AI Session Search (aise)

Ultra-fast Rust-based search of your local AI coding sessions, reindexing on the fly. Reads Claude Code, Claude Desktop, ChatGPT Codex, Cursor, Antigravity, Pi, Prime Agent, Google AI Studio, and Gemini CLI, on macOS, Linux, and Windows.

Search nine harnesses from one index

Useful AI coding history is already on disk. It is just hard to find, and in a range of totally different formats.

aise parses all nine formats into one index and searches it efficiently. Your fixes from last month, the files an agent overwrote, and the record of why a run failed become instantly findable.

Query that index from the command line, a Rust crate, a Python package, or an MCP server your coding agent calls mid-conversation.

Install

Install the global aise command with uv (recommended):

uv tool install ai-session-search && aise integrations install

or with Rust's cargo:

cargo install ai-session-search --locked && aise integrations install

Both install the newest stable release. The second command is the one that touches your tools: it registers the MCP server and writes guidance for each detected harness, and --dry-run previews every write.

Common tasks

Stop-hook feedback, a PreToolUse block, a local-command caveat, a task notification come from the harness, so ordinary results leave them out. They are usually why an agent stopped mid-task; --kinds harness-notice searches them.

A subagent's work lives in its own session, so the parent transcript shows only that it ran. --session-kinds subagent lists those sessions, and --parent-session narrows to one run's children.

Quick start

# Optional: build the index now rather than on your first query
aise reindex

# What have I been working on?
aise list
aise search "database migration" --provider codex

# Find a specific moment, then read around it
aise messages search "permission denied" --role user
aise show SESSION_ID
aise resume SESSION_ID

# Get a file back, into a .recovered sibling that never overwrites what is there
aise files history src/app.py
aise files extract src/app.py --restore

aise COMMAND --help documents every parameter and default. Date bounds accept ISO, EDTF, durations, and common natural-language forms; aise dates lists each accepted family with examples.

MCP server

Ask Claude Code why last Tuesday's run stopped, or which version of a file it wrote before the refactor, and it queries this index and answers inline without you leaving the conversation. The server is a subcommand of the same executable, so there is nothing else to install.

aise integrations install    # register the server with every detected harness
aise integrations status
aise mcp serve               # direct stdio use by a client

The server exposes search_sessions, search_messages, list_sessions, get_session, get_resume_command, get_index_status, run_skill_capability, and query_session_index for read-only SQL.

What a harness receives depends on what it can read, in the format and location it expects. --client limits the install to named harnesses, and no harness gets hooks.

No MCP tool changes your sessions or configuration, and publishing a bundle stays a CLI and library operation. A query does refresh the index, writing aise's own state; --index-refresh existing-only opens SQLite read-only. Tool inputs are closed schemas, so a misspelled field fails with its exact argument path.

A failed install rolls back what it had already written, and when it cannot finish the rollback it keeps the receipt and prints the exact aise integrations recover command. aise integrations status reads every target under the installer's own transaction lock, so it never shows you a half-applied state.

Speed

Measured on macOS with Apple silicon against an index of roughly 2.5 million messages, best of three warm runs. Each command also carried --index-refresh existing-only --format json, so the numbers are query cost with no refresh in the path:

Startup accounts for 43 ms of that first row, so aise list spends nearly all its time launching. Parsing, indexing, and querying are native code with no Python interpreter in the path; cargo install puts that binary straight on your PATH, while the uv install reaches it through a small Python console script.

Indexing is incremental. A refresh over this index walked 6,324 source files in 13 s and reparsed only the 6 sessions whose bytes had changed, so steady-state cost follows new data. Refresh runs on access by default, including when an MCP client asks a question, and --index-refresh existing-only skips it when you want the query cost alone.

A scoped query costs less than these rows (filters and scopes), and your index is probably smaller than this one, so treat these timings as a ceiling.

Session sources

Everything is read from local files on your own machine. Cloud-only ChatGPT, Claude, and Antigravity conversations are never copied from a vendor account, and are not searchable here.

Prime Agent is built on Pi but keeps its own root and child-run transcripts under ~/.prime/agent, so --provider pi never returns Prime sessions and the two keep separate IDs and resume commands. Codex desktop and Codex CLI/IDE share ~/.codex, so they index together.

The codex provider reads that local transcript format only. It does not parse Chat/Work browser state or sync cloud history, which OpenAI documents as separate histories in the merged desktop app.

These IDs are the TOML keys in provider configuration, and aise config paths prints each source's enabled state and effective roots.

Filters and scopes

# Narrow a search to one tool, one project, and one week at once
aise list --provider claude-desktop --path ~/source/project --since 7d

# Get the newest session in this directory or a descendant directory
aise list --path ~/source/project --limit 1

# Find what you asked for, without the compaction summaries repeating it back
aise messages search "regression" --role user --include-compaction false

# Find which tool call touched a file, by searching the argument rather than the prose
aise messages search "Cargo.toml" --field tool-argument --argument-path /path --tool-name-contains Edit

# Skim a long result page: keep each message's first 5 lines, or its last 5 with -5
aise messages get SESSION_ID --role user --lines-per-message 5

# Read the 75 most recent user turns. --order picks which 75; they print oldest-first
aise messages get SESSION_ID --role user --limit 75 --order newest

# Walk a long session in non-overlapping chunks instead of growing --limit
aise messages get SESSION_ID --seq-from 0 --seq-to 499
aise messages get SESSION_ID --seq-from 500 --seq-to 999

# Get every version of a file the session can account for, one JSON object per version
aise files extract path/to/file.rs --all --format jsonl

# See which setting or default decided a surprising result
aise messages search "database lock" --query-mode fuzzy --limit 20 --receipt-level full --format json

Date bounds test a session's span, created_at through updated_at: --since needs it to end on or after the bound, --until to start on or before, --when to overlap it. Spans can contain gaps, so a week-long match may hold one edit. Message, file, and event analytics use each event's timestamp.

With no configured operation or purpose default, omitting the limit on a Rust, CLI, or Python message search returns every literal, regex, or no-text match; --all-results says so explicitly in a script. Session-level and MCP searches page instead, so a client is never flooded. Fuzzy search always needs a finite page, and --limit 0 means the whole corpus only where that command's help says so.

Line windows share one sign convention: positive keeps the first N lines, negative the last N, 0 everything. aise show --transcript-lines windows a whole rendered transcript; --lines-per-message caps each returned message on its own. Both shorten what prints and nothing else, so a long page stays skimmable with no hit lost: matching, ranking, result count, pagination, and context membership are unaffected.

The index only narrows the scan: exact and regex modes recheck your pattern on every candidate, so indexing cannot change your results. --receipt-level summary shows how a surprising query was planned; full adds where each unpassed value came from. MCP returns the same receipt as structured data.

Generated message-search defaults

Search indexed AI-session messages while separating result selection, context, presentation, optional payloads, and receipts.

These are shipped defaults from an empty configuration. aise messages search --describe --describe-surface cli|mcp|python|rust resolves the same contract with the active configuration. Positive limit counts result rows; signed lines_per_message selects the beginning, end, or complete text of each already-selected message.

Export and analysis

# One session as Markdown you can paste into an issue; json and text also work
aise export SESSION_ID --format markdown

# Keep last week's sessions as files, before a harness rotates them away
aise export --since 7d --output-dir ./week

# What the whole corpus holds, as index.md, taxonomy.md, knowledge-graph.md; scopes narrow it
aise analyze --output ./analysis

# The same analysis, defined once as a reusable JSON policy
aise analyze --policy ./analysis-policy.json --output ./policy-analysis

A bundle destination must not already exist, so a run never merges into or replaces one you already have. aise publishes the complete directory in one atomic step and returns a receipt, so an interruption leaves no half-written bundle.

Python and Rust APIs

Use these to build on the index instead of shelling out to aise and parsing its output.

import ai_session_search as aise

search = aise.SessionSearch()
search.refresh()
messages = search.search_messages(
    "authentication",
    aise.MessageSearchRequest(role="user", limit=50),
)

use ai_session_search::service::SessionSearch;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let app = SessionSearch::load()?;
    let status = app.index().status()?;
    println!("{} repairs pending", status.repair_commands.len());
    Ok(())
}

A long refresh or search releases the GIL, so it will not block your other Python threads. The Rust crate keeps MCP, SQLite row, and PyO3 types out of its public contracts, so you work with provider-neutral types. The library guide has the full API, feature flags, and concurrency semantics.

Configuration

Nothing needs configuring: with no file, every provider is enabled at its platform-default paths. Google AI Studio is the exception, shipping no default root, so set [providers.aistudio].paths to search it. aise config init writes ~/.ai-session-search/config.toml with every setting commented out but the database and cache paths, so later releases' defaults keep applying to whatever you leave alone. aise config example previews it first.

aise config show      # effective settings
aise config origins   # where each setting came from
aise config paths     # resolved paths and per-source roots

Precedence runs from CLI argument, to environment variable, to TOML, to the platform default.

Search is unrestricted by default. To keep an agent out of your other projects, [search.scope] confines it to absolute roots you list, the invocation directory, and live MCP client roots. That mode refuses free-form aise db query SQL throughout, and when no authoritative root resolves it fails and names the fix instead of searching wider. See search access scope. Legacy settings import through aise migrate config.

Index maintenance

Never copy a live SQLite index by hand, because the file alone omits its WAL state. aise migrate database takes an online backup, verifies it, and publishes the destination atomically, never overwriting a path that already exists.

aise migrate database --help   # copy an index to a new location safely
aise migrate verify --help     # check a source and destination against the receipt
aise migrate recover --help    # finish or roll back an interrupted migration

aise doctor reports exact allocated and reclaimable bytes, and names aise compact when reclaiming them is worth an exclusive lock and temporary disk space. Migration never vacuums the source, so run aise compact against the published destination.

Command reference

aise COMMAND --help documents each one in full.

Architecture

One Rust implementation backs the CLI, the MCP server, the Rust crate, and the Python package, over one index lifecycle and one provider registry, so a new provider becomes searchable on all four at once.

Rust owns parsing, querying, migration, and filesystem publication. Filtering by provider, session, path, date, role, message kind, tool, sequence, or canonical tool-argument JSON pointer runs inside the index, and one --kinds filter owns message-class selection.

Reconstruction holds one version at a time, so --all never accumulates them; files extract --restore writes a .recovered sibling and never overwrites; export and analysis bundles publish atomically and are never replaced, and analysis bundles checksum every artifact.

Install details and uninstalling

Other paths to the same build: uvx --from ai-session-search aise --help runs it once without installing, uv add ai-session-search adds a project dependency, and python -m pip install ai-session-search also works.

Wheels cover CPython 3.12 through 3.14 on manylinux2014 x86_64 and aarch64, macOS x86_64 and arm64, and Windows x86_64. Building from source needs Rust 1.88 or newer and a C compiler, because SQLite is built from bundled C.

Installing the package never touches your configuration: no aliases, no MCP client edits, no instruction files or skills, and not one transcript read. aise integrations install is the second step that does all of that, adds the aisearch and ai_session_search aliases, and starts building the index in the background. Run aise doctor afterward for readiness and recovery guidance.

After an upgrade, the first aise command brings the skill and instructions it installed up to date. Set [integrations] refresh_after_upgrade = false to turn that off.

To remove it, take the integrations out before the executable, because the uninstaller is what knows which entries are its own:

aise integrations uninstall
uv tool uninstall ai-session-search    # or: cargo uninstall ai-session-search

Uninstalling removes what aise installed and leaves your index, configuration, and session files alone. The installation guide covers source builds, per-component flags, custom destinations, updates through aise package update, and recovery.

Development

uv sync --locked --all-extras
uv run maturin develop --uv    # required: Python tests import the compiled extension
./run_ci_local.sh              # authoritative gate; run before proposing a commit

The gate builds isolated config, cache, and database state, so it never touches a real user index. Run it with no environment prefix: prefixing RUSTC_WRAPPER= disables sccache and forces a cold rebuild.

While iterating:

cargo test -p ai-session-search <test name>
uv run pytest tests/<file>.py -k <test name>
cargo clippy --workspace --all-targets --all-features --locked -- -D warnings
uv run ruff check . && uv run mypy ai_session_search tests
uv run python -m mypy.stubtest ai_session_search --concise --ignore-disjoint-bases

CHANGELOG.md records what changed in each release. CONTRIBUTING.md covers review expectations, and the documentation index separates user guides from maintainer and migration records.

License

Apache License 2.0. Third-party dependencies keep their own licenses, and release artifacts ship dependency inventories and software bills of materials.