-aware features) so the warning stays.

macOS permissions

On first run, macOS will prompt for:

Pasteboard access (since macOS 15.4) — required.
Notifications — optional, only if you have any notifications.* flag set to true.

tl doctor reports the state of both. Pasteboard denial is a hard fail; notifications denial is informational.

FAQ / troubleshooting

`tl doctor` first, every time

tl doctor                          # always start here
cat ~/textlog/logs/stdout.log      # LaunchAgent stdout
cat ~/textlog/logs/stderr.log      # LaunchAgent stderr

Why isn't `"hi?"` showing up in `tl logs today`?

Default monitoring.min_length = 10. Anything shorter is dropped before storage. Either accept it as noise filtering, or:

sed -i '' 's/^min_length = .*/min_length = 1/' ~/textlog/config.toml

then restart the daemon.

Why do I see the same content multiple times in the MD file?

Storage::insert records every clipboard transition. If your clipboard manager (Raycast / Maccy / Paste / Alfred) re-asserts content, your terminal has copy-on-select enabled, or you re-copy the same text by hand, each write bumps NSPasteboard.changeCount and the daemon sees a new event.

These duplicates are never visible to Claude: query-time dedup collapses them by sha256. The MD file is the durable per-event audit trail; the SQL queries are what Claude actually consumes.

To reduce the MD-file duplication directly:

notifications.copy_log_path_on_complete is off by default from v0.1.6 (it's no longer needed for path discovery). Leave it off unless you explicitly want the path written back to your clipboard.
Disable copy-on-select in your terminal app's preferences.

Claude can't find my MCP server

claude mcp list                    # textlog should show ✓ Connected

If ✗ Failed to connect:

# Hand-drive an initialize request to see the actual error:
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"manual","version":"0"}}}' | tl mcp

If you see a JSON response, the server is fine — the issue is in Claude Code's registration. Try claude mcp remove textlog then re-add.

My screenshot OCR returns no text

Check ocr.min_confidence (default 0.4). On low-resolution screenshots or blurry text, Vision's confidence may sit below the threshold. Lower it to 0.2 to be more permissive, or set ocr.recognition_level = "accurate" if it's currently "fast".

The daemon stopped after I edited config

LaunchAgent caches ProgramArguments, not the config-file path or contents. The config is re-read on each tl start, so:

tl uninstall && tl install        # rewrites the plist + restarts cleanly

For tl start --foreground, just Ctrl-C and re-run.

After upgrading the binary (`cargo install textlog --force`)

tl stop then tl start usually works — KeepAlive respawns the agent and picks up the new binary on disk. If launchd trips its throttle (minimum runtime = 10) because of rapid respawns, you'll see tl status stuck on pid=None and launchctl print will report state = spawn scheduled / job state = spawn failed. Fix it with a full re-bootstrap:

tl uninstall && tl install        # clears the throttle, re-bootstraps

Nuclear reset

tl uninstall
rm -rf ~/textlog                  # config + db + MD archive all gone
tl doctor                         # confirms clean slate

Pasteboard permission denied

System Settings → Privacy & Security → Pasteboard → enable tl.

Notifications muted

System Settings → Notifications → tl → enable. (Or set notifications.enabled = false to never bother.)

Comparison with alternatives

	textlog	Raycast Clipboard History	Maccy / Paste	Rewind.ai
Background clipboard capture	yes	yes	yes	yes
Searchable history	yes (FTS5)	yes (UI only)	yes (UI only)	yes
Image OCR at capture time	yes (Apple Vision)	no	no	yes (cloud + local mix)
Exposes to LLM via MCP	yes	no	no	no
Local-only (zero network)	yes	yes	yes	no (cloud component)
Per-day Markdown archive	yes	no	no	no
Privacy regex filter at capture	yes	no	partial	no
Open source	yes (MIT)	proprietary	proprietary	proprietary
Built specifically for Claude Code	yes	no	no	no
Standalone clipboard UI	no (intentional)	yes	yes	yes

textlog is not a clipboard manager. It's an LLM-facing clipboard backend. Use Raycast or Maccy for browsing your history visually; use textlog for letting Claude reason over it.

Development

cargo test --bin tl                # 168 tests pass
cargo test --bin tl -- --ignored   # +5 live tests (NSPasteboard, Vision, notify-rust)
cargo clippy --bin tl --all-features --tests

Source layout:

src/
  main.rs               Entry point — clap parse + dispatch
  cli/
    args.rs             clap derive structs
    commands.rs         Per-command handlers
    doctor.rs           tl doctor health checks
    perf.rs             tl perf CPU/RSS sampling
  config/               TOML schema + load/save + env overlay
  error.rs              Top-level Error enum + Result alias
  filters.rs            Privacy filter (RegexSet)
  storage/
    mod.rs              Kind, CaptureRow, SearchHit, hex helpers
    sqlite.rs           rusqlite + FTS5 ring buffer
    markdown.rs         Daily MD archive writer
  ocr.rs                Apple Vision wrapper (objc2-vision)
  clipboard.rs          NSPasteboard read/write + ClipboardWriter trait
  notifier.rs           notify-rust + Notifier trait + CountingNotifier
  pipeline.rs           Pipeline::process_event + run loop
  mcp/
    mod.rs              run_stdio entry point
    tools.rs            McpServer + 7 #[tool] handlers
    schema.rs           Argument + result types (JsonSchema)
  service/
    mod.rs              install/uninstall/start/stop/status
    plist.rs            com.textlog.agent.plist generator
    launchctl.rs        LaunchctlRunner trait + System/Recording impls
  macos_perm.rs         Best-effort permission probes

Specs live in docs/superpowers/specs/ and the implementation plan in docs/superpowers/plans/.

Contributions welcome — open an issue or PR at github.com/xicv/textlog.

License

MIT. See LICENSE.

Name		Name	Last commit message	Last commit date
Latest commit History 34 Commits
docs/superpowers		docs/superpowers
src		src
tests/fixtures		tests/fixtures
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
LICENSE		LICENSE
README.md		README.md

Folders and files

Latest commit

History

Repository files navigation

textlog

Table of contents

Why this exists

Features

Release history

Quickstart

Registration scope

Real scenarios

1. The "what just blew up" loop

2. Screenshot a UI bug, get a CSS fix back

3. Compiler error from a different terminal tab

4. Recovering yesterday's API response

5. Dense docs you OCR'd from a PDF

6. "Why did this work yesterday?"

7. Pre-commit context dump

8. Privacy-respecting paste

9. Cross-session memory

10. Daily-archive paste trick (v0.1.1+ — no clipboard side-effect)

11. Code review prep

12. Slack/Linear ticket triage

13. Comparing two API responses

14. "Save this for later" without manual notes

15. AI coding pair-prog feedback loop

Architecture

Why these specific choices

CLI reference

MCP tools

Configuration

Tuning recipes

Minimal-noise mode

Maximum-recall mode

Multilingual capture

Strict privacy

Privacy

macOS permissions

FAQ / troubleshooting

tl doctor first, every time

Why isn't "hi?" showing up in tl logs today?

Why do I see the same content multiple times in the MD file?

Claude can't find my MCP server

My screenshot OCR returns no text

The daemon stopped after I edited config

After upgrading the binary (cargo install textlog --force)

Nuclear reset

Pasteboard permission denied

Notifications muted

Comparison with alternatives

Development

License

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

`tl doctor` first, every time

Why isn't `"hi?"` showing up in `tl logs today`?

After upgrading the binary (`cargo install textlog --force`)

Packages