AgentComparison
Cognition · product · software engineering agent

Devin

An autonomous software engineer with its own IDE, shell and browser.

Group: Coding agents

Cognition documents Devin as an autonomous software engineer that writes, runs and tests code, available as Devin Cloud sessions with an IDE, shell and browser, plus Devin Desktop and a Devin CLI that can also steer cloud sessions from a terminal.

Documentation researched, not testedChecked 11 October 2026Vendor documentation, not a test

The decision in plain English

Keep the surfaces separate: Devin Cloud runs on Cognition-managed VMs, while the CLI runs on your own machine under configurable approval modes. Cognition says customer data may be used for model training by default, so a paid plan's Data Controls opt-out matters before connecting private repositories.

Best fit

Engineering teams delegating well-scoped, verifiable tickets who will review the draft PRs it returns.

Poor fit

Hard-to-verify projects with no completion criteria, or anyone needing a numeric published usage quota or default-off model training without a paid-plan opt-out.

Documented capabilities

Vendor statements, not verified task outcomes.

  • Writes, runs and tests code in a cloud workspace Source · checked 2026-10-22
  • Embedded IDE, shell and interactive browser, with takeover Source · checked 2026-10-22
  • Devin CLI with tiered approval modes and optional OS sandbox Source · checked 2026-10-22
  • Start, steer and resume Devin Cloud sessions from the terminal Source · checked 2026-10-22
  • Admin security profiles restricting network, MCP, git and GitHub CLI access Source · checked 2026-10-22

Permissions and human control

  • Documentation shows IDE takeover and an interactive browser for cloud sessions
  • In the CLI's default Normal mode, read-only actions auto-approve while file writes and shell commands prompt; Smart mode auto-runs only actions a fast model judges safe and never auto-approves installs, rm, sudo, mutating git or access to dotenv and key material; Bypass auto-approves everything; Autonomous mode pairs with an OS sandbox where shell and fetches auto-approve but edit/write tools still prompt
  • Organisation-level deny and ask rules override every user permission mode
  • Admin security profiles can allowlist network destinations and MCP servers, set git read-only or full access, remove the GitHub CLI token and make Devin's own management tools read-only; sessions without a profile run unrestricted
  • Secrets can be scoped to the organisation, a person, a repository or one session, are encrypted at rest and are bound only to the commands that need them

Documented controls do not prove that an agent always follows them.

Limits and open questions

  • Teams pricing conflicts: pricing page says $80/month plus $40 per full seat; billing docs say an $80 minimum met by seats and credits. Confirm the offered total before payment.
  • Vendor explicitly says documentation may be out of date
  • Quotas are described as daily and weekly allowances but not published as numbers
  • Model training is on by default; only paid plans can opt out, and on Teams only an administrator can
  • Security profiles are admin-configured: without one, sessions run unrestricted, and on customer-run Outposts the network policy must be enforced by the operator's own infrastructure
  • Smart approval mode is rolling out gradually and may not be available

Privacy and security

Re-verified 22 October 2026 against Cognition's Platform Terms of Service (updated 30 June 2026) and privacy policy (updated 9 March 2026). Customer Data may be used for model training and service improvement by DEFAULT (ToS 3.3.1). Paid tiers can opt out; after opt-out, Customer Data is not used for training and zero data retention is enabled with Cognition's model providers. For the Teams tier only an administrator may opt out; Enterprise customer data is never used for training without express prior written consent (per Cognition's security page). ZDR does not preclude retention of data flagged by automated safety/abuse classifiers, safety/AUP review, or legal process (ToS 3.3.3). The privacy policy adds: voice-input audio is deleted after transcription unless stated otherwise, and deletion of personal information can be requested. Use restrictions prohibit building competing products or training competing models. Vendor documents, not independently audited.

Feature-by-feature: what is documented and what is not

Official sources checked 11 October 2026. A documented feature is a vendor statement, not proof of success.

FeatureEvidence statusDetail and limit
Cloud versus localSeparate controlsCloud uses managed VMs. Local CLI permissions must not be treated as proof of Cloud behaviour. Source
CLI Normal and Accept EditsDocumentedNormal auto-approves reads, prompts for writes/shell. Accept Edits auto-approves workspace edits, but shell/outside writes still prompt. Source
Smart and BypassDifferent riskSmart judges routine actions and prompts on uncertainty; sensitive categories never auto-approve. Bypass auto-approves all tool calls; organisation deny/ask rules still apply. Source
Autonomous sandboxNot universal write permissionSandbox allows shell/fetch automatically within OS limits. Direct edit/write tools still prompt and granted write scope can expand the sandbox. Source
Cloud security profilesMust be configuredProfiles can restrict network, MCP, git and GitHub CLI access. No-profile sessions run unrestricted; customer Outpost network controls depend on operator infrastructure. Source
Training and retentionPlan-dependent vendor claimTraining may be on by default. Paid users can opt out; Teams requires an admin. Enterprise training needs express written consent. Opt-out enables provider zero-retention, not deletion of Cognition-held records. Source
Teams pricingOfficial source conflictPricing page says $80 base plus $40/full seat; billing docs describe an $80 minimum met by seats/credits. Both cap self-serve Teams at 200. Confirm the actual offer. Source
Quotas and creditsNumeric quotas unknownPro/full seats have daily/weekly allowance; Max weekly only. Teams automations/review use shared credits, not seat quota. Numeric allowances are not published here. Source

How to start with Devin narrowly

  1. Choose Cloud or local CLI first, then one small ticket with tests and a reviewable diff.
  2. Before private repository access, review training opt-out and organisation permission rules.
  3. Use Normal mode for local review or an explicitly configured Cloud security profile; do not select Bypass as a safety default.
  4. Confirm quota, shared credits and the actual Teams quote before paid work.

Official product/setup source. These are documentation-based steps, not a hands-on setup test.

Common questions about Devin

What is Devin for?

Engineering teams delegating well-scoped, verifiable tickets who will review the draft PRs it returns. The vendor describes it as: Cognition documents Devin as an autonomous software engineer that writes, runs and tests code, available as Devin Cloud sessions with an IDE, shell and browser, plus Devin Desktop and a Devin CLI that can also steer cloud sessions from a terminal.

Who should look elsewhere?

Hard-to-verify projects with no completion criteria, or anyone needing a numeric published usage quota or default-off model training without a paid-plan opt-out.

How much does Devin cost?

Devin's pricing page (re-verified 2 November 2026) lists Free at $0 with a light quota, limited models, unlimited inline edits and Tab completions; Pro at $20/month adding full model availability, Devin Cloud access, and free use of SWE-2 (Cognition's latest model; the page now shows this without the previously advertised 16 October 2026 end date); Max at $200/month with significantly higher quotas; Teams at $80/month minimum plus $40/month per full seat; extra usage at API pricing. Plan structure confirmed against the self-serve billing doc (2 November 2026). The ChatGPT-subscription payment option claim was last verified 11 October 2026 and was not re-confirmed today.

What can it access and what needs my approval?

Documentation shows IDE takeover and an interactive browser for cloud sessions In the CLI's default Normal mode, read-only actions auto-approve while file writes and shell commands prompt; Smart mode auto-runs only actions a fast model judges safe and never auto-approves installs, rm, sudo, mutating git or access to dotenv and key material; Bypass auto-approves everything; Autonomous mode pairs with an OS sandbox where shell and fetches auto-approve but edit/write tools still prompt Organisation-level deny and ask rules override every user permission mode Admin security profiles can allowlist network destinations and MCP servers, set git read-only or full access, remove the GitHub CLI token and make Devin's own management tools read-only; sessions without a profile run unrestricted Secrets can be scoped to the organisation, a person, a repository or one session, are encrypted at rest and are bound only to the commands that need them

What happens to my data?

Re-verified 22 October 2026 against Cognition's Platform Terms of Service (updated 30 June 2026) and privacy policy (updated 9 March 2026). Customer Data may be used for model training and service improvement by DEFAULT (ToS 3.3.1). Paid tiers can opt out; after opt-out, Customer Data is not used for training and zero data retention is enabled with Cognition's model providers. For the Teams tier only an administrator may opt out; Enterprise customer data is never used for training without express prior written consent (per Cognition's security page). ZDR does not preclude retention of data flagged by automated safety/abuse classifiers, safety/AUP review, or legal process (ToS 3.3.3). The privacy policy adds: voice-input audio is deleted after transcription unless stated otherwise, and deletion of personal information can be requested. Use restrictions prohibit building competing products or training competing models. Vendor documents, not independently audited.

Has AgentComparison tested it?

No. This record is built from the vendor's own documentation, checked 2026-10-11. Vendor claims are attributed, not confirmed.

How to evaluate Devin yourself

  1. Pick one low-risk task that matches its documented capabilities: writes, runs and tests code in a cloud workspace.
  2. Write the expected output and what counts as failure before you start.
  3. Connect the minimum access needed, and check approval settings first.
  4. Run it twice and compare, then review every action it took.
  5. Check cost after any free allowance. Unknown price is not free.

Related

Documented use cases: Software development and repository work.

Sources and history

Change history

Documentation overview expanded with current pricing/access sources. No hands-on results claimed.

Report a correction