execkit

What you get that a built-in agent shell doesn't:

• Persistent sessions on local, SSH and Docker. cd and env carry across calls, and every command returns a structured result: split stdout/stderr, exit code, duration, cwd.
• Output that protects the agent's context. Secrets are redacted before the model sees them, and output budgets (tail, head, grep, a char cap) keep a noisy build from flooding the context window.
• An audit trail, a live viewer, and undo. Every command can go to a JSONL audit log, you can watch sessions live in a terminal or browser (above), and remote sessions can checkpoint and restore the workspace files.

Install

Zero-install, with uv. Add this to your MCP client config:

{ "mcpServers": { "execkit": { "command": "uvx", "args": ["execkit-mcp"] } } }

Or install it and let execkit print the config for your client:

pip install execkit-mcp && execkit-mcp setup claude   # or: cursor | gemini | codex | vscode | windsurf

Then execkit-mcp doctor checks your setup. More options (prebuilt binary, cargo install, building from source) are in the Quickstart.

Status: early 0.x. The API may change between minor versions. Read Limitations before pointing it at anything important.

Where it fits

execkit complements your agent's built-in shell or sandbox; it does not replace it. Use it when the agent needs to work on a remote host or inside a container, when you want a record of what ran, or when you want to undo file changes on a remote workspace.

The agent is the adversary. The LLM driving execkit can be prompt-injected by anything it reads, so execkit contains its own caller: a command passes the policy fence before it runs, secrets are redacted before output returns, and a changed SSH host key fails loudly instead of reconnecting into a MITM.

flowchart LR
    A([AI agent]) -->|command| F{policy fence}
    F -->|blocked| X([rejected, never runs])
    F -->|allowed| T[transport: local / SSH / Docker]
    T --> O[raw output]
    O --> R[redact secrets, bound output]
    R --> E([structured ExecResult])
    E -.-> A

Use it from an agent (MCP)

The agent gets session_create (local, ssh, or docker), session_exec, session_list and session_destroy, plus session_checkpoint / session_checkpoints / session_restore for remote undo.

State persists across calls, and every result is parsed, not scraped from a terminal:

// session_exec {"command": "cd /app && npm ci"}   -> { "exit_code": 0, "cwd": "/app" }
// session_exec {"command": "npm run build"}        // cwd is still /app
//   -> { "stderr": "Error: Cannot find module 'webpack'",
//        "exit_code": 1, "duration_ms": 3420, "cwd": "/app",
//        "truncated": false, "timed_out": false }

Commands time out after 120 seconds by default (timeout_secs per call, up to 3600). On timeout execkit interrupts the command with Ctrl-C and returns timed_out: true with exit code 124. The session keeps its cwd and env.

See crates/execkit-mcp/README.md for the operator security settings (host-key verification, key dir, audit, session limits).

Watch what the agent does

Set EXECKIT_MCP_AUDIT_DIR and every session is recorded. execkit-mcp watch shows it live in the terminal, and execkit-mcp watch --serve --open opens the read-only browser viewer shown at the top.

Search a transcript with / and jump between errors.	Rename, pin or keep a session, export it, or take a screenshot. Blocked commands show inline.

Use it as a library

[dependencies]
execkit = "0.9"                                           # local + SSH + Docker
# execkit = { version = "0.9", default-features = false }  # local + Docker only (no SSH; no russh/tokio)

use std::time::Duration;
use execkit::{Policy, Session};

fn main() -> Result<(), execkit::Error> {
    let mut s = Session::local()?
        .with_policy(Policy { allow: vec![], deny: vec!["rm".into()] })
        .with_timeout(Duration::from_secs(60));

    let r = s.exec("echo hi; echo err 1>&2; cd /tmp")?;
    // r.stdout == "hi"  r.stderr == "err"  r.exit_code == 0  r.cwd == "/tmp"
    println!("{} (exit {})", r.stdout, r.exit_code);

    let r = s.exec_with_timeout("sleep 30", None, Duration::from_secs(1))?;
    // r.timed_out == true  r.exit_code == 124; the session is still usable
    Ok(())
}

Runnable examples: cargo run --example local, EXECKIT_SSH="user:password@host:22" cargo run --example ssh, and EXECKIT_DOCKER=<container> cargo run --example docker.

Python

The same sessions from Python. pip install execkit (native bindings, no Rust toolchain needed):

from execkit import Session

with Session.local() as s:
    r = s.exec("echo hi; echo err >&2; cd /tmp")
    print(r.stdout, r.exit_code, r.cwd, r.stderr)   # hi 0 /tmp err

See crates/execkit-py/README.md.

What's in the box

• Persistent, stateful sessions over local PTY, SSH, or Docker. SSH accepts host aliases from your ~/.ssh/config.
• Structured ExecResult: split stdout/stderr, exit code, duration, cwd, truncated, timed_out.
• Base64 command framing. Comments, heredocs, !, trailing &, syntax errors and long commands do not hang the session.
• Timeouts that keep the session. A timed-out command is interrupted and the session carries on.
• Secret redaction of common token shapes (AWS, GitHub, GitLab, Slack, Stripe, Google, Anthropic, OpenAI, JWTs, PEM private keys), URL passwords, password=/token=-style pairs, and values the session assigned to secret-named variables. The echoed command is redacted too.
• Output budgets: tail/head/head+tail by line, a grep filter with context, and a char cap. Per call or a session default; the result reports what was kept.
• Undo for agent actions on remote sessions: snapshot the workspace and restore files if a command goes wrong (needs git on the remote and an explicit workspace; files only, not side effects).
• Audit log and live viewer, plus live MCP notifications to the client.
• Embeddable, never a service: cargo add, in your process; no daemon, no vendor.

Upgrading to 0.9

Breaking changes from 0.8. The details are in Upgrading to 0.9.

• SSH host keys are pinned in ~/.execkit/known_hosts, not ~/.ssh/known_hosts. Old pins are not read. The first connection re-pins, or copy them over with mkdir -p ~/.execkit && chmod 700 ~/.execkit then grep -E '^[^ ]+ SHA256:' ~/.ssh/known_hosts >> ~/.execkit/known_hosts. Old pins were keyed by bare host whatever the port: rewrite a line for a non-22 port as [host]:port, or a later port-22 connection to that host fails as a key mismatch.
• stdin is /dev/null for every command, and pagers are set to cat.
• The target needs base64.
• A timeout returns exit code 124 with timed_out: true and keeps the session, instead of an error that closed it. ExecResult has a new timed_out field.
• Session ids look like a3f9-1_local instead of 1_local.
• SshConfig has a new connect_timeout field (default 15 s). Use SshConfig::new.

Limitations

• Not a sandbox. The command policy is advisory string matching. It is easy to bypass: deny: ["curl"] blocks curl but not env curl, sudo curl or sh -c curl. The real control is a least-privilege environment: run the agent and SSH user with minimal rights.
• No interactive input. stdin is /dev/null, so prompts, REPLs and editors do not work. Use non-interactive flags (sudo -n, apt-get -y). Pagers default to cat, but running less or vim directly hangs until the timeout and closes the session. Shell history is off.
• Timeouts interrupt, they do not kill everything. execkit sends Ctrl-C. A command that ignores Ctrl-C ends the session. For long jobs, run them in the background (nohup CMD > /tmp/job.log 2>&1 &) and poll the log.
• The target needs a POSIX shell and base64. Local sessions use bash. Windows is not supported.
• Synchronous core. Fine for typical agent use; not tuned for thousands of concurrent sessions.
• SSH AcceptAny host-key mode exists for testing, behind an explicit insecure opt-in. Never use it in production.

Found something rough? Open an issue.

Contributing & security

• Contributions: see CONTRIBUTING.md.
• Found a vulnerability? Follow SECURITY.md. Please don't open a public issue for security reports.

License

Apache-2.0: embed it freely, including commercially. See LICENSE and NOTICE.