If your team runs AI agents for long, multi-step work and wants local-first memory, tracing, and scheduled activity, Raven 0.1.13 is worth evaluating as a preview tool, not yet a production platform. It fits people comfortable with a terminal, Python internals, and pre-1.0 change. Give the release notes, sandboxing, and provider behavior the closest look first.

For most teams, the question is whether this release can serve as the runtime for persistent, tool-using agents, and what to check before any real workflow depends on it. The decision-maker is a developer or operations manager who owns the host, credentials, and workflows, not someone buying a managed chatbot. At this stage you are deciding whether to pilot the project, not whether to sign a contract.

Key facts at a glance

Everything below reflects what the maintainers publish in the README, release notes, and issue tracker. This review did not install or run Raven, and nothing here should be read as a benchmark result, security clearance, or production validation.

What Raven 0.1.13 changes for operators

Raven is a terminal-first, self-improving agent harness with long-term memory, skills, local tracing, evaluation, and reusable workflows, per the repository README. Version 0.1.13, released 25 August 2026 as a public preview patch, is the snapshot under review. Release timestamps show roughly weekly patches from late July through late August 2026, which signals an active project and also a fast-moving one.

The v0.1.13 release notes say the patch addresses several operational pain points:

• Model choice is bound to the conversation rather than the process. Each session keeps its own model; switching in one session no longer changes others; and model, provider credential, and context window travel as one group so a switched model cannot run against a stale credential or window.
• One failed MCP server no longer cancels the whole turn. A transport or handshake failure becomes that server’s connection error, and the remaining servers still register.
• raven trajectory records a failing run and turns it into a deterministic regression test, with replay from recorded model responses and tool results instead of live tools. The failing trajectory minimizes into a redacted cassette that pytest can run, and reports strip config secrets, environment credentials, and sensitive paths.
• web_search is registered only when a usable key exists, so the model is not offered a tool it cannot run.
• raven doctor gained a tool capabilities section that names the config key behind each credential and distinguishes a genuinely reused credential from a missing one.
• ask_user now keeps one deadline for a whole batch, caps a call at four questions, and fails fast on duplicate questions, invalid options, or undeliverable prompts.
• Terminal input handling is more robust for SSH, tmux, busy event loops, IME input, and wide characters, and copy behavior names the mechanism actually used instead of claiming success.
• Config migration notices moved to stderr so JSON output stays parseable, and a new config saves only values you changed rather than freezing the default set.
• EverOS memory moved from 1.2.1 to 1.2.3 with no Raven adapter change and no data migration.

What this release does not promise

Several limitations matter more than the feature list for an adoption decision:

• Pre-1.0 status. The README calls the project pre-alpha, and the release notes state that CLI surfaces, plugin contracts, and runtime internals may continue to evolve.
• No PyPI publishing. The supported public install path is the GitHub Release wheel, so dependency management differs from an ordinary pip install package.
• Upgrades are manual and never automatic. Per the release notes, raven upgrade preserves configuration, sessions, and memory and installs only the latest stable release; editable source checkouts are never overwritten; and on native Windows the upgrade finishes in an external helper that you must wait on.
• The next-version architecture is not this release. The README is explicit that the Harness of Harnesses network is a future direction and that research-benchmark numbers come from an internal prototype, not the current public release.
• Open issue signals. At the snapshot, 77 issues were open. The issue tracker shows an unresolved Windows runtime problem (#119), a DeepSeek reasoning-mode failure for models missing from LiteLLM’s model map (#297), and a hardcoded episodic-log path that can point the agent at an empty file under the EverOS backend (#122), the last reported by a contributor.

Use cases the evidence supports

For a technical lead deciding whether to pilot Raven, the documented strengths map to specific workloads:

• Long-running, resumable sessions in which the model must stay bound to the conversation and concurrent sessions may use different models.
• Work that needs inspection: local tracing records LLM calls, models, token usage, cost, latency, tool inputs and outputs, subagent relationships, memory operations, and large prompts as separate artifacts.
• Persistent memory across sessions through EverOS, plus scheduled jobs and proactive activity.
• Delivery through chat gateways: the README lists twelve adapters covering Telegram, Slack, Discord, WhatsApp, Matrix, Feishu, WeCom, Mochat, QQ, DingTalk, email, and WeChat.
• Open-ended research that needs multi-source cross-checking through the Deep Research path. This depends on an external service (MiroThinker), and interactive surfaces ask before paid, minute-scale runs.

Checks before Raven carries real work

If the pilot moves forward, treat these as conditions rather than optional settings:

• Execution scope. Release notes for 0.1.12 state that choosing host execution over the sandbox warns that injected commands would run with full host privileges, and asks for explicit confirmation that defaults to No. Define what the agent may execute before first use.
• Gateway reach. Enabling a channel whose allow_from resolves to anyone triggers a warning that anyone who can message it can command the agent on that host. Keep the allowlist narrow.
• Provider and model fit. Open issue #297 shows that a reasoning-capable model LiteLLM does not recognize can lose reasoning content and fail mid-task. Confirm your provider and model combination, and run raven doctor after configuration.
• Operating system. Open issue #119 reports an embedded EverOS init failure on Windows that degraded memory store and recall. The documented path for Linux-style runs is Linux, macOS, or WSL2.
• Installer caveats. Native Windows PowerShell 5.1 rejects the short installer URL with a Permanent Redirect and needs the direct raw URL instead, per the release notes.
• Secrets and traces. Trajectory reports redact config secrets, environment credentials, and sensitive paths. Tracing is stored locally under ~/.raven/traces by default and can be disabled with RAVEN_TRACING=0; decide retention on shared machines.
• Version pinning. Patch releases have already carried breaking CLI changes, such as 0.1.12 moving interactive chat from bare raven agent to raven tui. Scripted automation should pin a release and re-run raven doctor after each upgrade.

Alternatives and evaluation questions

Raven is not the only way to build long-running agents, and the README itself positions it inside the EverMind ecosystem with plugins for OpenClaw, Hermes Agent, DeepSeek Harness, and Dify, with EverOS as the memory runtime. A team could start from a single-purpose agent and add memory, or choose a workflow platform instead of a harness, depending on whether the work is interactive, scheduled, or embedded in an existing product.

Useful questions before choosing:

• Which jobs need a persistent, resumable agent rather than a one-shot model call, and how long do they run?
• Do you need scheduled or proactive work and chat delivery, or only interactive terminal sessions?
• Which provider and model will carry the workload, and does the underlying LiteLLM layer recognize it, including reasoning support?
• Who can reach the agent, and what can it execute on the host?
• Can your team absorb pre-1.0 contract churn and re-test on patch releases?
• Is local-first memory and trace data an advantage for your compliance posture, or a burden on shared infrastructure?

Who should evaluate Raven now

Evaluate now if you are a developer or operations lead who can run a preview in a sandbox, you have workloads that genuinely need long-running agent work with memory and inspectable traces, and you accept that interfaces will change. Linux or macOS hosts, or a willingness to use WSL2, lower the friction given the Windows issues above.

Wait if you need a stable managed product for non-technical staff, you cannot scope what the agent may execute, production runs must sit on native Windows today, you are evaluating the Harness of Harnesses capabilities rather than the current harness, or you require a normal PyPI dependency with stable release management.

Verdict

Evaluate, do not deploy. On the evidence available at 3 September 2026, Raven v0.1.13 is a coherent pre-1.0 harness with preview surfaces aimed at real operator problems: per-conversation model binding, MCP failure isolation, a regression-trajectory path, honest doctor output, and local tracing. The conditions are equally clear: this was a source and documentation review with no runtime verification, the project is young by repository metadata (created 21 May 2026) and pre-alpha by its own README, 77 issues were open at snapshot, and gaps remain open in areas operators care about, including Windows support and provider model recognition.

What I can do next

If Raven or a similar agent harness is where your plans are heading, I can help turn this review into a decision through practical AI integration: first mapping the workflows you want automated, then reviewing Raven’s sandbox, provider, memory, and gateway configuration against your environment, and finally running a pinned pilot with regression checks before any headless process touches shared systems. If the review points you toward building your own agent stack instead, that work sits naturally in custom software development. I would need access to your target workflows, the host or container setup you intend to use, and the provider accounts involved, and I will not promise a result before seeing them. Contact me to start that review.

Sources

• EverMind-AI/Raven repository
• Raven v0.1.13 release notes
• Open issue 297: DeepSeek reasoning content dropped for unmapped models
• Open issue 119: model request failed on Windows
• Open issue 122: episodic log path under the EverOS backend