Back Original

Autolith: A programming agent with a live runtime

Autolith runs in a terminal and works directly in your repository. It can read and edit files, search the workspace, run commands and tests, keep project context, and use Common Lisp without leaving the conversation.

  • Repository work. Filesystem, shell, and search tools with visible results.
  • Oversized context. Recursive inference over corpora larger than the model window.
  • Continuity. Portable conversations, memories, agendas, checkpoints, and recovery.
  • Live Lisp. An SBCL runtime it can inspect, test, and extend.

For developers who want an agent they can inspect, extend, and recover.

Autolith · Resident agent

The Autolith mascot, rendered in black ink and bitmap dither on paper

Install Autolith

Autolith v0.35.0 runs on Linux x86-64, Linux aarch64, macOS arm64, FreeBSD x86-64, NetBSD x86-64, and OpenBSD x86-64. The binary release carries its own SBCL 2.6.6, Lisp dependencies, and native helpers; Linux is served in glibc and musl builds. Installed releases check for newer tags and ask before updating. Nix is the recommended route when you want the complete build pinned from the start.

Providers sign in with a browser device flow or an API key: ChatGPT Codex subscription, Grok subscription, Fireworks AI, Anthropic API (pay-per-token), and OpenCode (zen/go). Any other OpenAI-compatible endpoint can be registered from the REPL. The default model is gpt-5.6-sol.

Yes, piping a URL into a shell is evil. Read the installer before running it. Nix is the better installation path.

Autolith executes model-generated code with your user privileges. Its process boundaries protect reliability, not against hostile code. Use it as a development agent, not as a security sandbox.

Start Autolith · Binary, Nix, or source

After the binary installer
$ autolith auth
$ autolith

Recommended · Nix
$ nix run github:luciusmagn/autolith -- auth
$ nix run github:luciusmagn/autolith

Build from the repository
$ git clone https://github.com/luciusmagn/autolith
$ cd autolith
$ ./script/bootstrap
$ ./script/check
$ ./bin/autolith auth
$ ./bin/autolith

Three captured sessions

Real recordings against v0.35.0, driven by gpt-5.6-terra on the ChatGPT Codex subscription. Each window lists the exact prompt, wall time, token usage, the resulting artifact, and how to reproduce the session. Nothing is simulated and nothing is edited; long waits are only capped at two seconds during playback.

Captured session · a corpus larger than the context window
Prompt
Use rlm.complete over autolith-source.txt in the workspace root. It is the concatenation of this repository's 121 src/*.lisp files with ';;;; FILE:' headers, about 3.1 MB, larger than your context window. Task: list every condition class defined in the source, grouped by subsystem, with the defining file for each. Use a budget of 32 calls, 400000 tokens, depth 2. Do not read or search the file directly; if the budget exhausts, stop and report.
Provider
ChatGPT Codex subscription · gpt-5.6-terra (effort high)
Wall time
3 min 15 s for the answer; 4 min 36 s for the whole session
Tokens
59.6K in the root conversation (58.7K input + 915 output); the corpus itself never enters a prompt
Artifact
All 83 condition classes, grouped by 14 subsystems, each with its defining file
Trace
inference:s2Wb1o2, the complete frame conversation, readable in-session with resource.read
Reproduce
find src -name '*.lisp' | sort | while read f; do printf ';;;; FILE: %s\n' "$f"; cat "$f"; done > autolith-source.txt && autolith, then the prompt above
Captured session · break the running agent, watch it recover
Prompt
Show the source of terminal-ui--duration-text in your running image, then redefine it live so durations of 24 hours and more render as Dd H:MM:SS, for example 1d 2:00:00. Exercise the change with assertions for 59 seconds, 61 minutes, and 26 hours, then commit the mutation with a short message. Then deliberately break the function with a redefinition that signals an error, demonstrate the failure with self.eval, and recover by discarding the broken mutation. Verify with self.eval that the committed version is active again.
Provider
ChatGPT Codex subscription · gpt-5.6-terra (effort high)
Wall time
About 6 minutes, including the crash and the recovery boot
Tokens
289.0K total (286.1K input + 3.0K output), most of it the bounded crash context of the diagnosis turn
Artifact
Private image commit "Format long activity durations". The broken formatter is used by the live status row, so the deliberate break killed the active process outright: Autolith wrote crash capsule efa6aa16, booted the pristine recovery image, restarted from clean committed source with the private commit selected, restored the conversation, and opened a read-only diagnosis turn that asked before repairing anything. The committed mutation survived its own author.
Trace
Append-only mutation journal, the crash capsule, and the private commit's replay script
Reproduce
autolith in any workspace, then the prompt above
Captured session · persist the environment, quit, resume, continue
Prompt
Define autolith-demo-parse-fraction in the autolith package: parse a string like "3/4" into a rational. Exercise it, then persist it as a private image commit. Then add an agenda entry: handle zero denominators in autolith-demo-parse-fraction. After (quit) and autolith resume: Where were we? Check the agenda, confirm autolith-demo-parse-fraction is still defined in this fresh image by evaluating (autolith-demo-parse-fraction "3/4"), then finish the agenda item: signal a clear error for a zero denominator, exercise both paths, and mark the agenda entry done.
Provider
ChatGPT Codex subscription · gpt-5.6-terra (effort high)
Wall time
10 min 11 s across two processes; each persist runs the full validation suite before anything is retained
Tokens
601.1K total for the resumed conversation (598.5K input + 2.6K output)
Artifact
Private image commits 4048480f and 9eee3ae4. The function defined before (quit) answers (autolith-demo-parse-fraction "3/4") = 3/4 in the next process without any reloading; the agenda entry written in the first session is found, finished, and marked done in the second.
Trace
Append-only conversation store, the two commits' replay scripts, and the workspace agenda
Reproduce
autolith, the first prompt, (quit), then autolith resume and the second prompt

The agent and its tools run in one process

Autolith is a terminal programming agent, not a wrapper around another agent process. Its Common Lisp image contains the provider client, terminal interface, tool registry, conversation state, persistent memories, workspace agenda, and the code that decides what happens next.

Context larger than the window is an environment, not a prompt. rlm.complete interns the corpus as a content-addressed object; the model receives only its label, size, and digest, and drives a heap-isolated Lisp environment through bounded slices, searches, and sub-inferences under an explicit call and token budget. Every run leaves a readable inference trace.

Fast search, optional immutable mode

Workspace search runs in-process through fff, a fast Rust search library. Autolith keeps the index warm instead of starting a new search process for every query.

If you want inspection without active-image changes, start it with --immutable. The mode retains read-only inspection and recovery information while withholding evaluation, mutation, persistence, checkpoint, and rollback tools.

Update the running agent without restarting it

Autolith can inspect and replace complete functions, methods, classes, macros, conditions, and global settings in its running image. An exploratory change takes effect immediately and is recorded in an append-only mutation journal. It can be exercised, discarded, or retained as a private image commit.

A useful change can then become a private image commit. The commit contains a manifest and a complete executable Lisp replay script, retained in a separate private Git history. It changes the active agent without quietly patching the tracked source repository.

A representative mutation transaction

lisp.source TERMINAL-UI--DURATION-TEXT target self
← complete tracked DEFUNself.redefine
  (defun terminal-ui--duration-text ...)
← compiled and installed in the active image

→ self.exercise
  (assert (string= "1d 2:00:00" ...))
← journaled assertion passed

→ self.diff
← one reconstructible live mutation

→ self.commit "Show days in long durations"
← private image commit commit-id
  complete reconstruct.lisp retained in private Git

Different kinds of state stay separate

Source, conversations, useful facts, live mutations, exact heaps, and disposable experiments have different lifetimes. Autolith keeps them separate instead of pretending that one database or one saved core is everything.

Persistent surfaces · Reconstruction evidence

Conversations Append-only portable S-expressions with exact resume commands and crash-tail repair.
Memories Workspace or global facts, preferences, and decisions with bounded prompt recall.
Agendas Short workspace tasks and notes, available in full on every request.
Private image commits Complete replay scripts for durable user-specific definitions and settings, retained in private Git.
Generations A saved active core, exact tracked source commit, reconstruction script, manifest, and journal position.
Worker images Immutable experimental SBCL cores with parentage and durable notes, never selected as the active agent.
Recovery A separately built pristine image that can inspect a crash and select a known-working generation without loading the damaged core.

© 2026 Lambda Symbolics OÜ

Autolith · Revised 22 August 2026