DevCake runs a team of coding agents from your task board. Give it a goal; it breaks down large work, plans, implements, and reviews. You get pull requests, session transcripts, and recorded token usage and costs. You steer on the board and decide what gets merged. Every run is a clean room that starts from the ticket and the recorded artifacts alone, so what an agent knew is exactly what the record shows.
Bring a coding agent. Use your preferred agent as your setup and operations companion: have it configure the host, connect your tools, explain the queue, and diagnose failed runs. DevCake keeps the work moving between your sessions.
Clone this repository, open it in your agent, and start with:
Read
skills/devcake-ops/SKILL.md. Explain how DevCake fits my workflow, check this host, and help me set it up for one small mission. Keep auto-merge off. Verify the result and tell me what still needs my attention.
The operator skill is agent-neutral. Ask your agent to read it explicitly; automatic discovery is optional.
Start with one mission
You need a dedicated Linux or macOS host, Docker with Compose and Buildx, Python 3.12+ with uv (or pip), and model credentials. Ask your agent to check the deployment requirements, then, from this checkout:
uv tool install . # install the CLI from the version you will run
devcake doctor --json # inspect prerequisites and remedies
devcake up --bake # prepare secrets, build, start, and smoke-test
Have your agent connect a board and repositories with devcake setup
(--help lists options), or use the admin UI at
http://localhost:8080. Its login lives in the generated .env.
Configure model credentials there; the host baker builds the selected agent images.
Ask your agent: “Check my connections and branch protection, then walk me
through the first mission.” Start in
opt-in mode with one DEVCAKE-labeled ticket and inspect its PR together.
A fresh start for every run
The usual path is ONBOARD → PLAN → EXECUTE → REVIEW → merge. Large missions split into linked tickets; triage can supply the plan. REVIEW always runs. For work that produces a PR, Done means merged; auto-merge is off by default.
Each run starts a fresh session in a disposable container, with the ticket, prior transcripts and plans, relevant upstream work, and selected repositories and skills. Progress passes between runs through these artifacts; continuations within a run can resume its session.
Ask your agent: “Help me write a bounded mission with acceptance criteria, choose its repositories, and decide whether plans should need my approval.”
Give the team the right context
| Ingredient | What it does | Ask your agent |
|---|---|---|
| Work repositories | Targets for branches and PRs. Use GitHub, GitLab, Gitea, or the bundled Gitea for code, documents, and other deliverables. | “Where should this team's output go?” |
| Reference repositories | Consultation copies supplied across stages: shared libraries, specifications, examples. | “Which sources should this team consult?” |
| Memory repositories | Persistent notebooks shared across missions, bound to a board or Dev Type. Consumers get read-only context. | “Help me organize a notebook and its curation.” |
| Skills | Reusable instructions selected per Dev Type, from the editable built-in store or trusted external Git skill sources. | “Find useful skills and review their instructions with me.” |
| Dev Types | Harness, model, credentials, prompts, skills, and capacity assigned to pipeline roles. | “Staff my planning, execution, and review roles.” |
Memory is ordinary Git: a notebook's README defines its organization. Runs
contribute unverified leads to .claims/; a Memory Curator works through
them and proposes notes as PRs. You merge notes by default. DevCake
handles transport and review; learning depends on the notes you keep.
The implementation's comparative pilot is still outstanding.
Supported boards: Linear, GitHub Issues, GitLab Issues, Gitea Issues. Supported harnesses: Claude Code, Codex, Grok Build, Pi, OpenCode, Qwen Code. Mix them by role or team.
Put recurring work on a schedule
Scheduled Tasks create tickets from a template at an interval: dependency reviews, documentation checks, recurring reports, or your own maintenance routine. They use the normal mission pipeline. The same settings page holds the Memory Curator and the Relations Steward, which adds missing ticket dependencies on the board. Intake pause applies to scheduled work too.
Ask your agent: “Set up a recurring documentation check with a clear deliverable, a suitable interval, and human review.”
Operate it with your agent
- “Read
devcake status --json. Explain what's running, waiting, or blocked.” - “Investigate this failed mission using its ticket, run record, and logs. Fix the cause before granting another attempt.”
- “Review token usage and concurrency with me before we expand intake.”
- “Prepare backups and a quiet upgrade; verify health and a dispatch afterward.”
Ask your agent to use the operator skill and daily operations tutorial for these tasks.
Give your agent the tools directly. Install the CLI with its MCP extra
(uv tool install '.[mcp]') and add devcake mcp --read-only to your
agent's MCP configuration; drop the flag for a read-write connection. Every
admin operation becomes a tool, described by the API itself, so the tool
list grows with DevCake. Secret values never cross it, and every change an
agent makes is audited as such.
Ask your agent: “Connect to DevCake over MCP, read health and activity, and tell me what is waiting on me.”
You own the host and the trust. Ticket authors and repository contributors can influence agents holding your credentials. Docker access is root-equivalent; containers are not an injection-proof sandbox. Keep control ports private, protect default branches, and safeguard backups containing plaintext secrets. Auto-merge off constrains the app; forge branch protection constrains agent tokens. Ask your agent to walk through the security contract and operator duties before real work.
Evidence and next steps
DevCake is early-production software. In our own documented run, one board prompt became 54 tickets, 257 runs, and 42 human-merged PRs: self-reported evidence to assess with your agent against your own workload.
Ask your agent to explore the product overview, architecture, or design thesis with you. For changes to DevCake itself, start with Contributing and AGENTS.md.
Changelog · Roadmap and known gaps · Report a vulnerability · GPL-3.0 license
Release files for devcake-cli 0.1.4
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| devcake_cli-0.1.4.tar.gz | 51.8 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| devcake_cli-0.1.4-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 105.9 kB
Release files / devcake_cli-0.1.4.tar.gz
| Download URL | devcake_cli-0.1.4.tar.gz |
|---|---|
| Size | 51.8 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
fd15eb8fb75cdef581c930c9d14c2f624835e312ca05791007ebbcbf6ce00117
|
|
BLAKE2b-256 checksum How to use checksums |
7d0fec15836a4e343aa40901ea8223e69338aed93910178b53ac8d4ca1148a4d
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 7, 2026.
Transparency logRelease files / devcake_cli-0.1.4-py3-none-any.whl
| Download URL | devcake_cli-0.1.4-py3-none-any.whl |
|---|---|
| Size | 54.0 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
8f0210ea7f409c608427524e7d3e0a50db6f5502acfe103b62c9d1491c0ce91a
|
|
BLAKE2b-256 checksum How to use checksums |
57f6679e3231b0f3fb360a0c2164a0537dbc0f7d821bf8e96a1b9fd76900b94e
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 7, 2026.
Transparency log