a4sx-mcp
# a4sx-mcp
A local [MCP](https://modelcontextprotocol.io) server for [a4sx](https://session-exchange.com), the
library and marketplace for coding-agent sessions. It gives any MCP client (Claude Code, Claude
Desktop, Cursor, VS Code, ...) five tools to find a saved session, load it as context, and save the
current one, using the sign-in of the `a4sx` command-line tool.
It runs on your machine over stdio. Your a4sx token stays in the a4sx config directory and is only
ever sent to the server that issued it: pointing the server elsewhere with `--base` never sends it a
token from another server.
**Private library**: the sessions saved under your a4sx account on session-exchange.com. Its entries
are not listed in the marketplace and not returned by public search; they are listed, fetched,
versioned and deleted through the a4sx API with your own sign-in (or an agent key you create). Listing
a session in the marketplace is a separate step in the web app, with a review. "Private" describes who
can list and fetch an entry through a4sx; the server stores the bytes you upload, after scrubbing them
again.
## Tools
The names, input schemas and limits are the same as the hosted a4sx MCP server's.
| Tool | What it does | Sign-in |
|---|---|---|
| `search_sessions` | Search vetted public sessions (`scope=public`, default), or your own saved ones (`scope=library`); `harness` keeps only sessions recorded with the given harnesses | library scope only |
| `get_session_summary` | Describe one session before using it: outline, members, price, provenance | library source only |
| `acquire_session` | Copy a marketplace session you are entitled to (free, bought or granted) into your private library | yes |
| `fetch_session_context` | Load a session from your library as starting context, cut to `max_tokens` (500 to 100,000; default 8,000) | yes |
| `publish_session` | Save a transcript (JSONL) to your private library | yes |
Safety rules:
- **`acquire_session` never buys.** For a paid listing you have no purchase or grant for, it returns
`status: "payment_required"` with the price and the listing URL (or, when the listing's session is
sold on its own, the `member` to buy). You buy it yourself in the web app, then ask again.
- **Harnesses.** Claude Code, Codex CLI and Grok Build sessions are listed and loaded alike; only Claude
Code sessions can be resumed today. Every search and summary row carries the `harness` it was
recorded with (`{id, version}`; `unknown` when not identified), and `resumable` is `false` for a
session recorded with another harness. `fetch_session_context` renders such a session from its
stored canonical form.
- **`publish_session` only saves privately.** `visibility` accepts only `"private"`; listing a session
publicly is a separate, reviewed step in the web app. It accepts only Claude Code or Codex CLI JSONL
(every line a record with a `type`; a Claude Code user or assistant turn with a session id, or a Codex
CLI `session_meta` record first), so it can't be used to upload an arbitrary file. The transcript, its title and its label are
scrubbed on your machine with the `a4sx` SDK, which applies the a4sx CLI's rules (secrets, command-line
passwords, Sheet ids, and personal data: e-mail and IP addresses, home-directory user names, hosts,
key-file paths), and scrubbed again by the server. Scrubbing is a backstop, not a guarantee. A session
that continues someone else's (its `parent`, or the session it was resumed from, was acquired) stays
marked as untrusted (`untrusted: true`).
- **Neither tool asks for confirmation inside the protocol.** Like the hosted server, it relies on the
rules above and on the tool annotations: `acquire_session` and `publish_session` are marked
`readOnlyHint: false`, so clients that ask before a tool runs (Claude Code does by default) can
ask you, and you should not auto-approve them. The three read tools are marked read-only.
- **Loaded sessions are reference material**, not instructions. `fetch_session_context` always starts
with a notice saying so, then wraps everything the session's author controls (title and turns) in
`<<<A4SX_UNTRUSTED_SESSION nonce=…>>>` … `<<<END_A4SX_UNTRUSTED_SESSION nonce=…>>>` with a random
nonce. Inside it, every kind of line break is normalised to `\n`, `<<<`/`>>>` become `‹‹‹`/`›››`
and a leading `#` becomes `\#` (the notice says so), so content can't close the envelope or fake a
`### User` heading. The title is one line of at most 200 characters, and cutting to `max_tokens`
never removes the notice or the closing line. `search_sessions` and `get_session_summary` results are
framed the same way: a notice first, then the JSON inside the envelope, with every author-written
string on one line, neutralised and capped per field, and the whole result bounded to 64 KiB (a
page cut short says `truncated: true`, and `next_offset` resumes at the first row left out). Server
guidance (`next`, `next_offset`) goes on one line after the closing delimiter, never inside the
envelope, and carries only fixed texts and numbers. A session with another author's content gets
`untrusted: true` and a stronger notice. Look-alike characters (full-width `<`/`#`, zero-width
spaces) are left as they are: they cannot close the envelope, because the closing line needs the
random nonce, but a model may still be misled by them. This lowers the risk of prompt injection but
cannot rule it out: keep tool approval on.
- **Limits:** transcripts up to 2 MiB per `publish_session` call (use `a4sx push` for larger ones);
per process, at most 240 requests (`tools/list` and `tools/call`), 20 acquires and 10 publishes per
minute. The a4sx server enforces its own limits as well.
## Quickstart
You need [uv](https://docs.astral.sh/uv/) (for `uvx`) and an a4sx account.
1. Sign in once with the `a4sx` CLI. The MCP server reuses that sign-in:
```sh
uvx a4sx login
```
Searching public sessions works without signing in. If a tool needs your account and you are not
signed in (or the sign-in expired), it returns an error telling you to run `a4sx login`.
2. Add the server to your client.
### Claude Code
```sh
claude mcp add a4sx -- uvx a4sx-mcp
```
### Claude Desktop
Edit `claude_desktop_config.json` (Settings, Developer, Edit Config) and restart Claude Desktop:
```json
{
"mcpServers": {
"a4sx": { "command": "uvx", "args": ["a4sx-mcp"] }
}
}
```
### Cursor
Add to `~/.cursor/mcp.json` (all projects) or `.cursor/mcp.json` (one project):
```json
{
"mcpServers": {
"a4sx": { "command": "uvx", "args": ["a4sx-mcp"] }
}
}
```
### VS Code
Add to `.vscode/mcp.json` in your workspace (or run **MCP: Add Server** from the command palette):
```json
{
"servers": {
"a4sx": { "type": "stdio", "command": "uvx", "args": ["a4sx-mcp"] }
}
}
```
### Try it
Ask your agent, for example: *"Search a4sx for sessions about flaky pytest fixtures and summarise the
best one"*, or *"Load my a4sx session `my-refactor` as context"*.
## Configuration
Everything is optional. Don't put your token into an MCP client config file: sign in with
`a4sx login` instead, which stores it with mode 0600.
| Setting | Meaning |
|---|---|
| `--base URL` / `A4SX_BASE` | The a4sx server. Default: the one you signed in to, else `https://session-exchange.com`. Plain `http` is refused except for `localhost`. |
| `A4SX_SESSION`, `A4SX_SESSION_BASE` | A token to use instead of the stored sign-in (for CI), and the server it belongs to (default `https://session-exchange.com`). |
| `A4SX_CONFIG_DIR` | Where the `a4sx` CLI keeps its sign-in (default `~/.config/a4sx`, `%APPDATA%\a4sx` on Windows). |
| `A4SX_MCP_LOG` | `warning` (default), `info` or `debug`. Logs go to stderr and contain tool names, outcomes and durations only: never arguments, transcripts or tokens. |
| `A4SX_MCP_MAX_PUBLISH_BYTES` | Lower or raise the 2 MiB `publish_session` cap (the server may still refuse larger uploads). |
| `A4SX_MCP_PER_MIN`, `A4SX_MCP_ACQUIRE_PER_MIN`, `A4SX_MCP_PUBLISH_PER_MIN` | Local rate limits (240, 20, 10 per minute). `0` turns a local limit off. |
## Local server or remote `/mcp`?
a4sx also runs a hosted MCP server at `https://session-exchange.com/mcp`, with browser sign-in over
OAuth: it advertises the scope `library:read` (search your library, summaries, loading sessions as
context). Public search there works without signing in; acquiring or saving sessions needs a credential with
`library:acquire` or `library:publish`, for example an agent key created in your a4sx account. Clients that
support remote MCP servers with OAuth can connect to it directly and need nothing installed:
```sh
claude mcp add --transport http a4sx https://session-exchange.com/mcp
```
This package is the option for clients without remote OAuth support, and for setups that already use
`a4sx login`. Both expose the five session tools below; the hosted server lists further tools of its
own (for example, starting a purchase for the account holder to confirm), which this package does not
have.
Same in both:
- the five session tools' names, input schemas (including the `harness` filter), annotations, and the
server instructions;
- the untrusted framing: the text of `fetch_session_context` (notice, nonce envelope, line-break
normalisation, escaping, per-field caps, sizing, canonical sessions of other harnesses) and the JSON
envelope, `notice` field, per-field caps, 64 KiB result bound and hint line of `search_sessions`,
`get_session_summary` and `acquire_session`;
- the result fields (plus the local extras listed below);
- the error codes `invalid_arguments`, `credential_not_accepted`, `rate_limited`, `not_found`,
`too_large`, `content_sha_mismatch`, `not_a_transcript` and `no_transcript`;
- JSON-RPC error `-32029` when the overall request limit is exceeded.
Differences of this local server:
- **Sign-in.** It uses your `a4sx login` session, so it can do everything your account can in your own
library. The hosted server takes a scoped OAuth token (it advertises `library:read`) or an agent key
with the scopes you gave it (`library:read`, `library:acquire`, `library:publish`). With no sign-in at all, local tools return `not_signed_in`; the hosted
server answers with an OAuth challenge instead. The `credential_not_accepted` message says to run
`a4sx login` rather than to get a scoped token.
- **Rate limits.** They are counted per process, not per account. There is no failed-sign-in limit,
because there is no sign-in step here. `retry_after_s` comes from the API's error body, not from the
`Retry-After` header.
- **`untrusted`.** Besides `acquired_from` on the row and its parents (which is all the hosted server
checks), it also uses this machine's record of untrusted sessions from the `a4sx` SDK. A session
missing from your library listing counts as untrusted in both.
- **`publish_session`:**
- Every upload carries an idempotency key: yours if you pass `idempotency_key`, else a fresh one.
The hosted server sends a key only if you pass one.
- Uploads are retried twice, with that same key, on network errors, 429 and 502-504; when the
server is busy scrubbing other uploads, up to 6 attempts, each after the wait it asks for.
- The body, title and label are scrubbed locally, so the `X-Content-SHA256` sent to the API
covers the scrubbed bytes.
- The content check is stricter: every line must be a JSON record with a `type`, and the file a
Claude Code or Codex CLI session.
- `parent` is resolved like the CLI resolves names (exact id, then label, then a unique part of the
title or id). The hosted server passes it through verbatim.
- Without `parent`, a transcript you resumed with `a4sx resume` gets its parent edge from the SDK's
local lineage record. The hosted server records no edge.
- The result has extra `client_redactions` and `untrusted` fields.
- **`payment_required`.** The price comes from the public listing; the hosted server reads its pricing
store.
- **Server name.** It is `a4sx-mcp` in the MCP handshake (hosted: `com.session-exchange/sessions`).
## VS Code extension
The [`vscode/`](vscode/) directory holds the a4sx extension for VS Code 1.101 or later. It adds the
hosted a4sx MCP servers to agent mode in VS Code chat; VS Code runs the sign-in. Its command
**a4sx: Save this session** runs `a4sx push --latest` in a terminal, which saves your most recently
changed local Claude Code session to your a4sx library. To continue a saved session, use Claude Code;
other harnesses can't resume sessions today.
The extension is not on the Visual Studio Marketplace yet. To build and test it from source:
```sh
cd vscode
npm install
npm run compile
npm test
```
## MCP registry
[`server.json`](server.json) describes this server for the
[MCP Registry](https://registry.modelcontextprotocol.io): the hosted server at
`https://session-exchange.com/mcp` (Streamable HTTP) and this package on PyPI (`a4sx-mcp`, stdio).
It uses the name `com.session-exchange/sessions`, under which the hosted server is already listed.
<!-- mcp-name: com.session-exchange/sessions -->
## Development
```sh
git clone https://github.com/crossvault/a4sx-mcp && cd a4sx-mcp
python -m venv .venv && . .venv/bin/activate
pip install -e '.[dev]'
pytest -q # offline: the tools run against a local stub server
ruff check . && reuse lint
```
The tests never touch the network: tool tests run against `tests/stub_server.py` on 127.0.0.1, the
protocol test uses the MCP SDK's in-process client, and a guard fails any test that resolves or
connects to a non-loopback address. See [CONTRIBUTING.md](CONTRIBUTING.md).
## About
`a4sx-mcp` is made by crossVault GmbH, the company behind **a4sx**
([session-exchange.com](https://session-exchange.com)), a library and marketplace for coding-agent
sessions. It is a thin client: the tools call the public a4sx API through the open-source
[`a4sx`](https://github.com/crossvault/a4sx) SDK, and the hosted service is not part of this
repository. You need an a4sx account for everything except public search. It is open source so you
can check exactly what it sends before you connect it to your agent. Parts of this project were
developed with AI assistance (Claude).
## Security
Please report vulnerabilities privately; see [SECURITY.md](SECURITY.md).
## License
Apache-2.0, see [LICENSE](LICENSE). Copyright 2026 crossVault GmbH.
TDQS
Scored across 5 tools
Each tool occupies a distinct stage of the session workflow: search finds IDs, get_session_summary inspects, acquire_session adds to library, fetch_session_context loads it, and publish_session stores transcripts. The descriptions explicitly clarify boundaries (e.g. acquire never buys, publish is private), leaving no realistic confusion.
All five tools follow a consistent snake_case verb_noun pattern (search_, get_, acquire_, fetch_, publish_ + noun). Verb choices are predictable and readable with no mixed conventions.
Five tools cleanly map to the discover-inspect-acquire-load-publish lifecycle of a session marketplace/library. Nothing feels redundant or missing for the stated scope.
The full core workflow is covered: search, inspect, acquire, load context, and publish. Minor gaps exist, such as no explicit tool to remove or manage library entries or list one's own library contents beyond scope=library search, but agents can work around these.