errand
Run a coding agent from a chat channel, in a sandbox it cannot escape.
A message in one configured channel opens a thread and starts a session. The agent works in a project directory of your choosing and nowhere else. Replies in the thread are prompts; what the agent says, runs, and changes comes back to the same thread.
- One channel, one thread per session, several sessions at once.
- The agent runs sandboxed. There is no unsandboxed mode.
- A queue bounds how much work reaches the model provider at once, so a burst of messages cannot get an account rate limited.
- The chat token never enters a sandbox.
Full documentation, including every configuration field, is in
docs, built with VitePress from docs.
What you need
- Linux, with either bailey or rootless
podman on the host. The sandbox is not optional, so one of them has to be
there: without a backend the daemon refuses to start rather than running a
session unconfined.
sandbox.backendpicks which, and defaults tobailey. - A chat bot token, and one channel for it to serve.
- A credential for a model provider.
bailey confines a session as a host process, using Landlock for the filesystem and seccomp for the syscall surface, so sessions use the host's own tools and there is no image to build. podman runs the session in a container from an image you provide instead. Sandboxing covers what each one enforces, and what the daemon does when a host cannot enforce it.
Running it
There is a file to copy in config.example.json, and a schema beside it that gives an editor completion and checking.
The configuration file is read from ~/.config/errand/config.json, then
/etc/errand/config.json, then config.json in the working directory.
ERRAND_CONFIG names one outright and skips the search. A minimal one:
Everything else has a documented default. The daemon refuses to start rather
than run with a guarantee it cannot keep: if the backend cannot enforce
everything configured on this host, it says which and stops, unless
sandbox.requireFullEnforcement is set to false.
Two models, one session
A session can ask a cheaper model of the same provider one question about one thing that already exists: a long log, a large file, a diff. It is shown that one thing and nothing else, has no way to run or read anything, and answers in text, so what comes back is a description to check rather than a decision to follow. The point is what it keeps out of the session's own context, which is paid for again on every later turn.
Absent means no delegation at all. With it, the agent gets a delegate command
and is told how to use it; !status reports what it cost and how much it kept
out. To have the cheaper model do the work rather than describe it, use
!model instead.
The interface
An optional local web interface reads a session as it happens, browses the
project, and starts new ones. It has no login: the address it binds to is the
access control, and a public bind is refused rather than warned about. Build it
once (cd web && bun install && bun run build, then
cargo build --release), then add a web section to the configuration:
Set "observer": true to serve one that can watch and read but change nothing.
In a thread
!help lists what can be typed, !usage says how much of the provider's usage
window is left and when it resets, and !model moves the session to another
model of the same provider, keeping the conversation. Plan on the capable one,
switch to the cheap one to carry it out, switch back to review. The same commands are registered as slash
commands, so they can be picked rather than remembered. A message starting
!!! is an aside: the people in the thread see it and the agent is never told.
Full documentation, including every configuration field, is in
docs, built with VitePress from docs.
Running it as a service
Definitions for OpenRC and systemd are in packaging, along with what to prepare and what its exit codes mean.
Status
Early, but complete enough to run: chat, sandboxed sessions, the interface, and service definitions for both init systems.
Development
cargo build --release produces target/release/errand. The web interface
builds out of web into dist/web:
&& &&
The reference pages are generated from the schema, the command table, and the
character table; cargo test fails when what is committed no longer matches,
and cargo test regenerate_the_reference_pages -- --ignored brings them back
in step.
The documentation site builds with VitePress from docs, with bun:
cd docs && bun install && bun run build
Project rules are in AGENTS.md.
License
MIT OR Apache-2.0, at your option. See LICENSE-MIT and LICENSE-APACHE.